Image engine optimizes the image signal through a variety of specific algorithms, which plays an extremely important role in the display system. The traditional color optimization algorithm of image engine is composed of various image optimization algorithms, which can not optimize images adaptively and easy to amplifies the noise. Therefore, a full convolution neural network based on dilated convolution is proposed to construct the optimization algorithm, which focuses on optimizing images from the perspective of subjective perception. At the same time, a large-scale dataset is constructed to improve the generalization ability of the algorithm and prevent overfitting. The experiment results show that the proposed algorithm can effectively enhance the color of original images. Compared with the traditional method, the average peak signal-to-noise ratio is improved by 4.01 dB and the average structural similarity is improved by 0.04. The subjective comparison experiment shows that the proposed algorithm has a significant impact on the subjective perception quality of the image, with an average improvement of 61%.