Feature Map Splitting for Normalization Error Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Deep learning networks face challenges in training stability due to the 'Vanishing Gradient' problem and require efficient normalization methods, especially when dealing with tasks like object detection and crowd density estimation, where batch normalization is unsatisfactory.
Innovation Solution
An image processing method involving feature map splitting, normalization, and splicing, along with scale reduction and multi-scale fusion, is employed to enhance feature extraction and prediction accuracy. This method includes splitting feature maps based on dimension information, normalizing sub-feature maps, and splicing them to retain local information, reducing statistical errors and improving feature validity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If batch normalization is used for normalization, then training speed is improved, but normalization effectiveness deteriorates in tasks like crowd density estimation
Solution Approach 1:
The patent divides the feature map into multiple sub-feature maps based on spatial dimensions, then performs normalization on each sub-feature map independently. This segmentation approach allows the model to capture local statistical characteristics while maintaining training efficiency, resolving the contradiction between batch normalization's speed advantage and its effectiveness in certain tasks.
2Ease of operation
If feature maps are processed without splitting, then processing simplicity is maintained, but local feature information is lost due to statistical errors
Solution Approach 1:
The patent segments the feature map into multiple sub-feature maps along spatial dimensions before normalization. This segmentation preserves local feature information by computing statistics independently for each sub-region, thereby improving measurement precision without significantly complicating the processing pipeline.
Solution Approach 2:
After independent normalization of sub-feature maps, the patent merges them back into a complete feature map through splicing operations. This merging step recovers the global structure while retaining local feature precision, effectively combining the benefits of both localized and global processing.
3Productivity
If the feature map is normalized as a whole, then computational efficiency is improved, but statistical errors increase reducing feature validity
Solution Approach 1:
The patent divides the feature map into multiple sub-feature maps and performs normalization on each segment independently. This approach reduces statistical errors by capturing local variations, while the segmented processing maintains computational efficiency through parallelizable operations and避免了global statistical computations.
Data Source
AI summary
The present disclosure relates to an image processing method and device, an electronic apparatus and a storage medium. The method comprises: performing feature extraction on an image to be processed to obtain a first feature map of the image to be processed; splitting the first feature map into a plurality of first sub-feature maps according to dimension information of the first feature map and a preset splitting rule, wherein the dimension information of the first feature map comprises dimensions of the first feature map and size of each dimension; performing normalization on the plurality of first sub-feature maps respectively to obtain a plurality of second sub-feature maps; and splicing the plurality of second sub-feature maps to obtain a second feature map of the image to be processed. Embodiments of the present disclosure can reduce the statistical errors during normalization of a complete feature map.