Research On Fashion Landmark Detection Based On Attention Mechanism And Residual Neural Network

Fashion landmark detection aims at localizing functional keypoints or parts of clothing,which can help identify clothing styles,categories,etc.Fashion landmark detection is challenging due to the large spatial variances of the landmarks and the scale variations of the clothing images.To address the above challenges,fashion landmark detection models need to be able to autonomously focus on task-related features,and extract features containing multi-scale contextual information.Therefore,this paper proposes fashion landmark detection algorithms based on the attention mechanism and residual network.The main work includes:(1)A fashion landmark detection model based on spatial attention residual blocks and multi-scale context aggregation is proposed.First,a new spatial attention mechanism is designed,which is a direction-aware spatial attention module;a spatial attention residual block is constructed based on the residual idea,and a backbone network is built based on this module.Then,the top-to-bottom feature fusion method of the feature pyramid and the multi-scale feature aggregation method are used to extract the multi-scale contextual information.Finally,a deep residual spatial attention neural network is built.Training and testing are performed on two large-scale datasets,Deep Fashion and FLD,respectively.Compared with the best performing MDDNet among the comparison methods,the normalized errors are reduced by 14.54% and 3.86% on the Deep Fashion and FLD datasets,respectively.The model achieves advanced performance.(2)A fashion landmark detection model based on a lightweight deep residual spatial attention network is proposed.Aiming at the problems of large amount of parameters,large amount of calculation and high network training cost of the deep residual spatial attention neural network,a lightweight design is carried out.A lightweight deep residual spatial attention neural network is constructed by adopting different convolution decomposition methods.Training and testing are performed on two datasets,Deep Fashion and FLD,respectively.We compared with the best performing MDDNet among various existing methods,and found out the normalization error decreases by 14.48% and 0.77%,respectively.Compared with the deep residual spatial attention neural network,the amount of parameters is reduced by 45.94%,and the amount of calculation Flops is reduced by74.0%.(3)A clothing classification and attribute prediction model based on fashion landmark detection is proposed.In order to verify the effectiveness of the deep residual spatial attention neural network,we apply this model to the tasks of clothing classification and attribute prediction.Based on this model,a keypoint heatmap-guided attention mechanism is designed to enhance the features extracted by VGG16.Training and testing are performed on the Deep Fashion dataset.Compared with the model that does not include the attention mechanism guided by the keypoint heatmap,we found out the top-3 and Top-5accuracy of clothing classification are improved by 2.97% and 1.70%,respectively.The prediction is improved by 13.33% and 12.43% in Top-3 and Top-5 recall rates,respectively.The performance of the deep residual spatial attention network model is better than the GLE and MDDNet models in both clothing classification and attribute prediction tasks.In this work,two fashion landmark detection models are proposed.One improves the detection performance,and the other reduces the detection cost.And based on the first model,a model of clothing classification and attribute prediction is built,which further proves the effectiveness of this model.