The invention discloses a multi-resolution deep network image highlight removing method based on divide-and-conquer. The method comprises a training method and a testing method. Firstly, a highlight network removing model is constructed, and the model is composed of a pyramid structure, a nested residual network and a fusion structure. A pyramid structure uses a Laplace pyramid to grade image blocks, highlight is processed on different levels, a convolutional network and a residual network are adopted in a nested residual network to extract features of the image blocks of different levels, anda fusion structure is combined with output of the nested residual network to predict a non-highlight image. And after model training is completed, directly partitioning the test image, predicting a non-highlight image by the model, and finally splicing prediction results to obtain a non-highlight whole image. According to the model structure, the highlight phenomenon in the image can be efficiently removed in real time, and the model structure has high adaptability and high robustness to images with complex colors and textures.