A Weakly Supervised Image Semantic Segmentation Method Based on Spatial Pyramid Mask Pooling

Abstract

The invention discloses a weakly supervised image semantic segmentation method based on spatial pyramid mask pooling, comprising the following steps: selecting a convolutional neural network H, processing an input image X through the convolutional neural network H, and obtaining a classification feature map; The classification feature map establishes the spatial pyramid pooling module, and then performs spatial pyramid masking to obtain the output feature map; calculates the category activation vector and category probability vector according to the output feature map, and then establishes a competitive spatial pyramid to cover the pooling loss function; according to the competitive space The pyramid masking pooling loss function trains a convolutional neural network H and extracts segmentation feature maps. The invention realizes a weakly supervised image semantic segmentation model with richer local features, more perfect regional feature mining, and more robust target size and attitude, improves the extraction ability of local semantic information, and strengthens local targets or parts in weakly supervised semantic segmentation recognition ability.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and in particular relates to a method for semantic segmentation of weakly supervised images based on spatial pyramid masking pooling. Background technique [0002] Image semantic segmentation is a basic computer vision task, and its goal is to classify all pixels in the image. Since it can fully understand images at the pixel level, it is useful for other vision tasks such as image classification and object recognition. However, because the production of pixel-level label data requires a lot of energy, fully supervised image semantic segmentation is difficult to quickly achieve large-scale expansion, so weakly supervised image semantic segmentation methods that rely on image category labels have been widely studied. [0003] In past computer vision research, the pyramid model has been used a lot. The well-known SIFT algorithm extracts key point description features through the Laplacian...

AI Technical Summary

Problems solved by technology

On the other hand, although the pyramid model has the characteristics of multi-scale information and local information fusion, the problem of how to better mine the sub-regional semantic information on the basis of mastering the global information has not been completely solved.

Images