Channel Pruning for Deep Learning Object Detection Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection technologies based on deep learning models are slow due to their large size, making them unsuitable for devices with limited computing resources, such as unmanned distribution vehicles in smart logistics.
Innovation Solution
A model generation method and apparatus that employs model compression through channel pruning, where scaling coefficients of a batch normalization layer are used to identify and prune less important channels in the intermediate detection model, resulting in a faster object detection model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a deep learning model with large size is used for object detection, then the detection accuracy is improved, but the detection speed deteriorates
Solution Approach 1:
The patent extracts and removes redundant channels from the deep learning model based on importance assessment of batch normalization scaling coefficients. By taking out unnecessary channels while retaining important ones, the model size is reduced and detection speed is improved while maintaining detection accuracy.
Solution Approach 2:
The patent changes the model structure parameter by pruning channels based on scaling coefficient thresholds. This parameter change transforms the model from a dense structure to a sparse structure, improving detection speed while preserving accuracy through intelligent channel selection.
2Measurement precision
If a deep learning model with large size is used for object detection, then the detection accuracy is improved, but the device complexity increases
Solution Approach 1:
The patent extracts and removes redundant channels from the deep learning model based on importance assessment of batch normalization scaling coefficients. By taking out unnecessary channels while retaining important ones, the model size is reduced and detection speed is improved while maintaining detection accuracy.
Solution Approach 2:
The patent changes the model structure parameter by pruning channels based on scaling coefficient thresholds. This parameter change transforms the model from a dense structure to a sparse structure, improving detection speed while preserving accuracy through intelligent channel selection.
3Speed
If channel pruning is performed on the intermediate detection model, then the detection speed is improved, but the model structure is modified
Solution Approach 1:
The patent changes the model structure parameter by pruning channels based on scaling coefficient thresholds. This parameter change transforms the model from a dense structure to a sparse structure, improving detection speed while preserving accuracy through intelligent channel selection.
Solution Approach 2:
The patent uses scaling coefficients from batch normalization layers as feedback signals to identify and prune redundant channels. This feedback mechanism guides the pruning process to maintain model performance while improving speed.
Data Source
AI summary
Provided are a model generation method and apparatus, an object detection method and apparatus, a device and a storage medium. The method includes acquiring multiple scaling coefficients of a batch normalization layer in an initially-trained intermediate detection model, where the intermediate detection model is obtained by training an original detection model based on multiple training samples, and each training sample includes a sample image and a sample annotation result of a known object in the sample image; screening a to-be-pruned coefficient from the multiple scaling coefficients according to values of the multiple scaling coefficients; and screening a to-be-pruned channel corresponding to the to-be-pruned coefficient from multiple channels of the intermediate detection model and performing channel pruning on the to-be-pruned channel to generate an object detection model.