TCS-Net：一种核电安全壳细小裂缝分割网络

针对核电安全壳表面裂缝视觉检查任务面临的裂缝细小且占像素少、裂缝与背景对比度低、相似纹理干扰多、光照影响等问题，作者提出了一种细小裂缝分割模型TCS-Net(Segmentation network of tiny cracks)。该模型是编码——解码的网络结构，在下采样过程使用Soft Pooling减少编码过程池化导致的信息损失以保留图像边缘细节及位置信息；解码端在下采样过程中通过加入兼顾通道注意力和空间注意力的语义补偿模块(ResCRAM)以融合编码端的各层特征，可增强裂缝的多尺度细节信息；结合Bce(Binary Cross-Entropy)损失和Dice损失作为目标损失函数，以解决单一损失关注度倾向带来的训练不稳定的问题，也可优化Acc(Accuracy)、IOU(Intersection over Union)、Recall等性能指标。为了验证模型的有效性，在真实的安全壳图像对所提裂缝分割模型进行了测试。实验结果表明，与现有的主流语义分割模型相比，TCS-Net裂缝分割模型的IOU指标可提高5%-9%，Recall指标可提高9%-13%，由此说明该模型具有检测率和检测精度更高，能有效适用于目标与背景严重不平衡、背景复杂且干扰较多情况下的细小裂缝分割任务。

TCS-Net: A Tiny Crack Segmentation Network for Nuclear Containment Vessel

A Segmentation network of tiny cracks (TCS-NET) model is proposed to solve the problems faced by the visual inspection task of cracks on the surface of nuclear containment vessel, such as small cracks with few pixels, low contrast between cracks and background, much interference from similar textures, and the influence of illumination. The model is an encoding-decoding network structure. Soft Pooling is adopted in the downsampling process to reduce the information loss caused by the Pooling of the coding process to retain the details and location of the image edge. The multi-scale details of cracks can be enhanced by adding a semantic compensation module (ResCRAM) that combines channel attention and spatial attention to fuse the features of each layer at the decoding end. Combine Bce (Binary cross-entropy) loss and Dice loss as objective loss functions to solve the problem of training instability caused by a single loss of attention propensity. Acc (Accuracy), IOU (Intersection over Union), Recall and other performance indicators can also be optimized. In order to verify the validity of the model, the proposed fracture segmentation model was tested on a real containment image. Experimental results show that compared with existing mainstream semantic segmentation models, The IOU index of TCS-NET fracture segmentation model can be improved by 5%-9%, and the Recall index can be improved by 9%-13%, indicating that the model has higher detection rate and detection accuracy. It can be effectively applied to the fine crack segmentation task under the condition of serious imbalance between target and background, complex background and many disturbances.