Attribute Label Recognition And Key Point Location Of Clothing Image Based On Deep Learning

The huge profit potential of the fashion industry and the rapid development of clothing e-commerce have driven the intelligent fashion analysis for clothing to gain widespread attention in the fields of multimedia,computer vision and pattern recognition.The attribute label recognition and key point positioning of clothing images as the basic problems in intelligent fashion analysis have important research significance.A wide variety of garments and small differences between many categories pose challenges for efficient and accurate image attribute label recognition and key point positioning.The deep learning algorithm has already gained a lot of results in the field of image processing,but its identification and localization methods have a series of problems such as single model,excessive data volume,poor real-time performance,and no special field considerations.Continuous improvement,especially when it comes to clothing images,there is still much room for improvement in attribute label recognition and key point positioning.Therefore,this thesis mainly uses the deep learning algorithm to study the identification and key point location of clothing image attributes.(1)A new deep learning model based on residual neural network ResNet50,namely Res-FashionAINet,is designed for attribute label recognition of clothing images.The model mainly uses a convolution template and intersperses an identity template.The Dropout layer is added at the end of the model to avoid over-fitting,and the full-join layer is added to enhance the expression of the output features.The model learns from the lowest layer of the network,then learns the distribution of image features layer by layer,and abstracts its feature map into more dimensional feature vectors in the network’s Dense Layer.Finally,it will be abstracted.The feature vector is input into the classifier to predict the probability that the tag corresponds to each attribute category,and the highest probability is judged as the final attribute tag recognition result of the image.Experiments on data preprocessing,model training and attribute prediction are carried out.On the FashionAI_attributes dataset,the accuracy and processing speed of the Res-FashionAINet model in the recognition of clothing image attribute tags are improved.(2)For the key point positioning of clothing,a new model based on stacked hourglass network,Hg-FashionAINet,is designed to process features at all scales to capture various spatial relationships related to clothing,and to predict and locate key points of clothing.The model is based on successive steps of pooling and upsampling,used to generate the final prediction set,and uses repeated bottom-up,top-down processing in conjunction with intermediate supervision,while employing cross-phase feature fusion and From rough to fine supervision,two design ideas are used to improve performance.The experiment achieved high positioning accuracy on the large-scale high-quality fashion dataset FashionAI_key_point.The two deep learning models designed in this thesis have good feature extraction ability in the clothing image dataset,high recognition accuracy and fast training speed,which will be beneficial to the further development of intelligent fashion analysis.