Visualization And Pruning Of SSD Network

Convolutional Neural Networks?CNNs?have performed well in object detection.However,due to the large size of network parameters,there is much redundancy in CNNs.This thesis focuses on analysing and pruning the redundancy of the state of art real-time detection network Single-Shot multi-box Detector?SSD?.SSD network is built on top of a base network Visual Geometry Group?VGG?that ends with some convolution layers.Its base network VGG,designed for 1000 categories in Image Net dataset,is obviously over-parametered,when used for 21 categories classification in VOC dataset.The large base network not only leads to the waste of storage and the decrease of network running speed,but also leads to the overfitting problem,resulting in the accuracy reduction of SSD network.In this thesis,the redundant parameters in the base network VGG of SSD network are analyzed by network visualization methods.Due to the "black box" characteristic of CNNs,the predecessors can only blindly prune network parameters.In order to find the redundant parameters of SSD network effectively,this thesis visualizes the feature learning of SSD network and accordingly finds the invalid parameters with poor feature learning.An iterative network pruning method is proposed,based on the analysis of network feature learning,to effectively prune the redundant parameters while maintaining the accuracy of SSD network.The discriminative feature learned by last layer conv5₃ of VGG network is analyzed,and redundant kernels are prunned in the experiments.Results illustrate the efficiency of the proposed method.A reduced SSD network is obtained with even higher mean Average Precision?m AP?than the original one by 2 percent.When only 4% of the original kernels in conv5₃ is remained,m AP is still as high as that of the original one.