US 9619864

Let me also search for any CAFC docket activity involving this patent number, as requested.

Summary — U.S. Patent No. 9,619,864 (US9619864B2)

Bibliographic data

Sources: Google Patents (patents.google.com/patent/US9619864), Justia Patents (patents.justia.com/patent/9619864), USPTO-published PDF (patentimages.storage.googleapis.com/ec/97/47/00355339537f0d/US9619864.pdf).

Abstract (verbatim)

An image processing apparatus includes a high-frequency component translating unit, a high-frequency component extracting unit, a detail-gain generating unit and an image output unit. The high-frequency component translating unit extracts and translates first high-frequency components of an input image to generate a first image. The high-frequency component extracting unit extracts second high-frequency components to generate a second image. The detail-gain generating unit stores a conversion table and generates detail gains respectively associated with input pixels in the input image according to pixel values of the input pixels and the conversion table. The image output unit calculates a weighted superposition of the first image and the second image and generates a high frequency component of an output image according to the weighted superposition and the detail gains.

Plain-language overview of the technology

The patent addresses single-image super-resolution/sharpeness enhancement. Instead of using multiple low-resolution frames (memory-intensive) or a learning-based high-resolution database (computationally intensive), it adds synthetic high-frequency detail to an interpolated (blurry) image using two parallel filtering paths plus per-pixel "detail gains" that carry pseudo-randomness (from a periodic, continuous conversion table) so adjacent pixels get different gain values — increasing perceived contrast/sharpness while remaining temporally stable for video. The first path uses "overshoot" high-pass filters (positive coefficients at the mask center and periphery, negatives in between) to produce translated high-frequency detail; the second path uses ordinary high-pass filters to produce untranslated high-frequency components. The two filtered images are combined (weighted sum), multiplied by the detail gains, and blended back into the input image.

Independent claims overview

Claim 1 (apparatus): An image processing apparatus having at least one processor connected to a memory, configured to (a) extract and translate first high-frequency components of the input image to generate a first image; (b) extract second high-frequency components to generate a second image; (c) generate detail gains per input pixel using a conversion table and pixel values — where the gain for a given "first" pixel is tied to the summed pixel values of that pixel plus neighboring pixels that are separated from it by at least one pixel; and (d) compute a weighted superposition of the first and second images and produce the high-frequency component of an output image from that superposition and the detail gains. Key limiting feature: the non-adjacent (at-least-one-pixel-away) neighbor sampling, which makes the gain field vary more between neighboring pixels.

Claim 10 (method): A method performing the same four steps as claim 1: extract-and-translate first high-frequency components → first image; extract second high-frequency components → second image; generate detail gains from a conversion table and pixel values (with the same at-least-one-pixel-away neighbor-sum limitation for the first pixel's gain); and calculate the weighted superposition and generate the output high-frequency component from the gains plus the superposition.

Claim 19 (method): The same basic method as claim 10 but without the neighbor-sum limitation of claims 1/10; instead it affirmatively requires that generating the first image comprise filtering the input image with at least two overshoot high-pass filters to create multiple high-frequency translation images, then forming the first image by selecting the pixels with the highest absolute pixel values across those translation images. (Dependent claim 20 further requires the overshoot-filter masks to have positive coefficients at the center and periphery and negative coefficients elsewhere.)

CAFC 2026 docket check — result and uncertainty

I searched for a CAFC 2026 appeal docket and for PTAB/litigation involving this exact patent number. No CAFC appeal, IPR, or district-court case tied to patent 9619864 was found (the direct CAFC query returned no results; broader queries surfaced only unrelated matters). Absence of results is not proof of non-existence — public docket coverage can lag or be incomplete — but based on the searches available today, I have no evidence of any 2026 Federal Circuit proceeding involving this patent. All bibliographic data above is corroborated by multiple independent sources (Google Patents, Justia, and the USPTO image PDF) and matches the authoritative full-text provided, so I have high confidence in it.

Based on my searches of available litigation databases and general web sources, I found no known litigation involving US Patent No. 9619864 ("Image processing apparatus and method for increasing sharpness of images").

Findings

• Patent No.: US9619864B2
• Assignee: Realtek Semiconductor Corp.
• Filing / Grant: Filed Aug. 30, 2015; granted Apr. 11, 2017
• Status per Google Patents: Active (4th-year maintenance fee paid as of May 2020); anticipated expiration 2035-08-30

Litigation Search Results

Searches performed:

• General web searches for "9619864" patent litigation and "US9619864" lawsuit returned no litigation results.
• Search of Unified Patents portal (site:portal.unifiedpatents.com 9619864) returned no case listings.
• Searches combining the patent number with the assignee ("9619864" Realtek Semiconductor litigation, "9,619,864" patent asserted) returned only inventor/patent bibliographic pages (e.g., patentleaderboard.com) and no court dockets, complaints, or PTAB/IPR filings.

Caveats

• My searches were limited to publicly indexed sources. There could be matters filed in federal district courts (PACER), the ITC, or the PTAB that are not fully indexed by the search tools used, or that are filed under seal or under a different reference.
• The patent is not expired and appears to be maintained, so future assertion remains possible.
• If you need a more definitive clearance, I would recommend a paid docket search (e.g., Docket Navigator, RPX Insight, LexisNexis/Westlaw litigation databases) or a direct PACER query, which have more comprehensive coverage than general web searches.

Bottom line: Based on the information available to me, there is no known litigation involving US Patent 9619864 as of April 26, 2026.

I'll verify the structured data against public PTAB sources before finalizing the analysis.

Proceedings overview

There are zero AIA trial proceedings on US Patent 9,619,864 — no IPRs, no PGRs, no CBM reviews — per the USPTO Open Data Portal (the canonical structured data in this prompt reports no proceedings as of the most recent ingest), and independent web searches of the PTAB's public docket, PTABLitigationBlog, and general sources surfaced nothing naming this patent as the challenged patent. With no claims invalidated and no claims tested, the patent is fully intact and untested at the PTAB. For a defendant, that is a double-edged signal: the patent has never been weakened, but the invalidity arguments that would go into an IPR petition (e.g., single-frame super-resolution / sharpening art from the image-processing field) remain entirely available — no estoppel from prior proceedings binds you.

