Image colorization is one of the core problems of great interest in computer vision and has attracted more and more researchers in recent years. The colorization technique improves the recognition rate and scene comprehension of targets of human eyes in grayscale images, especially in low-light-level (LLL) images. However, the current colorization methods still have problems such as color bleeding, color deviation and boundary blurring. To solve the above problems, a colorization method for grayscale images combined with semantic segmentation was proposed. A colorization network composed of a classification sub-network and a semantic segmentation sub-network was designed to fuse the features extracted from the semantic segmentation sub-network and the classification sub-network into the colorization network, and guide the colorization network to assign an accurate color to each object and background in the image. Then the attention mechanism combining with the space and channel was introduced into the semantic segmentation sub-network to obtain a more accurate semantic segmentation results, thereby improving the perception ability of image boundaries and the processing ability of details of the colorization network. To verify the effectiveness of the proposed method, several sets of LLL images of different scenes were captured under LLL environment and colorized by multi-pixel photon counter (MPPC) experimental platform. The experimental results show that the peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image patch similarity (LPIPS) of the proposed method is improved by 6.12%, 11.65%, and 4.99%, respectively.