RGB-D saliency detection method based on asymmetric double-current network architecture

Abstract

The invention discloses an RGB-D saliency detection method based on an asymmetric double-flow network architecture. The method comprises the following steps: respectively obtaining input tensors IRGB and ID based on an RGB image in an RGB-D data set and a corresponding Depth depth map; inputting the input tensors IRGB and ID into an asymmetric double-flow network architecture to obtain multi-scale coding features based on RGB and Depth, wherein in the asymmetric double-flow network architecture, an RGB flow network further comprises a flow step module on the basis of a VGG, and four detail information transmission branches are adopted, and the Depth flow network adopts a detail information transmission branch; fusing the extracted depth features into the RGB flow through a depth attention module to obtain complementary features with rich position information; and performing feature decoding on the obtained complementary features through a decoder to obtain a final significance prediction result. According to the RGB-D saliency detection method based on the asymmetric double-flow network architecture, the RGB-D saliency detection model based on the asymmetric double-flow network architecture is constructed, the inherent difference between RGB and Depth data is fully considered, and an accurate prediction result can be obtained for many challenging scenes.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to an implementation method of RGB-D saliency detection based on an asymmetric two-stream network architecture. Background technique [0002] Saliency detection refers to the identification of the most attractive areas and objects in an image that can attract the user's visual attention and the most eye-catching. Due to the selection of the most visually characteristic information in the scene, it is widely used in computer vision. The application has attracted widespread attention. With the maturity of the salient object detection algorithm, its application has been involved in more and more industrial or academic fields. In the industrial world, for example, in the life scene, the product scene is photographed by a mobile phone or other camera equipment, and then processed to obtain the detailed information of the product concerned. In academia, for example, saliency detection is ...

AI Technical Summary

Problems solved by technology

[0004] The 2D saliency detection method is the most common saliency detection method based on static images. It uses image contrast, color, texture and other information for detection. Although it has achieved good detection results on the existing 2D saliency detection database , but appearance features in RGB data are less effective in predicting some challenging scenes (such as multiple or transparent objects, similar foreground and background, complex background, low-intensity environment, etc.)

Therefore, a symmetric RGB-D dual-stream network may ignore the inherent differences in RGB and depth data, resulting in false or false detections, etc.

In addition, existing RGB-D methods inevitably cause loss of detail information when adopting the stride and merge operations adopted in RGB and depth flow networks

Images