Research On Clothing Instance Retrieval Method Based On Local Feature Extraction Under The Framework Of Deep Learning

With the development of the Internet and e-commerce,online shopping has become one of the current mainstreams of consumption.Due to the case that the sales of traditional clothing stores have declined,more and more businesses are opening stores on the Internet to gain market share,resulting in more buyers starting to buy clothes online.Over the past decade,there are an increasing number of clothing images on the Internet,and retrieving effective clothing images from a huge number of clothing images has become a challenge.An excellent clothing retrieval system can improve consumers' experience and thus increase sales,which is crucial for e-commerce platforms.Motivated by this end,this paper studies the clothing retrieval method under the deep learning framework.The main contents of this paper are as follows:1.This work uses the most popular keypoint-based object detection model Center Net as the backbone network.In detail,the feature extraction network uses two different networks.The deep aggregation network can well integrate semantic and spatial features,and the stacked hourglass network with larger parameters can obtain more accurate keypoint information.The prediction network has three branches,including the prediction of the center point of the object,the center point offset,and the width and height of the object.Center Net uses Gaussian kernel function to predict the center point of clothing,and its border and offset are corrected by center point coordinate regression.Experiments show that the detection effect of the stacked hourglass network is better than that of the deep aggregation network in the cases of small clothing,occluded clothing,interference from the background,as well as the case that the tops and skirts that can be easily identified as dresses.2.Based on the object detection network model,this work adds keypoint detection-related prediction branches to construct a clothing pose estimation network model.Since the number of keypoints of different clothing types in the Deep Fashion2 data set is generally different,this work selects a large number and representative Clothing types: short sleeve top,long sleeve top,trousers,short sleeve dress for handling clothing pose estimation.Here,by loading the optimal target detection model pre-trained by the corresponding feature extraction network,the convergence speed of the clothing pose estimation model can be accelerated.In addition,the added prediction branches are keypoints prediction,keypoints offset prediction and keypoints to the center point distance prediction.The first two predictions can determine the location of the keypoints.While the last one re-optimizes the location of the keypoints through the relative position between the keypoints and the center point to ensure the effect of clothing pose estimation.In the case of occlusion,background interference,etc.,the keypoint positioning via the stacked hourglass network is more accurate,and the clothing pose estimation effect is better.3.This article uses Res Net-50 to perform feature extraction on the image of the target area for clothing image retrieval.In detail,the feature extraction in two different ways: extract global features of clothing through global pooling,or combine the information of keypoints to extract the local parts of clothing keypoints features and then fusion of global features and local features.Next,we train the model for classification on the extracted clothing features and use certain distance loss functions for fine-tuning.In this work,instead of utilizing the triple loss function for distance learning,we use another metric learning loss function N-pair loss,where the loss function is trained by loading N samples during training.Each sample corresponds to a positive sample and N-1negative samples,which is fully considered the effect of different negative samples on training.Experiments prove that the superiority of this loss function to the triple loss function.