A light field salient object detection method based on multimodal feature fusion

Abstract

This invention discloses a method for detecting salient targets in a light field based on multimodal feature fusion. The method uses a trained light field salient target detection model to detect the saliency map of a multimodal light field image. The model enhances the features between focal slices locally through a slice interleaving enhancement module to obtain enhanced focal flow features. A high-level feature fusion module fuses high-level multimodal semantic information features locally and globally, outputting a fused high-level feature map. Low-level cross-attention fuses low-level multimodal spatial-channel features through cross-enhancement, outputting a fused low-level feature map. Finally, a compact pyramid refinement module aggregates the above features and decodes them into an accurate saliency map. This invention solves the problems of insufficient feature learning between slices in the focal stack and insufficient fusion of multimodal information features in the light field by constructing a light field salient target detection model.

Images