Research on garment flat multi-component recognition based on Mask R -CNN

Abstract

The automatic recognition of garment flat information has been widely researched through computer vision. However, the unapparent visual feature and low recognition accuracy pose serious challenges to the application. Herein, inspired by multi-object instance segmentation, the method of mask region convolutional neural network (Mask R-CNN) for garment flat multi-component is proposed in this paper. The steps include feature enhancement, attribute annotation, feature extraction, and bounding box regression and recognition. First, the Laplacian was employed to enhance the image feature, and the Polygon annotated component attributes to reduce the interaction interference. Next, the ResNet was applied to realize identity mapping to characterize redundant information of components. Finally, the feature map was entered into two branches to achieve bounding box regression and recognition. The results demonstrated that the proposed method could realize multi-component recognition effectively. Compared with the unenhanced feature, the mAP increased by 2.27%, reaching 97.87%, and the average F1 was 0.958. Compared to VGGNet and MobileNet, the ResNet backbone used for Mask R-CNN could improve the mAP by 11.55%. Mask R-CNN was more robust than the state-of-the-art methods and more suitable for garment flat multi-component recognition.