Layerwise Convolution Channel Pruning for Faster Image Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image recognition technologies consume excessive calculation resources and have low efficiency due to redundancy in depth neural network parameters.
Innovation Solution
An image recognition method involving kernel set construction to update convolution kernel weights, clip redundant input channels, and generate a compressed neural network model without retraining, optimizing feature map reconstruction errors through multiple rounds of sampling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If depth neural network parameters are used for image classification and recognition, then recognition accuracy can be maintained, but calculation resources are excessively consumed and recognition efficiency is low
Solution Approach 1:
The patent extracts and removes redundant parameters from the original depth neural network model through parameter pruning technology. By identifying and eliminating unnecessary connection weights and neurons that contribute minimally to recognition accuracy, the model achieves reduced computational complexity while preserving essential recognition capabilities, thus resolving the contradiction between maintaining accuracy and improving efficiency
Solution Approach 2:
The patent transforms the neural network parameters from a dense state to a sparse state through pruning operations. This parameter change reduces the number of calculations required during inference while maintaining the functional integrity of the network, enabling faster recognition speed without significant accuracy loss
2Measurement precision
If depth neural network parameters are used for image classification and recognition, then recognition accuracy can be maintained, but calculation resources are excessively consumed
Solution Approach 1:
The patent extracts and removes redundant parameters from the original depth neural network model through parameter pruning technology. By identifying and eliminating unnecessary connection weights and neurons that contribute minimally to recognition accuracy, the model achieves reduced computational complexity while preserving essential recognition capabilities, thus resolving the contradiction between maintaining accuracy and improving efficiency
Solution Approach 2:
The patent discards redundant parameters that consume calculation resources but contribute minimally to recognition performance. Through systematic pruning, the model eliminates wasteful computational operations while recovering and optimizing the utilization of essential parameters, thereby reducing overall calculation resource consumption
3Quantity of substance
If traditional neural network compression methods are used, then model size is reduced, but retraining is required which increases time consumption
Solution Approach 1:
The patent performs parameter pruning during the model training phase as a preliminary action, integrating the compression process into the original training workflow. By pruning parameters and updating the model in the same training run without requiring separate retraining stages, the method achieves model compression while avoiding additional time consumption, thus resolving the contradiction between reducing model size and minimizing time loss
Data Source
AI summary
An image recognition method includes: inputting a sample image data set into an original neural network model in advance; and for each convolution layer of the original neural network model, using a feature map of the sample image data set in the current layer as a reconstruction target, firstly obtaining an update weight of a convolution kernel by using a kernel set construction method, then calculating an input channel combination having a minimum reconstruction error, cutting a redundant input channel to obtain a compression result of the current convolution layer, and finally, splicing compression results of the convolution layers to generate an image recognition model. An image to be recognized is obtained and is inputted into the image recognition model, and an output result of the image recognition model is used as an image recognition result of the image.