RGB-D image saliency detection method

Abstract

The invention relates to an RGB-D image saliency detection method, which comprises the following steps of: 1, inputting an RGB image and a depth image into a Resnet50 backbone network in which a full connection layer and a pooling layer are rejected, and extracting RGB features and depth features of each layer; 2, inputting the depth features of each layer into a corresponding depth improvement module, and performing fusion operations such as convolution and activation on the obtained depth improvement features and RGB features to form mixed features of each layer; 3, combining the mixed features in pairs from a high layer to a bottom layer, obtaining saliency maps of all layers through feature grouping supervision fusion modules, wherein the saliency map and a truth value map output by the upper layer are added into each feature grouping supervision fusion module for supervision, and finally the saliency map output by the fourth layer serves as a final result saliency map through iterative optimization. According to the method, noise interference of low-quality depth features is effectively suppressed, high-quality full fusion is performed on the features, a complete and accurate saliency map can be generated, and the method has relatively good detection performance.

Description

Technical field: [0001] The invention relates to the field of computer vision, in particular to a RGB-D image saliency detection method. Background technique: [0002] The key to the salient target detection technology is to extract the most attractive and important areas in the target scene. In recent years, many people have explored the salient target detection technology in the field of computer vision and applied this technology to semantic segmentation, target recognition, and image compression. and image segmentation. In the past few years, various salient object detection models based on RGB-D images have been proposed and all of them have achieved good performance. [0003] In the RGB-D image, the RGB image and the depth image appear in pairs. The RGB image provides detailed color texture information, and the depth image provides many spatial information such as the shape and position of the target area. In the past methods, the depth image is often used as the fou...

AI Technical Summary

Problems solved by technology

[0005] In the process of realizing the present invention, the inventor found that the prior art has at least the following shortcomings and deficiencies: (1) due to the limitation of the collection equipment, there will be edges in the data set Low-quality depth images that are blurred or disturbed by noise, the introduction of low-quality depth images reduces the detection performance; (2) The features containing noise are passed down, and useful features cannot be fully acquired and fused, and useful features are lost and wasted, thus affecting result quality

Images