Lightweight saliency target detection algorithm with semantic-guided feature fusion
Runhao Wang, Shuang Chen, Lin Luo, Yi He · 2025
U-shaped structures are widely used in the design of salient target detection networks. However, this structure commonly suffers from the problems of losing spatial location details and difficulties in capturing edge details, and is usually accompanied by an excessive amount of model parameters. To address these problems, this paper proposes a lightweight saliency target detection network with deep semantic information-guided feature fusion. First, the skip connections in the outer layer of the network are redesigned so that they can fuse different scales of feature information in this layer and all shallower layers above, thus enhancing the network's ability to capture edge details. Second, an MCA module is incorporated into the residual U-block to handle the last layer of features of the feature extraction network, to enhance its representational power and to serve as a semantic guide in the decoding process, facilitating the fusion of features between the decoding side and the encoding side. Finally, a depth-separable convolution is used to replace the traditional convolution in order to reduce the computational and parametric quantities of the network. The experimental results show that the proposed algorithm achieves excellent results in accuracy, precision, recall, mean-cross concurrency ratio, and F1 score, which proves that the algorithm has a better detection performance with more obvious boundaries.