(Note: my searches did surface PTAB matters captioned "Realtek Semiconductor Corp. v. ParkerVision, Inc." — e.g., IPR2025-00324 and IPR2025-00325 — but those challenge ParkerVision's U.S. Patent No. 7,865,177 and are unrelated to 9619864. Realtek appears there as petitioner, not as patent owner. Do not confuse those with proceedings on this patent.)

No per-proceeding sections are provided because there are no proceedings to report.

Strategic summary

Claim status: all 20 claims UNTESTED. Claims 1–20 of US9619864B2 have never been through an AIA trial. No claim has been canceled, no claim has been sustained by a Final Written Decision, and no petition has even been filed against this patent. The patent remains Active (4th-year maintenance fee paid May 2020; anticipated expiration 2035-08-30), so any assertion of it today rests on a full, unimpaired claim set. Because the claims have never been construed by the PTAB, the patent owner has no adverse claim-construction record, and you have no estoppel handicap — but you also have no favorable FWD to lean on.

Estoppel landscape: wide open. Because there have been no prior AIA trials involving this patent, § 315(e)(2) estoppel attaches to nobody. A defendant served with a demand letter or complaint today is free to raise any § 102/§ 103 ground that a skilled search would uncover — there is no prior petitioner whose grounds are off-limits, and no ground has been "used up." This is the cleanest possible posture from which to build an IPR petition or an invalidity defense in district court. Note the § 315(b) one-year bar: if you are sued, your IPR petition window runs from service of the complaint, so act early. (Realtek's own recent experience in the ParkerVision copycat-joinder cases, IPR2025-00324/-00325, is a cautionary tale: the Board denied institution for a time-barred petition seeking joinder, holding that time-barred petitions proceed "only in exceptional circumstances." Don't count on joinder as a fallback.)

Pattern signals: none on this patent. No petitioner has filed multiple IPRs against 9619864, the patent owner has never defended it at the PTAB, and no defensive aggregator (e.g., Unified Patents) appears in the chain for this patent. The complete absence of PTAB activity is itself informative: this is a Realtek semiconductor image-processing patent that has been maintained for nearly a decade without attracting a challenge — plausibly because it has simply never been asserted aggressively enough to provoke one. That means the first serious assertion may well be met with the patent's first IPR.

Recommended next steps

• There is no FWD to cite and no proceeding to track — the PTAB docket for this patent is empty. If you are facing an assertion, your job is to create the record, not react to it. Commission a prior-art search targeting: (1) single-image sharpening/unsharp-masking with overshoot filters (the "overshoot high-pass filter masks 360A/360B" of claims 7–8 and 16–17), (2) texture-detail injection / "super-resolution from a single frame" literature, and (3) gain-modulated detail blending. The cited art on the face of the patent — e.g., US 6,285,798 (Eastman Kodak, edge contrast gain-control), US 6,714,688 (edge contrast of interpolated images), US 9,123,140 (Pixelworks, recovering details in single-frame super resolution) — is a good starting map of where the closest prior art lives.
• Mind the § 315(b) clock. If litigation is filed against you, the petition deadline is one year from service. Given the untested claim set, an early, well-developed IPR is the highest-leverage move — the claims have never survived scrutiny, so there is no history suggesting they are robust.
• If you are the patent owner (Realtek): the clean PTAB record plus the payment of maintenance fees suggests the patent is considered commercially relevant. There is no immediate threat, but be aware the first assertion will likely draw the first IPR, and the ParkerVision-line decisions (including Realtek's own precedential loss in IPR2025-00324/-00325) show the Board is skeptical of late, copycat, or time-barred petitions — a stance that cuts in favor of an early, independent petitioner and against serial petitioners.
• Re-verify at filing time. The USPTO ODP data reflects the most recent ingest, and general web searches can lag dockets. Before making any filing decision, run a direct query of the PTAB's E2E/PRPS system and the USPTO Patent Center for "9619864" to confirm no petition was filed in the interim — but as of 2026-09-05, no PTAB activity exists on this patent, and that absence is a signal that the patent has not yet been stress-tested, not that it is strong.

Inventors

All four named inventors were employed by Realtek Semiconductor Corp. at the time of filing, per the recorded assignment (reel/frame 036462/0935) which runs from the inventors to Realtek, and per inventor bibliographic listings (e.g., patentleaderboard.com shows each inventor's patents under Realtek Semiconductor):

• Cheng-Yueh Chen — Realtek Semiconductor Corp.
• Chun-Hsing Hsieh — Realtek Semiconductor Corp.
• Jun-Yu Yang — Realtek Semiconductor Corp.
• Zhi-Gang Sun — Realtek Semiconductor Corp.

No unusual pattern observed: nothing in the available record suggests departure-from-assignee or a pre-filing consolidation of inventor rights. This is a conventional employee-inventor assignment to the employer.

Original assignee

• Realtek Semiconductor Corporation (Taiwan) — named assignee on the issued patent (Google Patents legal event, reel/frame 036462/0935).
• Line of business: fabless IC design / semiconductor supplier (audio codecs, network interface controllers, card-reader, TV/display SoCs, etc.). The image-sharpening claims are consistent with Realtek's display/scaler and video-processing product lines.
• Product embodiment: not verified for this specific patent, but Realtek is an operating semiconductor company, not a licensing vehicle.
• Current status: operating; large-entity 4th-year maintenance fee paid 2020-05-04 (Google Patents event MAFP), patent status "Active," anticipated expiration 2035-08-30.

