Enhanced Semantic Image Segmentation Through Convolutional and Atrous Convolution Techniques

Abstract

Enhanced Image content classification has improved dramatically with the advent of CNNs. This paper presents an enhanced method for semantic partitioning through merging traditional convolutional level and atrous (extended) convolution techniques. Our approach takes advantage of the hierarchical feature extraction capabilities of CNNs, while incorporating atrous convolutions to capture multi-scale contextual information without increasing the computational load. The proposed feature combines standard diffraction layers for detailed feature extraction that broadens the perceptive field, thus improving segmentation accuracy,
especially on multiscale features Extensive testing on the datasets including PASCAL VOC 2012 and Cityscapes.