SSS-YOLOv8: Track Crack Detection Based on Attention Mechanism and Convolutional Neural Networks

In practice, detecting track cracks is a challenging problem due to different sizes of cracks and complex backgrounds. In this paper, an improved YOLOv8n model is proposed. First, the SPD-Conv module is introduced into the model, which improves the model's detection accuracy of track cracks, especially fine cracks, by replacing the traditional jump convolution and pooling layers. Then, the Second-order Channel Attention module is introduced, which better utilizes the complex relationship between features and improves the generalization ability of the model in various complex contexts. Finally, the Selective Kernel Attention module is introduced, which adaptively adjusts the size of the receptive field through the input features to better handle features of different scales and further improve the detection performance. The experimental results show that the mean average precision(mAP) of the improved model is 3.4% higher than that of the original model. Therefore, the improved model can be better applied in detecting track cracks in actual detection.