添加L1正则来约束BN层系数
The semantic edge information can improve the performance of salient object detection. Specifically, semantic edge information enhances performance in the following aspects:
1. Clear edge structure: By embedding edge prior knowledge into the network, ENFNet better maintains the boundary clarity of salient objects. Traditional fully convolutional networks (FCNs) tend to blur spatial structures and edges due to successive strides and pooling operations, whereas ENFNet effectively embeds edge information into hierarchical feature maps through edge guidance blocks.
2. Precise saliency region localization: Edge guidance blocks not only perform feature operations but also spatial transformations to achieve effective edge embedding. This helps in more accurately locating the boundaries of salient objects in saliency detection.
3. High-quality saliency map generation: ENFNet generates saliency maps with high-quality boundary awareness, thanks to the network's layered embedding of detailed edge information.
4. Performance improvement: ENFNet achieves best-in-class performance across all datasets compared to existing state-of-the-art methods, indicating that edge information is effective in improving the accuracy of saliency detection.
5. Boundary preservation: By using IoU boundary loss, ENFNet further optimizes the accuracy of saliency boundaries. This loss function calculates the difference between the true boundary and the predicted boundary, aiding in the generation of saliency maps with clear boundaries.
6. Multi-scale feature fusion: Through a hierarchical edge-guided non-local structure, ENFNet combines local contrast features and global context features, enhancing its ability to detect salient objects.
BSDS500 轮廓检测与语义分割数据集_数据集-阿里云天池
Learning Spatial Context: Using Stuff to Find Things (stanford.edu)
阅读了三篇关于基于深度学习的边缘检测的综述论文，发现了12篇关于更流行的算法的论文。整理了模型架构和指标信息。
After reading three review papers on deep learning-based edge detection, I discovered 12 papers on more popular algorithms. I have organized the model structures and metric information.