At present, a large number of buildings in the world are threatened by wind and rain erosion and aging and inevitably crack. If not maintained in time, it will have a negative impact on the building and even endanger the structural safety and shorten the service life of the building. Due to the random shape and irregular size of cracks, the generalization and robustness of crack detection models still need to be improved. In this paper, the object detection algorithm based on deep learning takes YOLOv8 as the benchmark model to optimize the problems of excessive computation, low recall rate and mean average precision. The model further improves the EGAP-YOLO model from four aspects: feature extraction network, feature fusion network, detection head and loss function. This model proposes a new EGC attention module, which enables feature extraction network to extract more feature information. The feature fusion network is replaced by the midAFPN feature fusion network, and the dimensionality reduction operation before fusion is fine-tuned to retain more features of deep channel and suppress the features of shallow channel. In terms of detection head, the convolutional blocks are reduced, the model complexity is reduced, and the model generalization is improved. In terms of the loss function, replacing the original CIoU with WIoU v2 has strengthened the focus on the ground truth of common quality. Compared with YOLOv8n, the recall rate of this model increases by 4.6%, the mAP@50 index increases by 6.3%, the calculation volume decreases by 41.5%, and the number of parameters only increases slightly. It provides a new method for crack detection.