A segmentation network of Tiny cracks (TCS-Net) is proposed to solve the problems of small cracks with few pixels, low contrast between cracks and background, much interferences from similar textures, and uneven illumination. TCS-Net has a coding-decoding structure similar to that of U-Net network. The soft Pooling was used in down-sampling process to preserve the basic features of original images while maintaining the amplification of high-response features, thus reducing information loss and preserving the details and location of crack edges. In the up-sampling process, a semantic compensation module combining channel attention and spatial attention was added at the decoding end to fuse the features of each layer at the coding end, which could enhance the multi-scale details of cracks.In view of the classification problem of unbalanced data in crack segmentation task, in order to avoid the training process being dominated by most classes (background pixels), the binary cross entropy loss and Dice coefficient were combined by TCS-Net model as the objective loss function to solve the problem of training instability caused by attention tendency of single loss. It can also optimize performance indicators such as accuracy, crossover ratio and recall rate. Compared with the existing mainstream semantic segmentation models, TCS-Net fracture segmentation model improves the intersection over union ratio index by 5%~9% and recall ratio index by 9%~13%,which demonstrates that the model achieves higher detection rate and detection accuracy. The proposed TCS-Net can be effectively applied to the fine crack segmentation task under the conditions of serious imbalance between target and background, complex background with many disturbances