Enhanced Semantic Segmentation of Road Cracks and Potholes based on Improved SegNeXt

Automated pavement defect detection is essential for road maintenance, yet remains challenging due to the irregular geometry of cracks and potholes and the severe class imbalance between defect and background pixels. To address these challenges, this paper proposes DS-SegNeXt, an improved semantic segmentation framework built upon SegNeXt, with three principal contributions. First, a geometry-adaptive deformable encoder is introduced by replacing fixed-kernel depth-wise convolutions in the Multi-Scale Convolutional Attention (MSCA) module with Deformable Convolution v4 (DCNv4), yielding a Deformable MSCA (DMSCA) that enables spatially adaptive feature sampling conforming to the irregular morphology of road cracks and potholes. Second, a content-aware decoder is designed by substituting bilinear interpolation with DySample, a lightweight dynamic upsampler that preserves fine-grained defect boundary details during multi-scale feature reconstruction. Third, a composite loss function combining Cross-Entropy, Dice, and Lovász-Softmax losses is formulated to provide balanced, IoU-aligned gradient supervision under severe class imbalance, substantially improving detection sensitivity for minority defect categories. Experiments on a combined CRACK500 and UDTIRI dataset demonstrate that DS-SegNeXt achieves 74.14% mIoU, outperforming the SegNeXt baseline by 1.68% and surpassing U-Net, DeepLabV3+, and SegFormer under identical training conditions.