부산대학교 도서관

In the domain of railway safety, effective rail surface defects detection remains a pivotal challenge. This paper presents a novel convolutional architecture named Depthwise-Dilated Convolution (DDC) aimed at efficiently classifying rail surface defects. Leveraging the benefits of both depthwise and dilated convolutions, our DDC model showcases significant advantages in terms of computational efficiency and accuracy. Experiments on a comprehensive dataset of 4,698 rail surface defects images, sourced from passenger railways in Northwest China, resulted in an impressive test accuracy of 94.43%. Comparative benchmarks against state-of-the-art models, including CNN, CNN-SPP, Fractal Convolution, Transpose Convolution, and MobileNetV2, further substantiate the superiority of our approach. Through rigorous feature map evaluations, we illustrate the DDC's enhanced capability in discerning intricate rail defects features. Our findings position the Depthwise-Dilated Convolution as a frontrunner for real-world applications, especially in computational-constrained environments.