Image Segmentation Label Generation via Feature Response Enhancement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image semantic segmentation networks face poor training effects due to the use of coarse-grained class labels derived from Class Activation Mapping (CAM), which results in incomplete object coverage and lower accuracy.
Innovation Solution
An image segmentation label generation method that involves acquiring a feature map, determining a feature response map, increasing response values within a preset range, reconstructing the feature map, and determining a first-class activation mapping to generate high-precision segmentation labels, thereby enhancing the training effect of image semantic segmentation networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If Class Activation Mapping (CAM) is used as segmentation label, then the segmentation label can be obtained easily, but the accuracy and completeness of object coverage deteriorates
Solution Approach 1:
The method performs preliminary processing on the feature map by identifying and enhancing key response regions before generating the final segmentation label. This preliminary enhancement of critical features ensures that the subsequent CAM generation produces more accurate and complete object coverage, resolving the contradiction between ease of obtaining labels and their accuracy.
Solution Approach 2:
The method modifies the response values within the feature map by applying enhancement operations to specific regions. This parameter change in the feature representation allows the CAM to generate segmentation labels with improved accuracy and completeness while maintaining the overall simplicity of the process.
2Device complexity
If CAM is used as segmentation label, then the process remains simple, but the training effect of image semantic segmentation network deteriorates
Solution Approach 1:
The method performs preliminary enhancement of key response regions in the feature map before CAM generation. This preliminary action improves the quality of input features for the segmentation network, leading to better training effects while adding only minimal processing steps to the overall workflow.
Solution Approach 2:
By enhancing specific response values in the feature map, the method changes the parameter distribution to emphasize important features. This parameter adjustment improves the reliability of training without significantly increasing process complexity.
3Manufacturing precision
If response values are enhanced within preset range, then the accuracy and completeness of segmentation label improves, but the processing complexity increases
Solution Approach 1:
The method applies enhancement operations only to specific regions within the feature map where response values fall within a preset range. This localized processing approach improves segmentation accuracy by focusing computational resources on critical areas while minimizing the overall increase in processing complexity.
Solution Approach 2:
The method applies partial enhancement to only those regions that meet specific criteria (response values within preset range) rather than processing the entire feature map uniformly. This selective approach achieves improved accuracy without proportionally increasing processing complexity.
Data Source
AI summary
Provided in the present disclosure are an image segmentation label generation method and apparatus, and an electronic device and a storage medium. The image segmentation label generation method includes: acquiring a feature map of an original image, determining a feature response map of the feature map, wherein a response value in the feature response map represents a weight of a corresponding feature in the feature map in image classification; increasing a response value within a preset range in the feature response map, reconstructing the feature map according to a feature response map with the increased response value; determining a first-class activation mapping based on the first reconstructed feature map, and determining an image segmentation label according to the first-class activation mapping.