Small-scale Image Semantic Segmentation Method Based on Multilevel Superposition and Enhancement Fusion

Aiming at the problem of small-scale target loss and boundary discontinuity in image semantic segmentation, this paper proposes an image semantic segmentation method based on deep learning multi-level superposition and enhancement fusion. This method uses DeepLabV3+ network structure and constructs multi-level feature extraction. The method of multi-level superposition and jump connection improves the ability to acquire local features, so that while segmenting large targets, local feature information can also be obtained, and the segmentation accuracy of local target boundaries is improved. The network is trained, validated and tested on the public data set PASCAL VOC2012. Experiments show that the visual segmentation effect of this method on small-scale images is significantly better than most semantic segmentation methods that have appeared. At the same time, in the overall accuracy of all categories and the accuracy of several typical categories, the method in this paper is also superior to many previous methods.