Patent Drafting Analysis of Insurance Services Office’s Computer Vision Object Detection with Reinforcement Learning | US 12,067,644 B2
Patent Drafting Analysis of Insurance Services Office's Computer Vision Object Detection with Reinforcement Learning | US 12,067,644 B2
A structural and strategic analysis of US 12,067,644 B2 covering claim architecture, drafting quality signals, critical gaps, and prosecution positioning for Insurance Services Office's reinforcement learning-based multi-object detection system.
Spec Words
7,200
Across 6 sections
Draft now ↗
Total Claims
22
2 independent · 20 dependent
Draft now ↗
Figure Sheets
14
Prior art diagrams, system architecture, flowcharts, detection results
Draft now ↗
Published by PatSnap Insights Team · · 12 min read Verified by PatSnap Eureka Data
Overview
Structural Overview
The detailed description dominates at approximately 54% of total specification words (~3,900 of ~7,200), with the background section providing substantial prior art context at roughly 27% — an unusually high background ratio reflecting the inventor's effort to distinguish from tree-structured and hierarchical RL approaches. The claim set comprises 22 claims across 2 independent claims (Claim 1 as a system claim and Claim 12 as a method claim), yielding a 10:1 dependent-to-independent ratio that is notably high for the computer vision IPC class. The 14 drawing sheets blend 6 prior art figures with 8 inventive figures covering process flows, detection output images, performance graphs, and the hardware system architecture.
Section Word Distribution
↗ Click bars to exploreFigure Inventory — 14 Sheets
| Figure | Description | Role |
|---|---|---|
| FIG. 1 | Prior art hierarchical object detection showing iterative bounding box magnification steps applied to a single object image region.Search in Eureka ↗ | Other |
| FIG. 2 | Prior art deep reinforcement learning architecture for hierarchical object detection showing Image-Zoom and Pool45-Crops pipeline with Q-network layers.Search in Eureka ↗ | Other |
| FIG. 3 | Prior art sequential multi-object localization processing using VGG-16 CNN with ROI pooling layer, 4096-d image feature, RoI feature, and 650-d action history.Search in Eureka ↗ | Other |
| FIG. 4 | Prior art tree-structured reinforcement learning showing local translation and scaling action branches applied across image region candidates.Search in Eureka ↗ | Other |
| FIG. 5 | Prior art example detection results showing bounding boxes on face, body, and tool images across multiple zoom levels and scales.Search in Eureka ↗ | Other |
| FIG. 6 | Prior art tree-structured RL showing scaling actions (zoom-in, zoom-out) and local translation actions (up, down, left, right) applied to bounding boxes.Search in Eureka ↗ | Other |
| FIG. 7A | Inventive hierarchical object detection showing magnification actions (2, 2a) — the RL agent shifting and magnifying a bounding box within the image region.Search in Eureka ↗ | Key embodiment |
| FIG. 7B | Inventive top-down actions (4, 4a) showing the RL agent shifting a bounding box by compressing or expanding left, right, top, and bottom sides.Search in Eureka ↗ | Key embodiment |
| FIG. 7C | Inventive splitting actions (6, 6a, 6b) showing the RL agent splitting a bounding box into two sub-boxes when multiple objects are detected in one region.Search in Eureka ↗ | Claim support |
| FIG. 7D | Flowchart (method 10) showing the overall three-step process: set RL agent parameters (12), detect target objects (14), determine bounding boxes (16).Search in Eureka ↗ | Flow diagram |
| FIG. 8 | Sample VOC dataset image showing multiple detected objects (persons, dog, sheep) with bounding boxes and class labels output by the inventive system.Search in Eureka ↗ | Claim support |
| FIG. 9 | Performance graphs comparing training steps and test steps across four action set configurations (a-d): hierarchical, cut, fastcut, and mixed actions.Search in Eureka ↗ | Other |
| FIG. 10 | Performance comparison graphs showing RL agent results with and without history information at IoU thresholds a0 and a2, demonstrating the role of history vectors.Search in Eureka ↗ | Other |
| FIG. 11 | Performance graphs comparing system results under different terminal reward configurations ({3,−3} and {1,−1}) at IoU thresholds a0 and a2.Search in Eureka ↗ | Other |
| FIG. 12 | Performance graphs showing system results at different IoU thresholds (0.5 and 0.7) across epoch training and test runs at IoU levels a0 and a2.Search in Eureka ↗ | Other |
| FIG. 13 | System architecture diagram (102) showing computer vision system components: storage device (104), computer vision software code (106), network interface (108), bus (110), CPU/microprocessor (112), RAM (114), and input device (116).Search in Eureka ↗ | System architecture |
Analysis powered by PatSnap Eureka. Patent text and figures publicly available from USPTO. Draft a Similar Patent
Claims
Claim Architecture Analysis
The patent contains 2 independent claims: Claim 1 (computer vision system/apparatus) and Claim 12 (computer vision method), yielding a 10:1 dependent-to-independent ratio — substantially above the software/AI industry norm of 4–8:1. The dual independent claim structure provides apparatus and method coverage but notably omits a computer-readable medium (CRM) claim type, leaving a significant design-around vector. The dependent claims (2–11 depending from Claim 1; 13–22 depending from Claim 12) largely mirror each other in parallel, adding individual functional limitations such as bounding box splitting (Claims 6/17), aspect ratio selection (Claims 8/19), and trigger action training (Claims 10/21).