Assignment timeline

I was unable to pull the USPTO Assignment Center records directly during this session; the findings below rest on the USPTO-derived legal-events feed indexed by Google Patents (reel/frame 036462/0935), which is the same underlying record the Assignment Center would surface. The Google Patents legal-events list shows exactly one recorded assignment — the original inventor-to-Realtek conveyance — and no post-issuance assignments of any kind (no Merger, Change of Name, Security Agreement, License, or Release entries appear after grant).

• 2015-08-25 (executed; "effective date" per assignment) / recorded 2015-08-31 — Reel 036462/0935
  • Conveyance: Assignment of Assignors' Interest
  • Assignor: Cheng-Yueh Chen, Chun-Hsing Hsieh, Jun-Yu Yang, Zhi-Gang Sun (inventors)
  • Assignee: Realtek Semiconductor Corporation
  • Correspondent: not displayed in the indexed record; the correspondent of record for reel 036462/0935 would need to be read from the USPTO Assignment Center directly. Recurrence analysis is therefore N/A — there is no second link in the chain to compare against.
  • Context: standard pre-grant assignment from employee-inventors to their employer; the assignee named on the issued patent is the same entity, so no ownership change occurred at or after issuance.

Because no post-issuance assignments are recorded, Realtek Semiconductor Corporation remains the owner of record. (The 2015-08-31 recording predates the 2017-04-11 grant and simply perfects Realtek's title as original assignee.)

Timeline diagram

timeline
    title Ownership of US 9619864
    2015 : Filed by Realtek Semiconductor
         : Assigned by inventors to Realtek
    2017 : Patent issued
    2020 : Maintenance fee paid by Realtek

NPE / troll-pattern signals

1. Shell-entity transfer — not present. No transfer from Realtek to any IP-holding LLC, single-member entity, or licensing vehicle appears in the record. The only assignment runs to an operating company, not away from one.
2. Known asserter in the chain — not present. The only assignee ever named is Realtek Semiconductor Corp., a fabless operating company. No entity from the Acacia / Marathon / IV / Conversant / Round Rock / Spangenberg / MPHJ / etc. lists appears anywhere in the chain. The prior litigation review also found no assertion of this patent.
3. Repeat correspondent across the chain — not present / unclear. There is only one recorded conveyance (reel 036462/0935), so there is no chain in which a correspondent could recur. The correspondent's identity for that single entry was not retrievable from the indexed sources; flag for direct Assignment Center verification if needed.
4. Cascading transfers — not present. No consecutive assignments, no chained LLCs, no multi-hop transfers in any time window.
5. Pre-litigation transfer — not present. No infringement litigation naming this patent was found, and no assignment activity exists that could be timed to a filing.
6. Bankruptcy fire-sale — not present. Realtek is solvent and operating; no Chapter 7/11 or distress sale is associated with this patent.
7. Privateering — not present. No transfer of the patent to an NPE for assertion against Realtek's competitors is recorded, and no SEC/8-K disclosure of such an arrangement surfaced.
8. Defensive aggregator (anti-NPE) — not present. The chain does not terminate at RPX, AST, LOT, Unified Patents, or OIN; it terminates (and begins) at Realtek.

Verdict

Insufficient data — per the rubric, this is the "only the original assignment" case: the sole recorded conveyance is the 2015-08-31 inventor-to-Realtek assignment (reel 036462/0935), and the USPTO-derived legal-events feed shows no post-issuance assignments, so Realtek Semiconductor Corp. is the owner of record and no NPE appears in the chain. Every available signal points to ordinary operating-company ownership with no troll pattern, but I could not directly query the USPTO Assignment Center this session to exclude an unindexed later record, and the corresponding correspondent-of-record data for the one filing was not captured. Verify at https://assignmentcenter.uspto.gov/ (search patent number 9619864) or https://assignment.uspto.gov/patent/index.html to confirm the absence of later conveyances.

I'll verify the patent and each of its cited references against USPTO/public records before mapping them to claims.

§ 102 Prior-Art Mapping for US 9,619,864 B2

Verification of the target patent

As instructed, I confirmed that patent number 9619864 resolves to the correct, specific document — US 9,619,864 B2, "Image processing apparatus and method for increasing sharpness of images," Realtek Semiconductor Corp., filed 2015-08-30, granted 2017-04-11, CN priority 2015-03-25. No adjacent/similar numbers (e.g., 9619865, 9619863, 9619864A1) were substituted; the CN106161875B / TW I561068B counterparts are the same family and are not separate prior art against this document.

Note on scope: The reference set below is exactly the 13 entries on the face of the '864 patent (11 unique families, since Samsung US20020067862A1/US7130483B2 and Panasonic US20090232401A1/US7881549B2 are family duplicates), plus the 4 non-patent citations. Because the '864 effective filing date (2015-03-25) is after 2013-03-16, AIA § 102 governs.

A necessary threshold point for § 102

Anticipation under 35 U.S.C. § 102 requires a single reference that discloses every element of a claim, arranged as claimed. Reviewing all 13 cited references against the claim charts built in the prior sections, no cited reference — and no single non-patent citation — anticipates any of claims 1–20 in full. Every claim of '864 contains at least one of the following limiting elements that no single reference supplies:

• the two-path architecture (a translated/overshoot first-image path and an untranslated second-image path, both feeding one gain-scaled superposition);
• the overshoot high-pass masks with positive coefficients at the center and periphery and negative coefficients elsewhere (claims 7–8, 16–17, 20); and/or
• the non-adjacent (≥ 1-pixel-gap) neighbor-sum gain input (claims 1, 3–4, 10, 12–13).

Accordingly, the table below maps each reference to the individual claims it bears on and why it falls short of § 102, so its real role (§ 103) is clear. Where a reference is closest to a claim, I say so explicitly.

