RGBD image saliency detection method and system based on multi-level fusion

The invention relates to an RGBD image saliency detection method and system based on multi-level fusion. The method comprises the steps of: S1, carrying out data enhancement, and processing a color image, a corresponding depth image and a corresponding manual annotation image; S2, designing a multi-level convolutional neural network structure, extracting and fusing the features of the depth imageand the color image, and fusing the multi-level features to obtain a predicted saliency map Sk of each level; S3, designing a fusion optimization module, and then fusing the saliency maps Sk of all the levels to obtain a final saliency map S; S4, learning optimal parameters of an RGBD image saliency detection model by solving a minimum loss function to obtain a trained RGBD image saliency detection model; and S5, finally, inputting an RGBD image into the trained model, and performing calculation to obtain a saliency detection result, namely a saliency map, of the input image. Accordingto the method, the problem of rou