Abstract: Timely detection and precise localization of road cracks constitute an important prerequisite for performing pavement condition assessment and reducing road maintenance costs. However, most existing crack extraction techniques for road images suffer from limitations in recognition accuracy and generalization capability. An improved DeepLabv3+ based algorithm for crack extraction from road images was proposed. Based on encoder-decoder architecture of original DeepLabv3+, a multi-resolution feature fusion strategy was designed specifically for decoder module to enhance the representation of crack edge details. A strip pooling module was introduced to replace the parallel connections in atrous spatial pyramid pooling module with dense connections, thereby enhancing encoder’s feature extraction ability. Additionally, an attention mechanism was incorporated to improve model’s generalization ability for crack extraction under complex backgrounds and interference. Ablation and generalization experiments demonstrated that the proposed algorithm outperformed current mainstream algorithms in terms of both mean intersection over union (MIoU) and mean pixel accuracy (MPA).