A crack detection algorithm based on parallel extraction and attention fusion network is designed to address the problems of low accuracy, much noise and detail loss of existing methods for crack detection. First, the high and low-level features of the crack scene are extracted using multi-scale convolutional parallel neural networks with different depths; then, to improve the detection accuracy, the high and low-level features of the crack scene are effectively fused with the pixel attention mechanism for the features of the crack scene to obtain effective fusion features for crack detection; finally, the crack detection results are output using nonlinear mapping. The experimental results show that the proposed algorithm can obtain effective features for high-precision detection results, the details of crack detection results are clearer, and the supervised learning approach largely eliminates the noise interference of detection results and obtains detection results with better visual effects; the proposed algorithm also has good performance in the evaluation of quantitative indexes such as accuracy rate and recall rate, and the accuracy rate of crack detection reaches 85%.