Unmixing_Convolutional_Features_for_Crisp_Edge_Detection
2 METHODOLOGY
2.2 The Tracing Loss
- 模型训练过程中,LCE进行粗略的边缘学习,LBDRY通过特征分解进行边缘定位细化,LTEX对纹理区域进行较强的整体抑制。
- 实现了清晰的边缘生成,比单加权交叉熵具有更小的定位模糊性
2.2.1 WeightedCross Entropy
式(1)
2.2.2 Feature Unmixing byTracing Boundaries
boundary tracing function:式(2)
texture suppression function:式(3)
2.3 Context-Aware Fusion Block
CoFusion的机制:如图3所示
Z:代表了 L 侧的边缘热图;
W_context:通过注意力模块从Z中学习得出的权重图。
- 具体的说,从Z中提取出得分图 A_score ,通过 3 个 3 x 3 卷积层获取上下文信息用来推断 W_context ;A_score 随后通过 Softmax 激活来规范化,以获得权重图 W_context
P_final 通过下面两步进行计算:式(6)、(7)
3 EXPERIMENTS AND ANALYSIS
3.1 Datasets
- BSDS500:challenging edge detection bench- mark
- NYUDv2:challenging dataset for indoor scene pars- ing and is also a commonly used benchmark for edge detection evalu- ation
- Multicue Dataset: This dataset strictly distinguishes the defi- nitions of boundary and edge, and it thus consists of two sub data- sets: Multicue Boundary and Multicue Edge