RGB-D saliency object detection method based on bilateral attention mechanism

Abstract

The invention relates to the technical field of image processing, in particular to an RGB-D saliency object detection method based on a bilateral attention mechanism. The method includes: S10, cross-modal feature extraction, S20, salient region preliminary prediction, S30, prediction graph up-sampling, S40, bilateral attention-based residual generation and S50 residual compensation, and S60, the steps S30 to S50 are repeated, wherein the S10 cross-modal feature extraction comprises the step of extracting features of different levels of the RGB image and the depth map by using a neural network.The objective of the invention is to solve the performance bottleneck problem caused by indiscriminately learning from a foreground and a background based on an RGB-D method of deep learning. According to the method, an RGB-D saliency object detection method (BiANet) based on a bilateral attention mechanism is designed; the BiANet provided by the invention effectively learns the characteristics of the salient object from the foreground and the background through a bilateral attention mechanism, and experiments prove that the RGB-D salient object detection method surpasses the foremost edge onthe six public data sets.

Description

technical field [0001] The invention relates to the technical field of image processing, in particular to a RGB-D salient object detection method based on a bilateral attention mechanism. Background technique [0002] The purpose of saliency detection is to detect the most attractive object in a scene, which is widely used in many visual tasks such as visual tracking and image segmentation; most current methods focus on predicting salient objects from RGB images, these The method is easily disturbed by factors such as the similar color of the foreground and the background and the strong color contrast inside the foreground object, resulting in detection errors; with the popularity of low-cost depth sensors, the depth map of the scene is more and more easy to capture; the depth map provides The three-dimensional spatial relationship of the scene can effectively assist salient object detection algorithms to avoid ambiguity caused by foreground and background colors; therefore,...

AI Technical Summary

Problems solved by technology

[0004] The object of the present invention is to provide a RGB-D salient object detection method based on a bilateral attention mechanism, so as to solve the huge difference in the distribution of the existing foreground and background proposed in the above background technology, and learn from it without distinction. However, in the traditional methods, some methods propose the strategy of inferring salient regions from the foreground and background respectively, but in the deep learning-based method, this simple and effective idea has not been paid attention to.

Images