Core inventive concept: The claims solve the problem of duplicative object detection when a reinforcement learning agent fails to recognize previously detected objects by configuring the processor to perform reinforcement learning 'on a portion of the image appearing within the bounding box' and, upon receiving a positive terminal reward, either evaluating whether optimal rewards can be improved or determining a better fitting bounding box — as recited in the 'when the reinforcement learning agent receives a positive terminal reward' limitation of both Claims 1 and 12. The system further enables multi-object detection via bounding box splitting when 'more than one object in a region of the image' is detected, as recited in Claims 6 and 17.
Independent Claim Dissection
| Claim | Preamble | Transition | Key Body Elements |
|---|---|---|---|
| Claim 1 | A computer vision system for object detection with reinforcement learning | comprising | a memory storing at least one image; a processor in communication with the memory; processor setting a plurality of RL agent parameters; retrieving at least one image; detecting a target object based on RL agent parameters; determining a bounding box for detected target object; displaying bounding box on image; performing RL on portion of image within bounding box; when positive terminal reward received: (1) evaluating whether optimal number of rewards can be improved or (2) determining better fitting bounding boxSearch prior art ↗ |
| Claim 12 | A computer vision method for object detection with reinforcement learning | comprising the steps of | setting by a processor a plurality of RL agent parameters; retrieving by processor at least one image from memory; detecting by processor a target object in at least one image based on RL agent parameters; determining by processor a bounding box for detected target object; displaying the bounding box on image; performing RL on portion of image appearing within bounding box; when RL agent receives positive terminal reward: (1) evaluating whether optimal number of rewards can be improved or (2) determining better fitting bounding boxSearch prior art ↗ |
Claim Dependency Tree
1 Computer vision system — processor sets RL agent parameters, detects target object, determines bounding box, performs RL on bounding box region, responds to positive terminal rewardSearch Claim 1 prior art ↗
2 Adds: bounding box magnifies the portion of the imageSearch in Eureka ↗
3 Further: bounding box is shifted within image to magnify the portionSearch in Eureka ↗
4 Adds: bounding box is centered within the imageSearch in Eureka ↗
5 Adds: bounding box magnifies all of the imageSearch in Eureka ↗
6 Adds: bounding box is split into first and second bounding box when processor detects more than one object in a regionSearch in Eureka ↗
7 Further: first bounding box is processed if first and second bounding boxes overlapSearch in Eureka ↗
8 Adds: processor selects an aspect ratio and a size of a region of the imageSearch in Eureka ↗
9 Adds: processor opposes movement of the bounding box in response to a triggering actionSearch in Eureka ↗
10 Adds: processor trained with one or more of single class and multiple class object categoriesSearch in Eureka ↗
11 Adds: processor learns to initiate a trigger action for each class category if multiple class objects are detectedSearch in Eureka ↗
12 Computer vision method — processor sets RL agent parameters, retrieves image, detects target object, determines bounding box, performs RL on bounding box region, responds to positive terminal rewardSearch Claim 12 prior art ↗
13 Adds: magnifying the portion of the imageSearch in Eureka ↗
14 Further: shifting the bounding box within the image to magnify the portionSearch in Eureka ↗
15 Adds: centering the bounding box within the imageSearch in Eureka ↗
16 Adds: magnifying all of the image using the bounding boxSearch in Eureka ↗
17 Adds: splitting bounding box into first and second bounding box when more than one object is detected in a regionSearch in Eureka ↗
18 Further: processing first bounding box if first and second bounding boxes overlapSearch in Eureka ↗
19 Adds: selecting by processor an aspect ratio and a size of a regionSearch in Eureka ↗
20 Adds: opposing by processor movement of bounding box in response to triggering actionSearch in Eureka ↗
