DeepLabV3+ based lightweight segmentation for road cracks

ObjectiveTraditional road crack segmentation methods have two core problems, i.e., low segmentation accuracy and large number of model parameters. This study proposes a lightweight and efficient segmentation network for road cracks, aiming to balance accuracy and lightweight performance.MethodIt constructed a lightweight crack segmentation network based on the improved encoder-decoder structure of DeepLabV3+. The core improvements consisted of two aspects. First, replacing the original model's feature extraction network Xception with the lightweight MobileNetV3, which combined depthwise separable convolutions with channel attention mechanisms. This combination could extract crack features efficiently while reducing the number of model parameters. Second, embedding an efficient channel attention module to the output end of feature extraction network. This module adaptively assigned weights to each channel, thereby enhancing the network's focus on crack pixel features and improving the segmentation accuracy. The road crack images were the research carrier through entire network. It completed crack segmentation through feature extraction, attention enhancement, and encoding-decoding mapping.ResultThe proposed network improved the mean intersection over union by 11.6% on the Crack Forest Dataset compared with traditional SegNet method; and its F1-score was improved by 5.4%. The mean intersection over union of the proposed network was 1.1% higher than that of SegNet on the dataset CRACK500; and its F1-score was 17.5% higher than that of NestNet. The ablation test on dataset CRACK500 indicated that the proposed network improved the mean intersection over union by 0.7% compared with the original DeepLabV3+; and its F1-score was improved by 0.6%. Meanwhile, the model parameters were reduced by 89.5%; and the training time was shortened by 64.1%. These results verified the lightweight nature and effectiveness of the proposed network.ConclusionThe core innovation of this study lies in two aspects. One is the replacement of MobileNetV3. The other is the embedding of efficient channel attention mechanism, realizing the lightweight optimization on DeepLabV3+ network. The findings solve the contradiction between accuracy and parameter quantity in traditional methods.