1. Reference-by-reference analysis

US 6,285,798 B1 — Automatic tone adjustment by contrast gain-control on edges

• Assignee / inventor: Eastman Kodak Company / Hsien-Che Lee
• Dates: filed 1998-12-11 (App. 09/210,107); priority 1998-07-06; granted 2001-09-04
• Description: Decomposes an input image into detail (HF) signals and coarse signals; generates contrast gain-control (CGC) signals by detecting coarse-scale edges; modifies the detail signals according to the contrast gain signals; and adds the modified detail signals to the coarse signals to reconstruct an output image. The stated purpose is dynamic-range compression with edge-banding suppression, not sharpening per se.
• § 102 relevance: Closest single reference on the "filtered HF band, gain-weighted, then recombined" skeleton. It bears on the generic concept underlying claims 1, 2, 10, 11 (multi-band decomposition + gain on the detail band + recombination).
• Why it does not anticipate: (i) Its gain is applied at coarse-scale edge locations for tone/banding control, not per-pixel detail gains from a conversion table lookup of a pixel-value sum; (ii) it has no translated/overshoot first-image path, no overshoot masks; (iii) no non-adjacent neighbor-sum gain input. No claim is fully disclosed. Role: § 103, not § 102.

US 2002/0067862 A1 — Method for enhancing a digital image while suppressing undershoots and overshoots

US 7,130,483 B2 — same family (granted version)

• Assignee / inventor: Samsung Electronics Co., Ltd.
• Dates: US 2002/0067862 A1 filed 2001-12-12, published 2002-06-06; US 7,130,483 B2 granted 2006-10-31
• Description: Canonical unsharp-masking enhancement g = f + β(m,n)·α·h(m,n), where h is the high-pass-filtered image and β(m,n) is an explicit per-pixel gain computed from local pixel-value statistics (left/right luminance differences fL, fR, their minimum d, and a Laplacian Δ). β is a "shoot suppressing" gain whose stated purpose is to attenuate enhancement at edges so overshoot/undershoot are suppressed. The gain-scaled HF signal is added back to the input.
• § 102 relevance: This is the closest single reference to the gain-scaled HF add-back concept. It bears on claims 2/11 (blend HF component into the input), and on the generic gain+HF architecture of claims 1/10.
• Why it does not anticipate: A LUT ("conversion table") is not disclosed (β is an on-the-fly polynomial); the gain is a function of immediate-neighbor differences, not a sum over neighbors spaced ≥ 1 pixel away; there is one HPF path only, no translated/overshoot path, and the reference teaches suppression of the very overshoot '864's first path deliberately creates. No anticipation of any claim; strongest § 103 anchor against claims 1/2/10/11.

US 6,611,627 B1 — Digital image processing method for edge shaping

• Assignee: Eastman Kodak Company
• Dates: filed 2000-04-24; granted 2003-08-26
• Description: Method for shaping edges in a digital image, manipulating edge structure/transitions. Operates on edge statistics/transitions, not on overshoot-generating filter masks, and does not use a two-path gain-modulated fusion.
• § 102 relevance: Peripheral background on edge shaping/contrast. Bears at most tangentially on the "edge/detail enhancement" context of claim 1/10.
• Why it does not anticipate: No detail-gain conversion table, no translated HF path, no overshoot masks, no non-adjacent neighbor sum. Role: background only.

US 6,714,688 B1 — Method and apparatus for enhancing the edge contrast of an interpolated digital image

• Assignee / inventor: Eastman Kodak Company
• Dates: filed 2000-11-10; granted 2004-03-30
• Description: Supplies the exact preamble context of '864 — enhancing an interpolated digital image — using an "edge shaper," a sharpening filter adaptive to image morphology with parameter values derived from local statistics over a window around each pixel; bounds the output by the local max/min to reduce overshoot/undershoot and ringing.
• § 102 relevance: Closest reference on "gain/parameters computed from a neighborhood of pixel values, applied to an interpolated image." Bears on the contextual elements of claims 1, 10.
• Why it does not anticipate: It uses window statistics to shape the local tone scale, not a stored conversion table addressed by a pixel-value sum; its design goal is the opposite of overshoot creation; and it discloses no two-path translated/untranslated superposition and no non-adjacent neighbor sum. No anticipation; § 103 support (especially for claim 1/10 context).

US 6,735,330 B1 — Automatic digital radiographic bright light

• Assignee: Eastman Kodak Company
• Dates: filed 2000-10-17; granted 2004-05-11
• Description: Radiographic display/tone processing (bright-light rendering of digital radiographs) — a different technical field (medical imaging dynamic-range rendering).
• § 102 relevance: None material. Its only overlap is generic tone/gain processing.
• Why it does not anticipate: It does not disclose HF translation, overshoot masks, conversion-table detail gains, non-adjacent neighbor sums, or two-path weighted superposition. Role: remote background; not a § 102 or strong § 103 reference.

US 6,909,813 B2 — Contour correcting circuit and contour correcting method

• Assignee / inventor: Sanyo Electric Co., Ltd. / Amano et al.
• Dates: filed 2001-05-22; granted 2005-06-21
• Description: A contour (HF) component is extracted from the signal gradient around a pixel and multiplied by a per-pixel correction gain that varies with the level difference between the center pixel and adjacent pixels, then added back (with min/max clipping). A gain-calculating circuit (Fig. 2) produces the gain from signal levels.
• § 102 relevance: Independent corroboration that per-pixel gain-scaled HF add-back was mature, conventional architecture. Bears on claims 1/2/10/11 generically.
• Why it does not anticipate: Gain is derived from adjacent-pixel level differences (not a non-adjacent neighbor sum, not a stored conversion table); single contour path (no translated + untranslated dual path); no overshoot masks. No anticipation; § 103 support for the generic claim skeleton.

US 6,965,406 B1 — Image processor and image processing method

• Assignee: Sony Corporation
• Dates: filed 1999-04-16; granted 2005-11-15
• Description: General image processing (per the citation, broad image-processing apparatus/method). Background reference.
• § 102 relevance: Low. The full text was not retrieved in my searches, so I cannot map element-by-element with confidence (see caveats in the prior Obviousness section).
• Why it does not anticipate: On the record available, nothing suggests disclosure of the '864-specific combination (overshoot translation path + conversion-table detail gains + non-adjacent neighbor sum). Flagged: full-text verification required before any § 102 reliance.

