High-Quality Salient Object Detection and Optimization Method Based on Optical Sensor Data
Jing Jie, Jianzhong Cao, Gaopeng Zhang, Hao Ye, Dongyu Lu, Yujie Qiu, Xin Ma, Feng Wang · IEEE Sensors Journal · 2025
Salient object detection (SOD) is critical for computer vision tasks such as image segmentation and object tracking. A key challenge in SOD is the effective representation of the semantic properties of objects. However, current saliency models have large network sizes and computational costs, resulting in saliency maps with issues such as blurriness, uneven brightness, and loss of edge details. In this paper, we propose a new salient object detection and optimization method. First, we design an SOD network that includes a multi-scale feature extraction (MSFE) module and a multi-level feature fusion (MLFF) module to achieve thorough integration of features from different levels and multi-scale contextual information. Next, we perform exponential fusion between the generated saliency map and the saliency map produced by the DIPONet model to enhance the completeness of object detection. Finally, we design an edge detection network that includes the CFRL loss function and propose a feasible edge-aware optimization module to obtain saliency maps with uniform brightness and smooth edges. We use the NASNet-Mobile network as the backbone, which is pre-trained on the ImageNet dataset. Experimental results on four challenging optical sensor datasets—SOD, PASCAL-S, DUTS-TE, and DUT-OMRON—demonstrate that our method outperforms other approaches.