Significant progress has been made in salient object detection (SOD) based on deep neural networks. However, existing SOD methods are primarily designed for low-resolution image inputs, which often suffer from issues such as sampling depth, receptive field, and model performance when applied to high-resolution image inputs. To tackle this issue, this study introduces a lightweight feature extraction model by replacing the original ResNet network structure with a RepVGG-based model that incorporates the Efficient Channel Attention (ECA) module for lightweight purposes. In order to enhance both the model's accuracy and processing speed, we introduce the Effective Squeeze-and-Excitation (ESE) module for feature fusion. To tackle the issue of unclear boundaries of salient objects, we fuse the Weighted Binary Cross-Entropy, Structural Similarity (SSIM), and Shape-aware Loss into a combined loss function, which replaces the conventional cross-entropy loss. Experimental results demonstrate that the enhanced algorithm (RepPGNet) achieves a 2.8% increase in accuracy compared to the original algorithm, with reduced model parameters and improved clarity of salient object boundaries. The proposed algorithm is also shown to have improved speed and is suitable for high-resolution image scenes.
分享
引用
