Mode-selectable computer vision using reference pixel substitution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer vision technologies face challenges in efficiently switching between image classification, object detection, and image segmentation due to the need for significant changes in model architecture, high resource intensity, and varying label fidelity requirements.
Innovation Solution
A method that utilizes a single image classification model trained to disregard reference pixels, allowing for mode-selectable computer vision operations by varying the proportion of the image substituted with reference pixels, enabling efficient switching between image classification, object detection, and image segmentation without the need for extensive retraining or architectural changes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single image classification model is used for all computer vision tasks, then model complexity is reduced, but the model cannot natively support different tasks without architectural changes
Solution Approach 1:
The patent applies dynamics by making the model's effective architecture configurable rather than fixed. Through the search space definition and reference pixel substitution mechanisms, the same base model can dynamically adapt to different tasks (image classification, object detection, image segmentation) by adjusting parameters like search space size and reference pixel proportions, eliminating the need for multiple fixed architectural variants
Solution Approach 2:
The patent changes physical or computational parameters of the model's operation rather than its architecture. By varying parameters such as the proportion of reference pixels, search space dimensions, and window sizes, the model achieves different task behaviors without modifying its underlying neural network structure, thus reducing architectural complexity while maintaining versatility
2Reliability
If separate models are used for image classification, object detection, and image segmentation, then each task achieves optimal performance, but resource intensity and training time increase significantly
Solution Approach 1:
The patent implements universality by designing a single multi-functional model that can perform image classification, object detection, and image segmentation through different operational modes. The model uses a unified architecture with a single image classification head that can be configured for different tasks through search space definition and reference pixel substitution, eliminating the need for multiple specialized models and reducing overall computational resource intensity
Solution Approach 2:
The patent applies segmentation by dividing the image into different search spaces and regions with different levels of detail. For object detection and segmentation tasks, the image is segmented into multiple regions that are processed at different resolutions, allowing the single model to handle diverse tasks by selectively processing different portions of the image at different levels of granularity
3Productivity
If reference pixels are substituted in the image, then the model can focus on specific regions for detection, but the substituted regions may interfere with object identification
Solution Approach 1:
The patent uses reference pixels as an intermediary mechanism that temporarily replaces image data in specific regions. These reference pixels serve as placeholders that allow the model to focus computational attention on particular areas during detection tasks. The reference pixels are designed to be disregarded by the image classification head, acting as a mediator between the input image and the detection process without interfering with actual object identification in the search space
Data Source
AI summary
The present disclosure provides a method including defining a search space within a first image, which includes defining a plurality of image regions, generating second images corresponding to the image regions (for an image region, the corresponding second image substitutes reference pixels for the portion of the first image outside the image region), and applying an image classification model (trained to disregard the reference pixels) to the plurality of second images. The method further includes performing object detection on the search space using the image classification model.