Abstract

The semantic segmentation task is a basic task in the field of Mobile Edge Computing, which requires the classification of each pixel in the image, which has higher requirements for classification accuracy than the image classification task. Fine-grained classification tasks requires more detailed information, in addition to classifying according to the semantic information and spatial information of each pixel unit and the surrounding pixels, it is also necessary to distinguish from adjacent pixels, which is one of the main difficulties of the current segmentation task. However, high-resolution input images can bring more detailed information, but they are often accompanied by expensive computing costs, so smaller resolution images will be put in practical applications to ensure computing speed. As another task of computer vision, super-resolution recovery focuses on extracting information from low-resolution pictures and reasoning into higher-resolution feature maps. Its recovered detail features contribute to the high-precision classification of semantic segmentation tasks. Considering the complementarity of the two tasks, considering the use of transformer as a feature extractor, the design algorithm realizes semantic segmentation and super-resolution recovery tasks at the same time, multi-task learning can ensure that the backbone network obtains more common high-dimensional information, and then we use the results of super-resolution recovery branches to guide the semantic segmentation task to provide more detailed information and finally obtain an effective improvement on the original baseline.