CN 1917577 A — Method of reducing noise for combined images

• Assignee: 上海大学 (Shanghai University)
• Dates: filed 2006-09-01; published 2007-02-21
• Description: Noise reduction for combined (multi-image) composites — i.e., it addresses the multi-frame paradigm that '864's Background distinguishes and rejects.
• § 102 relevance: Very low. It is directed to noise reduction across combined frames, not single-image HF detail injection.
• Why it does not anticipate: Different problem (denoising multi-frame composites), no detail-gain conversion table, no overshoot path, no non-adjacent neighbor sum, no dual-path weighted superposition. Role: background only.

US 2009/0232401 A1 — Visual processing apparatus, display apparatus, visual processing method, program, and integrated circuit

US 7,881,549 B2 — same family (granted version)

• Assignee / inventor: Panasonic Corporation / Haruo Yamashita (per the record shown)
• Dates: US 2009/0232401 A1 filed 2005-10-12, published 2009-09-17; US 7,881,549 B2 granted 2011-02-01
• Description: Human-visual-system (HVS)–based visual processing, built on the proposition that controlled high-frequency detail/contrast improves perceived image quality — the same psychophysical premise '864 relies on.
• § 102 relevance: The nearest cited art to the perceptual rationale behind the periodic conversion table (claims 5/14).
• Why it does not anticipate: It does not teach a periodic, continuous gain LUT for detail enhancement, nor the non-adjacent neighbor sum, nor the overshoot path. No anticipation; its role is § 103 motivation (perceptual benefit of HF detail).

TW 201001334 A — Adjustment method of color tone for digital image and electronic apparatus thereof

• Assignee: Altek Corp
• Dates: filed 2008-06-20; published 2010-01-01
• Description: Color-tone adjustment for digital images.
• § 102 relevance: Low. Different processing goal (tone/color mapping vs. HF sharpness injection).
• Why it does not anticipate: No translated HF path, no overshoot masks, no conversion-table detail gains tied to a non-adjacent neighbor sum. Role: background only.

US 9,123,140 B1 — Recovering details in single frame super resolution images

• Assignee: Pixelworks, Inc.
• Dates: filed 2013-09-25; granted 2015-09-01
• § 102 availability note: Because it issued after the '864 priority date (2015-03-25), it is not § 102(a)(1) art. It is available only under § 102(a)(2) as a U.S. patent effectively filed (2013-09-25) before the '864 effective filing date.
• Description: Recovers detail by feature-based copying from the 1× layer into an initial SR result, using directional feature maxima (horizontal/vertical/diagonal/anti-diagonal), patch-SAD matching, and inter/intra-difference fusion. Solves the identical single-frame upscaling problem stated in '864's Background.
• § 102 relevance: Confirms that single-frame detail injection into an upscaled image and directional max-selection of filtered features were known. Bears on the context and motivation for claims 1/10/19.
• Why it does not anticipate: Its mechanism is patch/feature copying, not overshoot high-pass filtering with positive-periphery masks; there is no conversion-table gain, no non-adjacent neighbor sum, and the "max" selection is over feature matches, not over overshoot-filter translation images. No anticipation; strong § 103 motivation for the single-frame problem.

2. Non-patent citations (all § 102(a)(1) printed publications)

None of the four non-patent publications discloses any structural element of claims 1–20, and none is a § 102 reference against '864.

3. Consolidated § 102 conclusion

Bottom line: On the cited record, no reference is an anticipatory ( § 102 ) reference. Every cited patent and publication supplies at most individual elements of the '864 claims — the gain-scaled HF add-back skeleton (Samsung, Sanyo), multi-band detail gain control (Kodak '798), the interpolated-image context (Kodak '688), directional feature fusion (Pixelworks '140), and the single-frame-SR problem framing (Pixelworks '140 and the four non-patent papers). The claim-distinguishing features — the overshoot high-pass "translation" path with positive center-and-periphery masks (claims 7–8, 16–17, 19–20) and the non-adjacent-neighbor sum gain input (claims 1, 3–4, 10, 12–13) — are absent from every cited reference. These references therefore support a § 103 obviousness attack (best primary: Samsung '483/‘862, in view of Kodak '798, Kodak '688, and Pixelworks '140), not a § 102 anticipation.

Caveats (carry forward)

• I could not fully retrieve the specifications of US 6,965,406 (Sony), US 6,735,330, CN 1917577 A, and TW 201001334 A in this session; the "no anticipation / background" characterizations for those four rest on their titles/abstracts and the examiner's citation context and should be confirmed against full text before being relied on in a petition.
• The prosecution history was not reviewed; the examiner's stated reason for allowance may identify a point of novelty beyond those I have inferred.
• Because US 9,123,140 B1 issued after the '864 priority date, it is § 102(a)(2) art only; do not cite it as § 102(a)(1) or as "prior art by publication" on the '864 date.
• The legal-status/date data were confirmed against Google Patents, FreePatentsOnline, Justia, and the USPTO image PDF; interpret all identifiers literally, and note that the two Samsung entries and the two Panasonic entries are family duplicates, not independent references.

§ 103 Obviousness Analysis — U.S. Patent No. 9,619,864 (US9619864B2)

Scope note: This analysis builds on the Prior Art section previously generated (the 13 examiner-cited references verified there) and additional confirmation obtained this session on US7130483B2 (Samsung) and US9123140B1 (Pixelworks). Because the patent's effective filing date (Mar. 25, 2015) post-dates Mar. 16, 2013, AIA § 103 governs. Every reference below is § 102(a)(1) art except US9123140B1, which is available only under § 102(a)(2) (filed Sep. 25, 2013; issued Sep. 1, 2015 — after the '864 effective filing date).

