RGB-D saliency detection method based on dynamic filtering decoupling convolutional network

Abstract

The invention discloses an RGB-D saliency detection method based on a dynamic filtering decoupling convolutional network. The method comprises the following steps: acquiring an RGB image tensor and a depth image tensor; respectively inputting the RGB image tensor and the depth image tensor into an encoder network to extract a single-mode feature group, and generating an RGB feature group and a depth feature group according to the characteristics of the encoder network and hierarchical division; respectively putting the features of the RGB feature group and the features of the depth feature group into a respective modal-specific overall guide dynamic enhancement module (MGDEM), and performing single-modal specific feature enhancement; inputting the enhanced RGB features and depth features into a scene perception cross-modal dynamic fusion module (SCDFM), and performing feature fusion between modals, wherein the MGDEM and the SCDFM are both based on a decoupled dynamic filtering convolution structure; and inputting the fused features into a decoder to obtain a predicted significance result. According to the method, accurate saliency prediction is realized.

Description

technical field [0001] The invention relates to the technical field of computer vision, in particular to an RGB-D saliency detection method based on a dynamic filter decoupling convolutional network. Background technique [0002] The goal of saliency detection is to find the object or region that most attracts human visual attention in an image. This task plays a preprocessing role in different vision tasks, allowing the computer to first detect areas of interest to humans, and then perform subsequent processing on the detected areas. For example, when performing target recognition, it is not necessary to traverse the entire image with a sliding window, but to perform saliency detection first, and then only recognize the target in the saliency area. Accurate and reliable saliency detection can save computation while increasing accuracy, benefiting many image tracking and recognition processing tasks in visual graphics. [0003] Many current methods perform saliency detecti...

AI Technical Summary

Problems solved by technology

However, reusing fusion modules inevitably incurs additional computation and storage consumption

This further extends the challenge of cross-modal feature fusion

Images