21 Adds: training processor with single class and multiple class object categoriesSearch in Eureka ↗
22 Adds: learning to initiate trigger action for each class category if multiple class objects detectedSearch in Eureka ↗
| Metric | This Application | Software / AI Industry Norm |
|---|---|---|
| Total claims | 22 | 15 – 25 |
| Independent claim count | 2 | 2 – 4 |
| Dependent : Independent ratio | 10.00 : 1 | 4 – 8 : 1 |
| Method claims present? | Yes — Claim 12 | Common |
| System / apparatus claims? | Yes — Claim 1 | Common |
Analysis powered by PatSnap Eureka. Patent text and figures publicly available from USPTO. Draft a Similar Patent
Drafting Quality
Drafting Quality Signals
The patent's strongest quality attribute is its comprehensive dependent claim fallback structure — Claims 2–11 and 13–22 systematically enumerate the bounding box manipulation actions (magnify, center, split, oppose) described in the detailed description, giving the examiner and any litigant a layered set of narrowing positions. However, the absence of a computer-readable medium (CRM) independent claim leaves a structural gap that a sophisticated competitor could exploit without infringing either Claim 1 or Claim 12.
✅
Antecedent Basis
All claim elements in dependent claims properly trace back to antecedents established in the independent claims. For example, Claim 2's reference to 'the bounding box' properly traces to 'determining a bounding box for the detected target object' in Claim 1. Claims 6 and 7 introduce 'a first bounding box and a second bounding box' in Claim 6 and then reference 'the first bounding box and the second bounding box' in Claim 7, maintaining clean antecedent basis throughout. No orphaned 'the' references were identified across all 22 claims.
✅
Spec–Claim Consistency
The key independent claim limitations map reliably to specific figures and paragraphs. The three-step process in Claims 1 and 12 (set parameters, detect object, determine bounding box) maps directly to FIG. 7D steps 12, 14, and 16. The splitting action recited in Claims 6 and 17 is illustrated in FIG. 7C showing bounding boxes 6a and 6b. The 'performing RL on a portion of the image appearing within the bounding box' limitation maps to the detailed description's discussion of magnification actions 2 shown in FIG. 7A. The hardware system claim in Claim 1 (memory, processor) is supported by FIG. 13 (components 104, 112, 114).
✅
Transition Word Usage
Both independent claims use 'comprising' — the correct open-ended transition for a software/AI system and method where additional components or steps should not negate infringement. Claim 12 uses 'comprising the steps of,' which is slightly narrower than a pure 'comprising' formulation but remains open-ended and does not limit the method to exactly the enumerated steps. No 'consisting of' or 'consisting essentially of' limitations appear anywhere in the claim set, which is appropriate for a computer vision system with inherent additional hardware and software components.
✅
§112(f) Means-Plus-Function Risk
No 'means for' or 'step for' language appears in any of the 22 claims. The claims use concrete structural and functional language tied to identified hardware — 'a processor' performing specific operations — which avoids §112(f) invocation under the post-Williamson presumption. The 'reinforcement learning agent' term is defined in the specification as a software construct executed by the processor rather than a structural 'means,' providing further insulation from §112(f) challenge. Overall §112(f) risk is low for this claim set.
⚠️
§101 Eligibility Risk
The claims face moderate Alice/Mayo §101 exposure because the core inventive concept — evaluating whether optimal rewards can be improved after a positive terminal reward — is an abstract mathematical optimization process. However, Claims 1 and 12 are tied to concrete hardware ('a memory storing at least one image' and 'a processor in communication with the memory') and the displayed bounding box output, which provides a meaningful §101 defense under the 'practical application' prong. The 'displaying the bounding box on the image' limitation in both Claims 1 and 12 also adds a concrete output step. A future examiner or challenger may still argue the processor limitations are generic, so dependent Claims 8 (aspect ratio selection) and 10/21 (training with class categories) add useful subject matter eligibility arguments.
