RESUMEN
In long-term use, cracks will show up on the road, delivering monetary losses and security hazards. However, the road surface with a complex background has various disturbances, so it is challenging to segment the cracks accurately. Therefore, we propose a pavement cracks segmentation method based on a conditional generative adversarial network in this paper. U-net3+ with the attention module is used in the generator to generate segmented images for pavement cracks. The attention module highlights crack features and suppresses noise features from two dimensions of channel and space, then fuses the features generated by these two dimensions to obtain more complementary crack features. The original image is stitched with the manual annotation of cracks and the generated segmented image as the input of the discriminator. The PatchGAN method is used in the discriminator. Moreover, we propose a weighted hybrid loss function to improve the segmentation accuracy by exploiting the difference between the generated and annotated images. Through alternating gaming training of the generator and the discriminator, the segmentation image of cracks generated by the generator is very close to the actual segmentation image, thus achieving the effect of crack detection. Our experimental results using the Crack500 datasets show that the proposed method can eliminate various disturbances and achieve superior performance in pavement crack detection with complex backgrounds.