RGB-D saliency detection based on multi-scale feature fusion

The introduction of the depth map provides a wealth of position clues for RGB saliency detection, but low-quality depth maps will misguide the model's feature fitting, and due to the large changes in the scale of salient objects in the real world, the network will be in the process of prediction It is more difficult and the error becomes larger. To solve the above problems, this paper designs a new RGB-D saliency detection model based on deep learning. This paper uses VGG19 as the backbone network to extract the features of the two modalities of the RGB map and the depth map; then uses the serial adaptive fusion module to perform cross-modal fusion of the extracted features, so that the advantages of the RGB map and the depth map complement each other. Automatically select depth features; then use the multi-scale feature aggregation module of joint edge detection to fuse the cross-modal fusion features with edge information; finally, use the global guidance module to guide the model with global features to obtain the prediction result. Using this method to predict the images on 4 public datasets, and compared with 6 different methods, the prediction result of this method is closer to the artificially calibrated truth map. PR (Precision-Recall) curve, S (S-measure) index, F (F-measure) index and MAE (Mean Absolute Error) index show that the overall performance of the method in this paper is higher than that of the six methods.