基于RCF的细化边缘检测模型

Refined edge detection model based on RCF

边缘检测是图像处理和计算机视觉中的基本问题。针对基于深度学习的边缘检测技术(如RCF网络)存在生成的边缘线模糊粗糙及边缘信息不全等问题,本文提出一种基于RCF网络的细化边缘检测模型RED。该模型在RCF模型的基础上,去除主干网络中部分下采样,并在主干网络中引入CA注意力模块和混合扩张卷积;在深监督模块改变压缩层的数量和参数,采用平滑压缩的方式进行特征降维;在最后的融合模块,采用跨层交叉融合的方式来融合高低层间的信息。改进后的模型在扩充后的BSDS500数据集上进行了训练和测试,通过在BSDS500基准上进行评估得到数据集最优尺度(ODS)和单图最优尺度(OIS),分别为0.809和0.832。实验结果表明,该模型提取的边缘轮廓更加清晰细致,提取到的边缘信息也更加全面丰富。

Edge detection is a fundamental method in image processing and computer vision.Aiming to address the issues of roughness and blurriness in edges generated by deep learning-based edge detection technology,a refined edge detection(RED)model based on richer convolutional features(RCF)for edge detection was proposed.In this model,RCF was used as the baseline network.Some downsampling operations in the backbone network were removed,and the coordinate attention(CA)module and hybrid dilated convolution were added to the backbone network.The number and parameters of the compression layers were changed in the deep supervision module,and smooth compression for reducing feature dimensionality was adopted.In the final fusion module,a cross-layer cross-fusion module was used to fuse the information from high and low layers.The RED model was trained and tested on the extended BSDS500 dataset.The optimal dataset scale(ODS)and the optimal image scale(OIS)of the dataset were 0.809 and 0.832,respectively,as evaluated on the BSDS500 benchmark.The experimental results showed that RED model extracted clearer and more detailed edge contours,and the extracted edge information was more comprehensive and abundant.