Dilated convolution and spatial pyramid fusion in the image segmentation problem

Abstract

Image segmentation is one of the important tasks in the computer vision, where the goal is to segment different regions in an image into semantically meaningful parts. However, due to the presence of target and contextual information at different scales in an image, traditional segmentation methods face the challenges of information loss and lack of accuracy when dealing with images at different scales. To address this problem, this study proposes an innovative approach that combines dilated convolution and spatial pyramid pooling to improve the processing power and accuracy of segmentation models for images of different scales.