Low-resolution (LR) grayscale images with multi-wavelength information are difficult to fully demultiplex. High-resolution (HR) colored images reconstructed from LR images are prone to channel crosstalk. To reconstruct HR colored images that are not prone to channel crosstalk, we propose an HR colored image reconstruction algorithm based on a three-dimensional convolutional neural network (CNN). The principal component analysis method is used to extract structural information from HR monochromatic images and LR colored images, and then the CNN is trained based on the structural information to establish a mapping relationship between the HR monochromatic image and LR colored image. Consequently, a HR colored image is generated. The experimental results show that the proposed algorithm can obtain HR colored images without channel crosstalk and color distortion. The quantitative evaluation indexs show that the root mean square error and structural similarity parameter are less than 0.1 and greater than 0.9, respectively.