Research On Clothing Image Detection Algorithm Based On Deep Learning

Clothing detection is the use of models to determine the location and specific category of clothing in an image or video.Image-based clothing detection requires the network to extract accurate location information of clothing and the key parts of the clothing that distinguish between different clothing categories.However,on the one hand,there is a serious imbalance in the number of images with different sizes of clothing in the dataset,making it difficult for the detection network to be fully trained.On the other hand,clothing has the problem of many different forms of clothing in the pictures,being occluded,and the shooting viewpoint changes,making it difficult to give accurate positioning of clothing and extract distinguishing features.Therefore,accurate clothing positioning and extraction of clothing features has become one of the challenges of clothing recognition.At present,the approaches to clothing detection are mainly divided into two categories: one is top-down,which is anchor-based and needs to calculate the IOU between the anchor box and the bounding box,but it is limited by the setting of the anchor box and does not perform well when the clothing scale is variable;the other is the bottom-up method,which uses the feature extraction network to get the keypoints and calculates the location and size of the clothing by the keypoints,but the prediction of the keypoints often has a slight error.In order to address the above issues,the following work was completed:(1)To address the difficulty of accurately locating clothing,this thesis proposes a method for the linear combination of central keypoint.The linear combination of central keypoint improves the original algorithm’s prediction of central keypoint,allowing for more accurate central keypoint to be predicted.Multi-keypoint matching network,which belongs to the bottom-up approach.It detects clothing to get three keypoints(top-left corner keypoint,bottom-right corner keypoint,and center keypoint),performs corner keypoint matching by calculating the distance between the embedding vectors of different corner keypoints to get the initial bounding box,and finally gets the final bounding box by matching the center keypoint.The experimental results demonstrate that this method improves the mAP metric by 4.9% compared to DeepMark++ on the DeepFashion2 dataset.(2)To address the problem of difficulty in accurately extracting highly distinguishable clothing features,this thesis proposes a multi-region sampling technique.The key part feature extraction network accurately captures clothing features using deformable convolution,then captures the key part features of the clothing using block division and special pooling in multi-region sampling technique.Finally,FPN plus feature fusion is used to deepen the features of the clothing part and accurately acquire the parts of the image that contribute significantly to clothing recognition.The experimental results show that the structure improves the mAP metric by 3% on top of MKMnet and 7.9% compared to DeepMark++ in terms of mAP metric.In summary,by proposing a linear combination of central keypoint and a multiregion sampling technique,this thesis solves the difficulties of accurately locating clothing and accurately extracting clothing features,and clothing detection accuracy is further improved.