Super Network Merging for Target Detection NAS
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing target detection Neural Architecture Search (NAS) methods result in sub-networks with simple structures, limiting the performance of the finally searched network structure in image processing.
Innovation Solution
Merging a pretrained backbone network with the rear end of a target detection network to form a super network, which is then trained and subjected to NAS to obtain a target detection neural architecture, enhancing the search space and training efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a pretrained backbone network is merged with the rear end of a target detection network to form a super network, then the search space becomes more abundant and training efficiency improves, but the device complexity and computational resources required increase
Solution Approach 1:
The patent merges a pretrained backbone network with the rear end of a target detection network to form a merged super network. This combination integrates the feature extraction capabilities of the backbone network with the detection capabilities of the target detection network, creating a unified structure with abundant search space for NAS while maintaining organized modularity between the two components.
Solution Approach 2:
The super network is divided into distinct functional modules: the backbone network for feature extraction and the rear end network for target detection. This segmentation allows independent optimization of each module while maintaining their integrated functionality, managing complexity through structured division of labor.
2Manufacturing precision
If the merged super network is trained and subjected to NAS, then the image processing performance improves, but the training time and computational resources increase
Solution Approach 1:
The backbone network is pretrained separately before being merged with the target detection network. This preliminary training establishes strong feature extraction capabilities in advance, providing a solid foundation that accelerates subsequent joint training of the merged super network and reduces overall training time.
Solution Approach 2:
The pretrained backbone network's learned features are continuously utilized during the joint training of the merged super network. Rather than retraining features from scratch, the system maintains and refines the pretrained features while integrating detection capabilities, ensuring continuous useful action and reducing redundant computation.
Data Source
AI summary
The present disclosure relates to an image processing method and apparatus based on a super network, and a computer storage medium. The method can include that a pretrained backbone network is merged with a rear end of a target detection network to obtain a merged super network, the merged super network is trained, Neural Architecture Search (NAS) is performed based on the trained super network to obtain a target detection neural architecture, and an image to be processed is processed by using the target detection neural architecture to obtain an image processing result.