1. The claims in issue and where their real patentable weight sits

Claim 1 (apparatus) / Claim 10 (method) require four functional elements:

• (a) extract and translate first HF components → first image;
• (b) extract second HF components → second image;
• (c) generate per-pixel detail gains using a conversion table and pixel values, where the gain of a given pixel corresponds to the pixel-value sum of that pixel plus neighboring pixels separated from it by at least one pixel;
• (d) weighted superposition of first + second images → output HF component, scaled by the gains.

Claim 19 (method) drops the neighbor-sum limitation but affirmatively requires (a) generating the first image by filtering the input with at least two overshoot high-pass filters into plural high-frequency translation images and selecting pixels with the highest absolute pixel values across them; claim 20 adds the mask geometry (positive coefficients at center and periphery; negative elsewhere).

Realistically, the features a § 103 challenger must supply from the art are: (i) the two-path arrangement (a translated/overshoot-filtered path plus an ordinary high-pass path feeding a single gain-scaled output); (ii) the overshoot-filter mask geometry (claims 8, 17, 20); (iii) gains produced by a periodic continuous conversion table (claims 5, 14); and (iv) gain inputs computed as sums over non-adjacent (≥ 1-pixel-gap) neighbors (claims 1, 3–4, 10, 12–13). The "conversion table" itself and the "multiply a weighted HF combination by per-pixel gains and add back" skeleton are the weakest features — those were decades old by 2015.

2. Person of ordinary skill in the art (PHOSITA)

A PHOSITA would be a designer of digital image/video enhancement pipelines (image-processing engineer or IC/display-scaler architect) with a B.S./M.S. in electrical engineering or computer science and 2–4 years of experience implementing filtering, unsharp masking, edge enhancement, and super-resolution on embedded display/video hardware (typical of the Realtek/Pixelworks/Samsung/Sanyo/Kodak space). The person would be fluent in: FIR filtering and mask design; multi-band (detail/coarse) decomposition; per-pixel or per-region gain control; LUT-based point operations; and the memory/cost constraints of real-time display processors. That profile matters because several "missing" features are implementation choices (LUT vs. formula; window geometry) that this person would treat as routine.

3. Strongest combinations and element mapping

Combination A — Samsung US7130483B2 / US20020067862A1 (primary) + Kodak US6285798B1 (+ optional Pixelworks US9123140B1 for problem framing)

Best against: claims 1, 2, 10, 11, and the generic gain-scaled-HF architecture; also claims 6, 15.

What Samsung teaches (verified this session from the '483 text): the canonical unsharp-masking enhancement g = f + β(m,n)·α·h(m,n), where h is a high-pass-filtered version of the input and β(m,n) is an explicit per-pixel gain computed from local pixel-value statistics — left/right (or up/down) luminance differences fL, fR, their minimum d, and a Laplacian Δ(m,n) = f(m,n+1) − 2f(m,n) + f(m,n−1) — via the function β(m,n) = 1 − (1−x(m,n))ᵖ·(y(m,n))ᑫ. The gain-scaled HF signal is added back to the input. That is a direct disclosure of elements (b) (extract HF by high-pass filtering), the generic form of (c) (per-pixel gain derived from surrounding pixel values), (d) in part (HF content scaled by a per-pixel gain and recombined), and claims 2/11 (blend the HF component into the input).

What Kodak '798 adds: a filter-bank decomposition into detail (HF) and coarse signals, with contrast gain-control (CGC) signals computed from edge structure and applied to the detail band before reconstruction — i.e., a second, independent filtering path whose output is gain-modulated and then combined. That supplies the notion of multiple, separately filtered HF signals being gain-weighted and superposed, closing part of the gap to claim 1's two-path weighted superposition.

What Pixelworks '140 adds (problem framing and motivation): it addresses the identical problem stated in '864's Background — producing a detail-enriched high-resolution frame from a single low-resolution/interpolated input, without multi-frame memory cost or a learning database. Its mechanism (verified from the Justia text: feature-based copying of detail from the 1× layer into the initial SR result, directional feature maxima in horizontal/vertical/diagonal/anti-diagonal, patch-SAD and inter/intra-difference fusion) confirms that single-frame detail injection into an upscaled image was a known, active problem and that directionally filtered features and per-pixel/per-block blending decisions were conventional.

Motivation to combine (KSR): A designer building a single-frame SR/sharpen block (the Pixelworks problem) would naturally start from classic unsharp masking with content-adaptive gain (Samsung) and multi-band gain control (Kodak '798) because those are the standard tools for exactly the operation '864 performs — taking filtered detail and deciding, pixel-by-pixel, how much of it to add. Combining Samsung's per-pixel β with Kodak's multi-band detail handling yields the claimed "weighted superposition of two HF images × per-pixel detail gains → output HF component" as the predictable union of familiar elements (KSR, 550 U.S. at 416). Implementing the β(x,y) gain function as a lookup table rather than an on-the-fly polynomial would be an obvious implementation choice to a PHOSITA (LUTs are the default hardware realization of pointwise gain functions in display ICs) — which substantially deflates the patentable weight of "stores a conversion table" in claims 1/10 and the "0 to N" range of claim 6/15.

Where Combination A fails: it does not account for (i) a first HF path that is translated by overshoot filters with positive center and periphery coefficients — indeed Samsung and Sanyo are expressly directed to suppressing overshoots, i.e., a possible teaching away from creating them (see § 5); (ii) the two-path combination as claimed (a translated path plus an untranslated path both feeding the gain stage — Kodak '798's two bands are not a translated/untranslated pair); and (iii) the non-adjacent-neighbor sum used to derive the gain input.

Combination B — Sanyo US6909813B2 + Samsung US7130483B2 (+ Kodak US6714688B1)

Best against: claims 1, 2, 10, 11 and the per-pixel gain-with-content-adaptivity concept.

