Research On Object Detection Algorithm And Pruning Optimization Based On Improved SSD

Object detection is one of the core topics in the field of computer vision and the basis of many other visual tasks.Thanks to the advancement and development of deep learning technology and modern GPU computing capabilities,target detection technology based on deep learning has already achieved some results in some areas,such as intelligent security,unmanned vehicle driving,and medical image processing.The main task of target detection is to detect and locate specific targets from image information.It is a challenging subject that incorporates cutting-edge technologies in many fields such as image processing,pattern recognition,feature extraction,and deep learning.Therefore,this thesis further studies the SSD target detection algorithm based on deep learning technology,and uses the channel pruning algorithm to prune the backbone network of the SSD algorithm,making the SSD target detection algorithm more efficient.The specific research contents of this thesis are as follows:?1?An SSD target detection algorithm based on the fusion of dilated convolution and feature map is proposed.This algorithm improves the original SSD algorithm and raises the detection ability of the SSD algorithm.The original SSD algorithm needs to select candidate boxes on the six-layer feature maps of different scales for prediction,but the target information provided by these six-layer feature maps is limited.In order to solve this problem,this thesis uses Res Net50 with stronger feature expression ability to replace the VGG16 network in the original SSD algorithm.At the same time,two modules are designed.One is the feature fusion module,which merges the deep feature maps with large fields and rich semantic information and the shallow feature maps with small fields and detailed information to enhance the overall expression ability of the feature map of an image.The other is a hollow convolution to further expand the receptive field of shallow feature maps and enhance semantic information.Experimental results show that the design effectively improves the ability of the SSD algorithm to detect targets.?2?In order to further reduce the storage space and computing resources required by the improved SSD algorithm,we used the channel pruning algorithm to compress the backbone network Res Net50 of the improved SSD model in model compression,so as to realize the efficient operation of the SSD network model on devices with limited computing and storage.In the past,when the channel pruning algorithm was used,in order to better obtain the smallest reconstruction error,even the input channels that are completely unrelated to the feature discrimination capability of the convolutional neural network are likely to be incorrectly retained.In view of this problem,we introduce the channel attention mechanism to more accurately judge the importance of features in its input channel.For the input feature map of the channel attention unit,it is necessary to first learn the importance of different channels,that is,assign weights to different channels.Therefore,the importance C?a 1×N vector?of each layer in the feature map can be calculated by the channel attention unit,and the channel corresponding to the former6)₇)?6?₇)represents the number of channels to be retained after pruning in the l layer)larger values in the vector C is saved to achieve the purpose of pruning.Experimental results show that the proposed method effectively reduces the redundancy of network parameters and achieves network acceleration and compression.