Classification of pothole pavement based on pseudo-sample generation augmentation

Potholes can be a nuisance to the vehicle and can affect the decision-making of the intelligent driving system. However, most autonomous driving algorithms are currently trained using datasets collected from normal road conditions, as datasets containing pothole roads are scarce. This limitation reduces the robustness and sensitivity of autonomous driving algorithms in recognizing pothole pavement. To address the aforementioned limitation, we propose an Pseudo-Samples Generation strategy based on Improved Cycle Generative Adversarial Network (PSG-ICGAN) to enhance the model’s accuracy in recognizing pothole pavement. In PSG-ICGAN, we enhance the adversarial sample generation algorithm of CycleGAN to produce pseudo-samples with highly similar semantic information to the original images, yet indistinguishable by the classification model. Adding these challenging pseudo-samples to the training dataset significantly enhances the model’s robustness in recognizing pothole road images. In addition, we introduce a channel attention mechanism into the pothole classification model, which helps the model to capture subtle pothole features. Experiments show that our model achieves superior performance in the pothole road dataset.