Object Detection Model Training via Candidate Region Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition technologies face low detection accuracy due to the large size of images containing target objects and small target objects within them, which hampers the effectiveness of machine learning models during training.
Innovation Solution
A detection model training method that segments sample images into candidate image regions, uses multiple detection models to assess the inclusion of target objects, and selects a target image region based on relationship degree change values to improve object detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If the entire sample image is used for model training, then the model can learn from comprehensive image data, but the detection accuracy decreases because the target object is relatively small within the large image
Solution Approach 1:
The sample image is segmented into multiple candidate image regions, and the model is trained to select the target object region from these candidates. This segmentation approach allows the model to learn from comprehensive image data while focusing on specific regions containing target objects, thereby improving detection accuracy without losing the benefit of full-image training data.
2Measurement precision
If the image size is reduced to highlight the target object, then detection accuracy may improve, but important contextual information is lost
Solution Approach 1:
The patent introduces a new dimension by generating multiple candidate image regions from the full-resolution sample image. Instead of simply resizing the image, the system creates multiple regional views that preserve original image quality while providing focused attention on potential target areas. The model learns to select the correct candidate region, thereby maintaining contextual information from the full image while achieving high detection accuracy.
3Measurement precision
If multiple detection models are trained to evaluate candidate regions, then detection precision improves, but computational complexity and training time increase
Solution Approach 1:
The training process is segmented into stages: first training on the entire sample image to learn global features, then training on segmented candidate image regions to learn local target object features. This segmentation of the training process allows the model to achieve high detection precision while managing computational complexity through progressive learning rather than requiring all candidate regions to be processed simultaneously.
Solution Approach 2:
The system performs preliminary action by generating candidate image regions and their relationship degrees before the final detection stage. By pre-processing and evaluating candidate regions with relationship degree calculations, the system reduces the complexity of the main detection task, allowing the trained model to focus on making accurate predictions rather than searching through all possible regions.
Data Source
AI summary
A computer device segments a first sample region to obtain a candidate image region set that includes a plurality of candidate image regions. For each of the candidate image regions, the device obtains a first relationship degree corresponding to each candidate image region and obtains a second relationship degree corresponding to the candidate image region. The device obtains a relationship degree change value based on the first relationship degree and the second relationship degree. The device selects, from the plurality of candidate image regions, a first candidate image region as a target image region in accordance with a determination that the first candidate image region satisfies a condition in the relationship degree change value. The device performs model training based on the target image region to obtain a target detection model.