Visual Saliency Extraction Method of Stereo Image Based on Parameter Sharing Deep Learning Network

Abstract

The invention discloses a stereoscopic image visual saliency extraction method based on a parameter sharing deep learning network, which constructs a parameter sharing deep learning network in the training stage, which includes an input layer, a parameter sharing feature extraction framework, a saliency map generation framework, and a parameter sharing feature The extraction framework consists of the first resnet-50 convolutional network block, the second resnet-50 convolutional network block, the third resnet-50 convolutional network block, and the fourth resnet-50 convolution in the resnet-50 network The network block and the fifth resnet-50 convolutional network block are set up in sequence, and the color map feature and the disparity map feature are extracted by using the parameter sharing feature extraction framework; in the test phase, the parameter sharing deep learning network training model is used to train the left side of the stereo image to be tested. The viewpoint image and the left disparity image are predicted, and the human gaze prediction map is obtained, which is the visually salient image. The advantage is that the extracted stereoscopic features conform to the salient semantics, and it has strong extraction stability and high extraction accuracy.

Description

technical field [0001] The invention relates to an image visual salient extraction technology, in particular to a stereoscopic image visual salient extraction method based on a parameter sharing deep learning network. Background technique [0002] Image saliency detection is a technique to search for more important regions in an image, most of which are used in the image preprocessing stage to find the image parts to be processed preferentially, and it is a bionic mechanism that imitates the human visual mechanism. Stereo vision saliency detection is an extension of image saliency detection, and the problem it faces is how to apply additional depth information to assist saliency detection. [0003] The classic method of stereo vision saliency detection using manual features directly uses filtering and other methods to extract depth features when using depth information, such as the stereo vision saliency detection method proposed by Qi Feng et al. This method first converts...

AI Technical Summary

Problems solved by technology

Although this type of method can quickly obtain a visual saliency map with a certain quality, due to the limited input features, the ability to distinguish visual saliency is limited and cannot cope with complex scenes.

Images