Salient target detection method based on complementary branch network

Abstract

The invention provides a saliency target detection method based on a complementary branch network, and the method comprises the steps: enabling labels of a plurality of images to pass through a label redefinition module, and obtaining expanded diversified saliency labels; processing the extracted features of the backbone network by an information exchange module to obtain saliency features with discrimination capability; and finally, enabling different features to correspond to different labels respectively, and obtaining a saliency result graph through connected flow loss respectively. According to the invention, through the proposed label redefinition module, given labels are sufficiently expanded, useful priori knowledge is used as auxiliary labels, complementary feature information from different branch structures is fused and captured through the proposed information exchange module, and total loss of branch merging is calculated through proposed connected flow loss. Therefore, the optimization of the whole network is supervised, and the accuracy of saliency target detection of a complex scene is ensured.

Description

technical field [0001] The invention belongs to the field of computer vision, in particular to a salient target detection method based on a complementary branch network. Background technique [0002] In recent years, salient object detection has become a research hotspot in the field of computer vision, and its purpose is to obtain the most salient target area in any given image or video. This technology has a wide range of application prospects, such as: image retrieval, video compression, visual tracking, robot navigation, etc. [0003] Traditional salient object detection methods mainly rely on manually extracted limited information, such as color and stripes. At present, due to the development of deep learning, new salient object detection methods emerge in an endless stream. These methods can be roughly divided into two categories, namely aggregation-based methods and edge-based methods. [0004] Usually, the former extracts multi-layer features through the encoder a...

AI Technical Summary

Problems solved by technology

[0004] Usually, the former extracts multi-layer features through the encoder and fuses them with the decoder to obtain saliency maps, while the latter obtains saliency maps by supervising mask labels with the help of edge label priors, although these methods have achieved remarkable results. progress, but some challenges remain

Specifically, these methods often fail to achieve satisfactory results when facing ambiguity, occlusion, illumination, rotation, and salient object detection at different scales.

Images