⚠️
Dependent Claim Fallback Quality
The dependent claims mirror each other almost perfectly across the two independent claims — Claims 2–11 (depending from Claim 1) add the same limitations as Claims 13–22 (depending from Claim 12), respectively. While this creates enforcement symmetry, it does not add genuinely new technical fallback positions. Claims 6/17 (bounding box splitting) and Claims 9/20 (opposing movement in response to triggering action) are the most strategically valuable because they add the most distinctive technical limitations not present in the prior art. Claims 4/15 (centering the bounding box) and Claims 5/16 (magnifying all of the image) are weaker fallback positions as they likely exist in the prior art hierarchical approaches described in FIGS. 1–6.
⚠️
Abstract Quality
The abstract describes the system at a high functional level — 'the reinforcement learning agent determines a bounding box for each of the detected objects' — but does not identify the novel distinguishing mechanism: the post-terminal-reward evaluation step that prevents duplicative detections. An examiner reading only the abstract would understand this is an RL-based object detection system but would not identify the specific 'when the reinforcement learning agent receives a positive terminal reward, performing at least one of (1) evaluating whether an optimal number of rewards can be improved; or (2) determining a better fitting bounding box' limitation that survived prosecution and forms the core of the granted claims.
✅
Figure Support Quality
Every structural claim limitation in Claims 1 and 12 has corresponding figure support. The memory and processor system architecture is shown in FIG. 13; the three core method steps map to FIG. 7D; bounding box magnification maps to FIGS. 7A–7B; splitting into two bounding boxes maps to FIG. 7C (boxes 6a, 6b). The performance-related dependent claim concept of 'evaluating whether optimal number of rewards can be improved' is supported qualitatively by FIGS. 9–12 (performance graphs). The only minor gap is the 'opposing movement in response to a triggering action' limitation of Claims 9/20, which is described in the specification but not explicitly illustrated with a dedicated action diagram.
Analysis powered by PatSnap Eureka. Patent text and figures publicly available from USPTO. Draft a Similar Patent
Scorecard
Strategic Intent Scorecard
Multi-dimensional assessment of this application's patent strategy quality, based on claim structure, specification depth, and prosecution positioning.
Claim Breadth
3.2
Prosecution Defensibility
3.5
Spec–Claim Consistency
4
Dependent Claim Coverage
3
Claim Type Diversity
2.5
Figure Support Quality
3.8
Key observation: Spec–Claim Consistency scores highest (4.0/5.0) because every independent claim limitation maps to a specific figure — FIG. 7D for the three-step method, FIG. 7C for the splitting action, and FIG. 13 for the hardware system, providing robust written description support. Claim Type Diversity scores lowest (2.5/5.0) because the filing contains only system and method claims — the absence of a computer-readable medium (CRM) independent claim means a competitor who distributes the RL software on disc or via download without operating a system could avoid infringement of both independent claims. A stronger filing strategy would have added at least one CRM independent claim to claim the software itself as stored on a non-transitory medium.
See how your own draft compares — Open Eureka IP Drafting →
Critical Gaps
3 Critical Gaps in This Claim Set
A senior-attorney lens on the three highest-priority structural weaknesses — what each exposes in prosecution and litigation, and what a stronger filing would have done differently.
🔒
3 Critical Gaps in This Claim Set
See the full attorney-level analysis of what this application leaves unprotected — and how to draft it more defensively for your own filings.
Missing CRM independent claim Terminal reward condition narrows scope IoU threshold and history vector unclaimed
Unlock Full Analysis — Free Frequently asked questions
US 12,067,644 B2 — key questions answered
Still have questions? PatSnap Eureka can answer them from patent data instantly. Search in Eureka
Disclaimer: This analysis is generated by PatSnap Eureka AI based on publicly available patent data from the USPTO. It does not constitute legal advice and should not be relied upon as such. Patent data may be subject to change as prosecution progresses. Scores and assessments reflect automated analysis and may not capture all relevant legal or technical nuances. Always consult a qualified patent attorney for formal legal opinions on patentability, freedom to operate, or infringement.
Ask anything about this patent.
PatSnap Eureka searches patents and data to answer instantly.
PatSnap Eureka searches patents and data to answer instantly.
Powered by PatSnap Eureka
Link copied to clipboard