DWDM swin transformer-based multi-scale sampling aggregation network for defect image segmentation of composite materials

Composite materials play a crucial role in aircraft manufacturing, as their performance and reliability directly affect the safety and efficiency of the aircraft. However, in practical applications, many composite material image segmentation tasks face problems such as highly similar shapes and colors between different types of materials, defects, and peripheral components, imposing significant limitations on the improvement of segmentation performance and thus restrict the development of composite material defect detection technology. Therefore, this paper proposes a multi-scale sampling aggregation network based on Dynamic Window Downsampling module (DWDM) Swin Transformer, which uses Swin Transformer as the basic architecture and introduces two innovative modules: DWDM downsampling module and cross entropy loop integral loss function. The DWDM downsampling module effectively reduces the dimensionality of the feature map by dynamically adjusting the window size, reducing the complexity of the model and the risk of overfitting. By reducing the spatial dimension of the feature map, downsampling significantly reduces the computational complexity of subsequent layers, thereby significantly improving the computational efficiency of the model. The cross entropy loop integral loss function is an innovation based on the traditional cross entropy loss function. By introducing the cyclic integration method, this loss function not only considers pixel level prediction accuracy, but also takes into account the cumulative effect along specific paths along the image boundary. This innovative loss function design makes the model more accurate in handling boundaries and complex shapes, effectively solving the problems of inaccurate edge segmentation and low segmentation accuracy in traditional methods. The experimental results indicate that the DWDM Swin Transformer network achieves advanced performance. Compared with U-Net, YOLOV8 and other networks, the detection rate of this network has increased by 3.2%, and the missed detection rate and false detection rate have been reduced by 1% and 2%, respectively.