Salient object detection has become an important research direction in image processing technology, so for the challenges of inference efficiency and feature fusion in bimodal salient target detection, this paper proposes an improved salient target detection method for multimodal images. Firstly, a parameter-free SimAM module is introduced into the early feature extraction convolutional laver of the CDINet model to emphasize the key features, which is also used to enhance the ability of the subsequent layers to capture the key information. Then, a bidirectional feature fusion (BiFPN) machine is introduced at the decoder stage system, replacing standard convolution with depth-separable convolution to reduce computational effort while achieving efficient feature fusion and improving multi-scale detection. Finally, the model is enhanced to handle unbalanced data by using focus loss and weighted cross-entropy loss function, as well as combining the cross-merge ratio loss function to improve the accuracy of boundary prediction. The experimental results show that the improved model (SblCDINet) improves the F-measure by 1.07%, 1.15%, 1.06%, 0.16%, and 1.44%, reduces the computational complexity of the model by 19.8% over the five datasets compared to the original CDINet model data. Compared with other models, the SblCDINet model not only achieves the improvement of inference efficiency and the enhancement of the feature fusion effect, but also effectively reduces the computational complexity, and the experimental data also verifies the effectiveness and superiority of the method in multi-modal image saliency object detection.
Salient object detection has become an important research direction in image processing technology, so for the challenges of inference efficiency and feature fusion in bimodal salient target detection, this paper proposes an improved salient target detection method for multimodal images. Firstly, a parameter-free SimAM module is introduced into the early feature extraction convolutional laver of the CDINet model to emphasize the key features, which is also used to enhance the ability of the subsequent layers to capture the key information. Then, a bidirectional feature fusion (BiFPN) machine is introduced at the decoder stage system, replacing standard convolution with depth-separable convolution to reduce computational effort while achieving efficient feature fusion and improving multi-scale detection. Finally, the model is enhanced to handle unbalanced data by using focus loss and weighted cross-entropy loss function, as well as combining the cross-merge ratio loss function to improve the accuracy of boundary prediction. The experimental results show that the improved model (SblCDINet) improves the F-measure by 1.07%, 1.15%, 1.06%, 0.16%, and 1.44%, reduces the computational complexity of the model by 19.8% over the five datasets compared to the original CDINet model data. Compared with other models, the SblCDINet model not only achieves the improvement of inference efficiency and the enhancement of the feature fusion effect, but also effectively reduces the computational complexity, and the experimental data also verifies the effectiveness and superiority of the method in multi-modal image saliency object detection.