Study On Edge Detection With Feature Re-extraction Deep Convolutional Neural Network

Edge detection can extract the contours of objects and visually prominent edges in natural images,hence it is widely used in image segmentation,object detection,target recognition,3D reconstruction,medical images and other fields.In recent years,with the successful application of deep learning in the image processing,it solves the problem that traditional edge detection algorithms cannot effectively extract high-level information in images in a large part,which leads to low robustness of detection results.Nevertheless,the edge detection algorithm based on deep learning still has the disadvantage that the features extracted from the backbone neural network cannot be fully utilized effectively.At the same time,for the imbalance of training samples,loss function cannot effectively guide the training process of the model.In this paper,the edge detection algorithm based on deep convolution neural network is studied experimentally.In order to effectively utilize the features extracted by backbone neural network for edge detection,an edge detection algorithm based on deep convolution neural network feature re-extraction is proposed.The algorithm mainly consists of three modules: backbone network(VGG-16),feature re-extraction and feature fusion.In the feature extraction module,residual learning mechanism is introduced to eliminate potential gradient vanishing/exploding,which enables the module to more stably map the features from different levels of backbone network to edge pixel space.Finally,the edge image is obtained by feature fusion.The co-distributed generalization ability of the model is validated on BSDS 500 dataset,and obtain the ODS F-score of 0.807.The cross-distribution generalization ability of the model is validated on NYUD V2 dataset,and the ODS F-score is 0.701.In order to further verify the effectiveness of the proposed feature re-extractor,when the backbone neural network is frozen,the model can achieve ODS F-score of 0.792 after only 5.4k iterations on BSDS 500 dataset.To solve the disadvantageous effect of the imbalance of positive and negative samples during training process,the original loss function based on binary classification is optimized.By multiplying the positive and negative sample loss terms of the binary loss function by different weight coefficients,the network can focus more on training hard examples.In order to verify the effectiveness of the optimization scheme,the scheme proposed is compared with the loss function in HED,RDS and RCF.The experimental results show that the optimization scheme can effectively improve the detection accuracy of the model,and the ODS performance index are improved by 0.006,0.005 and 0.005 respectively.