Revisiting Generative Adversarial Networks for Binary Semantic Segmentation on Imbalanced Datasets

Anomalous pavement surface conditions detection aims to detect pixels representing anomalous states, such as cracks, on pavement surface images automatically by algorithms. Recently, deep learning models have been intensively applied to related topics with outstanding performance. However, most existing deep learning-related solutions rarely achieve a stable performance on diverse datasets. To address this issue, in this work, we propose a deep learning framework based on conditional Generative Adversarial Networks for anomalous region detection on pavement images at the pixel level. In particular, the proposed framework is developed to enhance the generator's ability to estimate the probability feature map from heterogeneous inputs with two training stages and multiscale feature representation. Moreover, several attention mechanisms are incorporated into the proposed framework to mitigate the performance deterioration of model training on severely imbalanced datasets. We implement experiments on six accessible pavement datasets. Extensive qualitative and quantitative experiments demonstrate that the proposed framework can achieve SOTA results on these datasets efficiently and robustly.