Sanyo '813 teaches a contour-correcting circuit in which a contour (HF) component is extracted from the signal gradient around a pixel and multiplied by a per-pixel correction gain that varies with the level difference between the center pixel and adjacent pixels, then added back with clipping. This is independent corroboration that per-pixel gain-scaled HF add-back was a mature, conventional architecture by 2005 — making the generic claim 1/10 combination (extract HF → per-pixel gain → gain-scaled combination) look like routine engineering rather than invention. Kodak '688 supplies the exact preamble context of '864 (enhancing the edge contrast of an interpolated digital image) and teaches deriving pixel-modification parameters from local statistics over a window around each pixel — conceptually the same "gain from a neighborhood of pixel values" idea.

Motivation: Sanyo and Samsung solve the same problem (controlling how much HF detail is added per pixel) by the same means (content-adaptive per-pixel gain). Kodak '688 establishes that applying such techniques to interpolated images — the precise input of '864 — was known. Combining them yields the claim 1 skeleton with strong motivation and no surprises.

Where Combination B fails: same three gaps as Combination A (overshoot/translation path, dual-path superposition, non-adjacent-neighbor gain input). Sanyo and Samsung are both anti-overshoot; neither suggests deliberately synthesizing translated detail with positive-periphery masks.

Combination C — for the overshoot high-pass path (claims 7–8, 16–17, 19–20)

Weakest case on the cited art alone.

Claim 19/20's "overshoot high-pass filters" with positive center-and-periphery coefficients and negative coefficients in between are the most technically distinctive feature. No examiner-cited reference on the face of '864 discloses such a filter:

• Samsung '483 and Sanyo '813 teach gain functions whose entire purpose is to attenuate enhancement at edges to prevent overshoot/undershoot — the opposite sign of intent from '864's masks, which produce overshoot ringing deliberately.
• Kodak '798 (edge-shaping) and Kodak '688 (edge shaping of interpolated images) operate on edge statistics/transitions, not on overshoot-generating filter masks.
• Pixelworks '140 recovers detail by feature/patch copying, not by overshoot filtering.
• US6965406B1 (Sony), CN1917577A, TW201001334A, and US6735330B1 are remote background on the record before me.

A § 103 challenger could still argue: the "translate ... first high-frequency components" step is just a peaking/LTI (luminance transient improvement) operation — deliberately adding pre-shoot/post-shoot to edges to create the illusion of higher resolution — which was well known in TV/display contour-correction circuits predating 2015. The two overshoot masks 360A/360B are simply horizontal and vertical peaking kernels, and the "select pixels with the highest absolute pixel values" fusion of directional filter outputs is a standard directional-max detail-fusion technique (a concept visible even in Pixelworks '140's "select the direction with the maximum absolute feature"). However, that argument requires importing art outside the examiner-cited list (e.g., peaking/LTI patents from the 1980s–2000s). On the cited references alone, claims 19–20 — and dependent claims 7–8, 16–17 — are the most resistant to a § 103 challenge, because none of the cited references discloses an overshoot-producing mask, and Samsung/Sanyo arguably teach away from it.

Combination D — for the periodic-continuous conversion table (claims 5, 14) and the non-adjacent-neighbor sum (claims 1, 3–4, 10, 12–13)

Weakest link in the challenger's map — genuinely inventive-looking features.

• Periodic continuous table (claims 5/14). The specification's own rationale is that a periodic function gives bright and dark pixels "approximately equal opportunity" to receive high gains — i.e., it injects controlled randomness into the gain field to improve perceived texture/contrast, while continuity keeps video temporally stable. The nearest cited art is Panasonic US20090232401A1/US7881549B2, which is grounded in human-visual-system models and the proposition that controlled high-frequency detail/noise improves perceived sharpness. A challenger would argue: a PHOSITA aware of the perceptual benefit of moderate random HF detail (Panasonic) and needing a gain function that varies between adjacent pixels would choose a periodic, continuous mapping to avoid DC bias toward bright or dark regions — an "obvious to try" design choice among a small number of known periodic functions (triangle wave, sine, semicircle). But there is no cited reference teaching a periodic gain LUT for detail enhancement, and the specific benefit (equal opportunity across tones + temporal stability under noise) is articulated only in '864. This is a plausible but moderately weak obviousness case.
• Non-adjacent neighbor sum (claims 1, 3–4, 10, 12–13). The claimed gain input — a sum of the center pixel's value with pixels separated by at least one pixel (claim 4 concretizes: the up/down/left/right pixels two steps away) — is designed so that the neighbor sets of adjacent center pixels do not overlap, maximizing gain differences between neighboring pixels and thus local contrast. No cited reference teaches a gain computed from a sparse, non-overlapping ring sum; Sanyo and Samsung use immediate-neighbor differences, and Kodak '688 uses window statistics but not for this decorrelation purpose. This is the single feature most likely to carry claim 1/10 over the art, and it is essentially undisclosed in the cited record. To knock it out, a challenger would need additional art on gain fields derived from dilated/eroded or staggered neighborhoods (e.g., dithering or checkerboard-pattern gain modulation), which is not in the cited list.

4. Element-level summary of the best § 103 case

Bottom line on vulnerability ranking (on the cited-art record):

1. Most exposed: the generic combination of claims 1, 2, 6, 10, 11, 15 to the extent the claim scope is construed broadly — Samsung '483 + Kodak '798 (or Sanyo '813) renders the "per-pixel gains from a conversion table × HF content, added back" architecture highly vulnerable, since a LUT implementation of Samsung's β function is an obvious implementation detail and the two-HF-band gain-controlled recombination is in Kodak '798. A determined challenger's best single theory is Samsung '483 as primary, in view of Kodak '798 and Kodak '688, with Pixelworks '140 supplying the single-frame-SR motivation and the interpolated-input context.
2. Intermediate: claims 5/14 (periodic table) — attackable via Panasonic '401 plus the "obvious design choice" rationale, but no direct teaching.
3. Most robust: claims 7–8, 16–17, 19–20 (overshoot-filter path and mask geometry) and claims 3–4, 12–13 (non-adjacent neighbor sum). On the examiner-cited references alone, these features are undisclosed and Samsung/Sanyo's anti-overshoot purpose arguably teaches away from the '864 overshoot-generating masks.

