基于语义信息的精细化边缘检测方法

Refined Edge Detection Method Based on Semantic Information

边缘检测是在图像中准确地提取视觉上显著的边缘像素,以得到图像的边缘信息,然而传统基于全卷积网络的边缘检测方法通常存在预测边缘粗糙、模糊等问题。提出一种语义信息指导的精细化边缘检测方法。通过图像分割子网络将学习到的图像语义信息传递给边缘检测子网络,同时利用图像语义信息指导边缘检测子网络,其引入具有注意力机制与残差结构的特征融合模块,以生成精细的图像边缘,增强不同尺度的特征融合。在此基础上,结合图像分割任务和图像边缘检测任务中的代价函数定义新的模型代价函数并进行训练,进一步提高网络边缘检测质量。在BSDS500数据集上的实验结果验证了该方法的有效性,结果表明,该方法的固定轮廓阈值与图像最佳阈值分别达到0.818和0.841,相比HED、RCF等主流边缘检测方法,能够预测更精细的边缘图像,且鲁棒性更优。

Edge detection is to accurately extract visually significant edge pixels from the image to obtain the edge information of the image. Traditional edge detection methods based on Full Convolution Network(FCN) usually require rough and fuzzy edge prediction.This paper proposes a refined edge detection method guided by semantic information.The learned image semantic information is transmitted to the edge detection subnetwork through the image segmentation subnetwork.The image semantic information is used to guide the edge detection subnetwork.A feature fusion module with attention mechanism and residual structure is also introduced to generate fine image edges to enhance feature fusion at different scales.On this basis,the cost function in image segmentation task is combined with the image edge detection task,to define a new model cost function,which is further trained to improve the quality of network edge detection. The experimental results on the BSDS500 dataset verify the effectiveness of the proposed method.The optimal dataset scale and image optimal scale attained by this method are 0.818 and 0.841,respectively. Compared with mainstream edge detection methods,such as HED and RCF,the proposed method can predict finer edge images with improved robustness.