JOINT SEGMENTATION MODEL FOR CRACKS AND JOINTS BASED ON Deeplabv3+

This paper addresses the problem of low segmentation accuracy of cracks and joints in complex scenarios and proposes an improved model, ViR-Deeplabv3+, based on DeepLabv3+. First, the model replaces the traditional backbone network with the Vision Transformer (ViT) with global perception ability. This enables the model to no longer be limited to extracting local information when processing image data, but to capture the global context features of the image more efficiently, thereby enhances subsequent segmentation tasks. Secondly, residual connections between the ViT and the Hollow Space Pyramid Pooling (ASPP) module are ingeniously introduced. The design concept of residual connection effectively solves problems such as the gradient vanishing problem, ensuring that the rich feature information from ViT can be smoothly and unobstructedly transmitted to the ASPP module for further fusion and mining of multi-scale features. Finally, we conducted model training and ablation experiments based on the self-built dataset (including crack and seam samples). The results showed that the mean intersection and union ratio (mIoU) of ViR-Deeplabv3+ reached 75.27%, which was 2.97% higher than that of the baseline model Deeplabv3+. This scheme provides an effective solution for precisely detecting and segmenting cracks and joints in complex scenarios, and has important practical application value.