AI Model Candidate Box Size Refinement for Grid Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition systems face challenges in accurately detecting objects in images arranged in a grid format, often resulting in inaccurate location estimation and inconvenience in manually drawing patterns.
Innovation Solution
An electronic device equipped with a processor and memory that applies images to an artificial intelligence model to generate candidate boxes for objects, with a loss function used to adjust the sizes of these boxes to reduce size differences, thereby improving detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If objects are arranged in a grid format for image recognition, then the structured organization facilitates processing, but the location estimation accuracy deteriorates
Solution Approach 1:
The patent divides the grid-based detection problem into multiple candidate boxes for each object, where each candidate box represents a possible location and size. The AI model generates multiple segmented hypotheses rather than a single grid-aligned detection, allowing precise location estimation while maintaining grid processing benefits.
Solution Approach 2:
The patent introduces dynamic adjustment of candidate box sizes through a loss function that optimizes the difference between adjacent boxes. This dynamic sizing mechanism allows the system to adapt box dimensions to actual object boundaries rather than being constrained by fixed grid cells, improving location accuracy.
2Adaptability or versatility
If manual pattern drawing is required for object detection, then customization is possible, but operational convenience deteriorates
Solution Approach 1:
The patent enables the system to automatically generate and optimize candidate boxes without requiring manual pattern drawing. The AI model self-adjusts the detection boxes based on the loss function that minimizes size differences between adjacent objects, providing both customization and convenience simultaneously.
Solution Approach 2:
The patent performs preliminary generation of multiple candidate boxes before final detection, allowing the system to pre-compute possible object locations and sizes. This preliminary action eliminates the need for manual pattern drawing while maintaining detection accuracy through automated optimization.
3Reliability
If candidate boxes are generated for each object, then detection completeness is improved, but the size difference between adjacent boxes increases
Solution Approach 1:
The patent implements a feedback mechanism through the loss function that continuously monitors and adjusts the size difference between adjacent candidate boxes. The loss function provides feedback signals to the AI model during training, enabling the system to learn optimal box sizing that maintains both detection completeness and size consistency.
Solution Approach 2:
The patent changes the optimization parameters by introducing a loss function that specifically targets the size difference between adjacent boxes. This parameter change shifts the optimization goal from purely detection accuracy to also including size consistency, resolving the contradiction between completeness and precision.
Data Source
AI summary
An electronic device for detecting a target object is configured to obtain a plurality of first candidate boxes corresponding to a first object and a plurality of second candidate boxes corresponding to a second object by applying an image is provided. The electronic device includes the first object and the second object to an artificial intelligence model, wherein the artificial intelligence model is trained to use a loss function for reducing a size difference between candidate boxes corresponding to two adjacent objects to determine sizes of the candidate boxes.