RGBD image saliency detection method for regenerating three-stream convolutional neural network by using saliency map

Abstract

The invention discloses an RGBD image saliency detection method for regenerating a three-stream convolutional neural network by using a saliency map. The method comprises the following steps: establishing a gate structure form through a double-flow end-to-end network, decoding to generate an initial significance prediction graph, establishing a single-flow light-weight network by taking the initial significance prediction graph as input, and in order to save memory and computing power, guiding single-flow network decoding by a decoding part by utilizing decoding information of the double-flownetwork, wherein the previous information utilizes the previous experience to guide the subsequent feature information, and the saliency prediction graph generated by the single stream and the initialsaliency prediction graph are added by setting the weight to obtain the final saliency prediction graph. Two layers of networks are established to effectively utilize previous information and subsequent information to enhance an initial significance prediction graph, experiments prove that the method is effective, and experiments prove that prediction results on two test sets of the method are clear in boundary, and significant objects are complete in structure.

Description

technical field [0001] The present invention relates to the technical field of salient object detection, and more specifically relates to a RGBD image saliency detection method in which a saliency map is regenerated into a three-stream network. Background technique [0002] With the rapid improvement of computer hardware equipment, the complex calculation of neural network has gradually shifted from CPU operation to GPU operation. Nvidia has successively launched CUDA and other acceleration packages for adaptive optimization, which makes the development of neural network reach an unprecedented height. The in-depth development of neural networks has also brought earth-shaking changes to some computer vision, such as target detection, pedestrian tracking, semantic segmentation, etc. Saliency detection is also the direction of computer vision affected in this trend. Saliency detection is the target scene. Detection of artificial salient regions, which are regions of human inter...

AI Technical Summary

Problems solved by technology

Many previous methods often only use an end-to-end network to generate the final saliency prediction map, but in fact, in the human brain, more than one calculation is performed on the scene seen, and humans tend to optimize what they want to see. The scene received by the eyes is also one of the principles that the human eye is often easily deceived

Images