Abstract:

Aiming at the low accuracy of small size clothing and occlusion clothing image segmentation, a clothing image segmentation method based on improved multi-scale feature learning strategy and attention mechanism was proposed, using Mask R-CNN as the basic framework. Augmentation feature pyramid network was used to optimize the feature learning process of the model, the semantic gap between different levels was narrowed through unified supervision of multi-scale clothing features, the residual feature enhancement module was introduced to reduce the loss of high level features, and soft region of interest selection was used to adaptively obtain the optimal region of interest characteristics. Then, the channel attention module was introduced into the classification prediction branch, the spatial attention modules were introduced into the bounding box regression and mask prediction branches respectively, and the clothing area features that need to be focused on in the image were extracted. The experimental results show that the proposed method can improve the detection rate and the segmentation completeness in dealing with small size clothing and occlusion clothing images, and the extracted clothing instance is more accurate comparing with the other methods, and the mean average precision is 3.8 points higher than that of the original model.

Key words: clothing image segmenttation, multi-scale feature learning, consistent supervision, residual feature augmentation, soft region of interest selection, attention mechanism