Research On Neural Network Compression And Optimization Algorithm Based On Object Detection

For a long time,object detection is a hotspot and difficulty in the field of computer vision.In recent years,with the rapid development of deep learning,the object detection algorithm based on convolutional neural networks has made a breakthrough.However,convolutional neural networks have the disadvantages of computationally intensive and parameter redundancy,especially object detection algorithms that require a large number of proposals to be extracted in advance.In this paper,for the problem of heavy computation and redundant parameter of convolutional neural network,we take the object detection algorithm based on convolutional neural network(Faster RCNN)as a research object,and combine network structure optimization and coarse-grained and fine-grained pruning method,the algorithm is successfully improved.And we also explore the applicability of network compression and optimization methods,the research content and main contributions of this paper include:Firstly,for the problem that the computation and model parameters in the Faster RCNN are mainly concentrated on the feature fusion module,this paper proposes an efficient feature fusion module,which first reduces the number of channels by using dimensionality reduction convolution,and then fuses the local channel feature information by pointwise convolution and the group convolution extracts the spatial feature information.The group fully connected layer fuses feature information extracted by the convolution layer.The module effectively classifies and locates proposals,and reduces the amount of computation and the number of parameters.Secondly,for the problem of insufficient feature extraction ability in region proposal network of the Faster RCNN,this paper proposes a multi-scale dilation region proposal network,which combines multi-scale feature information of shallow,middle and high level,expands the local receptive field,enhances the more effective channel feature information and suppresses the more invalid channel feature information.The subnetwork improves quality of proposals and recall rate,and finally improves the detection accuracy.Finally,for the problem that the computation is concentrated in the convolutional layer and parameters are concentrated in the fully connected layer,this paper studies coarse-grained pruning and fine-grained pruning.The coarse-grained pruning adopts the method of filter weighting and sorting,and cuts the unimportant channels in the convolutional layer.The fine-grained pruning cuts the unimportant weights in the fully connected layer according to the absolute value of the weight,which greatly reduces the computation and parameters.In addition,we propose a fine-grained pruning that can automatically learn the sparse network structure,control the sparsity unchanged,recover the miscutting weight by cutting and splicing,and automatically learn and train to get a good sparse fully connected layer,which ensures the detection and recognition accuracy.Through experimental verification and analysis,our network structure optimization,coarse-grained and fine-grained pruning method not only effectively reduces computation,parameters,memory usage and power consumption of Faster RCNN,but also improves its detection accuracy.At the same time,the method can be expanded to other object detection algorithms and classification network models,and optimize the performance of other network structures.