5. Motivation-to-combine analysis (KSR / Graham)

• Graham factor 1 — scope and content of prior art: As mapped above, Samsung '483, Sanyo '813, Kodak '798/'688, and Pixelworks '140 collectively occupy the exact technical space of '864: adding content-adaptive, gain-controlled high-frequency detail to an (interpolated) image with a single input frame. The differences between the claimed invention and this art are narrow (dual-path overshoot fusion; LUT periodicity; sparse-neighborhood gain inputs), not conceptual.
• Graham factor 2 — differences: Detailed above; the meaningful differences are the overshoot mask geometry, the dual-path architecture, and the non-adjacent-sum/periodic-table gain generation.
• Graham factor 3 — level of ordinary skill: High in this field (display/video IC design); a PHOSITA would routinely substitute LUTs for gain formulas, add directional filters, and try alternative window geometries — all of which cuts against patentability of the "implementation-detail" features.
• Graham factor 4 — secondary considerations: None of record. No commercial-success, long-felt-need, copying, or unexpected-results evidence appears in the file. Realtek has paid maintenance fees, but maintenance alone is not a secondary indicator of non-obviousness. The absence of any licensing or litigation activity (see the Litigation/PTAB sections) means there is no marketplace evidence of non-obviousness either.

KSR-style motivations a challenger would invoke:

1. Design need / known problem: '864's own Background concedes the prior techniques (multi-frame SR; learning-based SR) and their drawbacks. Pixelworks '140 proves that solving "single-frame detail recovery" was a recognized design need in the same timeframe. A PHOSITA seeking a low-cost single-frame solution had a direct incentive to combine existing unsharp-masking-with-gain (Samsung/Sanyo) into the single-frame framework.
2. Predictable combination of known elements: Gain-scaled HF add-back (Samsung/Sanyo), multi-band detail gain control (Kodak '798), and application to interpolated images (Kodak '688) were all known; combining them yields the claimed result with predictable success — the hallmark of KSR obviousness.
3. Obvious implementation substitution: A "conversion table" is the standard hardware realization of a pointwise gain function (Samsung's β or Sanyo's gain); substituting a LUT for a formula is the kind of "ordinary innovation" KSR says is not patentable.
4. Market pressure / known alternatives: Display scaler and TV-chip vendors (all four assignees here — Samsung, Sanyo, Kodak, Pixelworks, Realtek) competed on perceived sharpness; the incentive to add controllable per-pixel detail while managing edge artifacts (shoot suppression) was industry-wide, providing motivation to arrive at the claimed gain-modulated detail-injection architecture.

Countervailing arguments (why the examiner allowed the claims, and why a challenger still faces hurdles):

• Teaching away / design-choice tension: Samsung and Sanyo exist to suppress overshoot; '864's first path affirmatively creates translated (overshoot) HF content to push the output's maximum spatial frequency beyond the input's. A challenger must explain why a PHOSITA would deliberately add the very artifact the closest prior art teaches to avoid — a significant hurdle for claims 7–8/16–17/19–20.
• No single reference or obvious pair discloses the dual-path fusion: Even combining Samsung (one HF path + gain) with Kodak '798 (multi-band) does not produce the claimed "translated first image + untranslated second image, both gain-scaled together" structure; the combination requires hindsight reconstruction of '864's specific two-filter topology.
• The sparse-neighborhood gain input is unexplained by the cited art: The non-overlap/decorrelation benefit (spec ¶ [0026]) is not suggested anywhere in the record; claiming it is "obvious to try" is weaker post-KSR than a showing of a finite, predictable set of solutions, and the record contains no art pointing to a two-pixel-radius gain window.
• Claim 19/20's mask geometry (positive center and periphery, negative in between) is a concrete, unusual filter shape with no counterpart in the cited masks (standard HPFs have negative periphery; the cited art's whole thrust is to avoid the ringing such masks produce).

6. Caveats and open items

• Full-text verification gaps: The complete specifications of US9123140B1 (Pixelworks), US6965406B1, US6735330B1, CN1917577A, TW201001334A, and US20090232401A1/US7881549B2 were not fully reviewed in the prior-art search run. Element-level conclusions against those references — particularly whether Pixelworks '140 anywhere discloses gain-LUT or overshoot-filter processing — remain provisional and should be confirmed against full text before relying on them in a petition or litigation.
• No file-wrapper review: The prosecution history (why the claims were allowed over this exact art) was not reviewed. The examiner's reasons for allowance could materially inform the obviousness analysis, particularly which features the examiner regarded as the point of novelty.
• Outside art: A complete § 103 case against the overshoot-filter claims (19–20) and possibly the periodic-table claims (5/14) would need art not on the face of the patent — peaking/LTI contour-enhancement patents (for overshoot-generating filters), and dithering/gain-modulation art (for the non-adjacent and periodic gain features). The examiner-cited list is the right starting map, but it is not sufficient to prove the strongest claims obvious.

Overall: The claims of US9619864B2 are partially vulnerable under § 103 on the cited art. The broad, gain-scaled, HF-detail-injection concept of claims 1/2/10/11 (and the LUT/0–N implementation details of claims 6/15) is the most attackable — best combination: Samsung US7130483B2/US20020067862A1 in view of Kodak US6285798B1, Kodak US6714688B1, and Pixelworks US9123140B1, with Pixelworks supplying the single-frame-SR motivation. The features that most plausibly carry the patent — the overshoot high-pass "translation" path with positive center-and-periphery masks (claims 7–8, 16–17, 19–20) and the sparse, non-adjacent-neighbor-sum gain input (claims 1, 3–4, 10, 12–13) — are not disclosed in the cited record, are not rendered obvious by any combination of the cited references without hindsight, and for the overshoot feature the closest cited art arguably teaches away.