Clothing Classification And Object Detection Based On Convolutional Neural Network

The wide variety of clothing and the small degree of distinction between many categories are the challenges of efficient and accurate image search and analysis.The tag information obtained from images can help solve the problem of image search and analysis.We can obtain the tag information from clothing image by image classification and object detection based on convolutional neural network,but the tag information obtained in this way is still lacking in accuracy.This thesis studies the image classification and object detection of clothing image in view of the above problems.Main work of the thesis:1)This thesis studied and improved four kinds of convolutional neural network models to classify the clothing pictures.Four kinds of convolutional neural network models,namely AlexNet,GoogLeNet,VGG16 and VGG19,were trained in the DeepFashion dataset published by the Chinese University of Hong Kong.And the model was adjusted according to the experimental results and the characteristics of different models.For the AlexNet model,this thesis added a convolution layer to extract the features better and increase the expression ability of the model;the adjusted image classification accuracy has improved significantly.For the VGG19 model,this thesis added an extra classifier to the network to enhance the gradient signal of the backward conduction.After the training,the accuracy of the image classification of the model had also been improved,and the increased classifier had shown a good effect.For the VGG16 model,in order to fuse the multi-layer features,Pooling layer and 1×1 convolution layer.After adjustment,the results of the VGG16+Concat model performed was the best in multiple sets of experiments.2)Experiment on clothing object detection based on SSD model.Firstly,this thesis processed the data and uses the dataset to train the SSD model.Then,we add the structure of the multi-layer feature fusion in the SSD model and carry on the experiment again.Finally,this thesis adjusts the training method to get the better object detection model.3)This thesis tested the real-time performance and practical application of the model obtained by training,and selected the appropriate model to develop the application of image classification and object detection for garment pictures.Firstly,the image classification model trained is tested,the image processing time of each model is obtained,and the real time performance of the model is evaluated according to the time of image processing.Then,developed testsets to test the accuracy of these three image classification models in practical applications.According to the real-time performance and the actual image classification accuracy,to select the appropriate image classification model.Making a testset to test time of image processing of the SSD model in applications and evaluated real-time performance.Finally,we used the selected image classification model and SSD model to produce a garment image processing application.In this thesis,we used the existing image recognition model to carry out the image classification and object detection experiment of clothing data,adjusted the model structure and training methods,and carried out several experiments,and got the appropriate model and training methods,and finally apply the model.