Research On Clothing Detection Method Based On Improved SSD Algorith

China is a multi-ethnic country,and many ethnic groups have unique costumes,reflecting the different cultural backgrounds of each ethnic group.This thesis collects a large amount of clothing image data for 8 kinds of common clothing in my country,and uses the deep learning method to study clothing recognition.The advantages and disadvantages of commonly used deep learning methods are analyzed and improved.The main contents of this thesis are as follows:(1)Construct a clothing dataset,including Tibetan,Manchu,Miao,Korean,Tibetan Buddhism,and Zhuang national costumes,and other costumes of police and doctors.The SSD algorithm is selected as the detection method for clothing recognition,and the original ordinary convolution block is replaced by a depthwise separable convolution block,and the downsampling method and network structure of the SSD are improved to complete the lightweight design of the SSD.A comprehensive evaluation of improved lightweight SSD is performed using the VOC dataset.Finally,a lightweight SSD is used to complete the experiments on the clothing dataset,and compared with Mobilenet V1-SSD and SSD.(2)In view of the problem that the improved lightweight SSD algorithm has relatively poor effect in identifying clothing with insignificant features,parameter optimization and network structure changes are carried out.Compare the performance of Leaky Relu,Mish,and Relu6 activation functions on the clothing dataset,and analyze the performance of adding two different position residual blocks on the clothing dataset.Experiments show that the two residual blocks have better recognition effects,and the Relu6 function in the activation function has better effects,and its calculation is relatively simple.Finally,fusion experiments are performed on all the improvements.(3)The overall framework of the circuit is designed for the improved SSD algorithm.Complete the circuit design of channel-by-channel convolution and point-by-point convolution respectively.The function simulation of the circuit is completed by C language and Verilog,and the logic synthesis of the channel-by-channel convolution circuit and the point-by-point convolution circuit is completed by using DC software,and its timing,area and power consumption are analyzed.