Crack detection is an important indicator of structural safety assessment. Traditional crack detection methods based on image processing have the disadvantages of large noise and low accuracy in a tunnel environment with uneven illumination and serious noise pollution. Aiming at this problem, a tunnel crack recognition algorithm based on machine vision is proposed. First, the tunnel image is filtered in frequency domain and is differenced in spatial domain to enhance the image texture features; then the image segmented by the above steps is extracted by setting the area parameter Tv, the saturation parameter Ts and the special parameters Tv`, Ts` to extract the background noises and delete them, in order to detect the complete tunnel crack images; finally, combining the non-mutation and development regularity of the application scenario in this paper, designing a lightweight crack connection algorithm to connect the breakpoints in crack images to avoid the undetected phenomenon. Experiments show that the images processed in this paper can effectively extract complete cracks, and the accuracy of image recognition of tunnel cracks reaches 94%, the recall rate reaches 98%, and the detection accuracy can meet the actual engineering needs.