Three-modal salient object detection method and system based on mamba dual attention

Abstract

The method comprises the following steps: extracting multi-scale feature maps of each modality respectively; in each level, inputting three modality features into a dynamic weighting module for preliminary fusion and spatial attention enhancement processing, and obtaining re-weighted fusion features according to a dynamic weight generation mechanism; inputting the re-weighted fusion features into a double attention enhancement module, capturing long-distance dependency relationships by using visual state space block features, and obtaining double attention enhancement features based on a constructed channel attention branch and a spatial attention branch; inputting the double attention enhancement features into an edge guiding auxiliary module, capturing multi-scale boundary information by using a boundary branch through discrete wavelet transform, and strengthening global structure by using a global main branch through dense connection convolution, and obtaining fusion features; and generating multi-size prediction maps by using a target detection network according to the fusion features.

Images