Research On Multi-scale Edge Detection Based On Deep Learning

Edge detection aims to extract perceptually salient edges of natural images,which is the basis of high-level computer vision tasks,such as image segmentation,object detection/recognition,image sketch,and even directly determines the upper limit of the task accuracy.So,it is very meaningful to study edge detection.Traditional edge detection methods distinguish edge pixels according to manual features such as brightness,color,gradient and texture.However,it is difficult to express high-level semantic meaning with low-level visual cues.In recent years,with the advent of artificial intelligence,deep learning has become a mainstream method in the field of image processing.But up to now,the application of deep learning in image edge detection has just begun,and the research results are limited.At the same time,the training of deep learning algorithm needs a lot of data operations,and the multiscale change of the detected object brings new challenges to edge detection.To solve the above problems,this paper does research on the multi-scale edge detection algorithm based on deep learning.In edge detection based on deep learning,similar low-level features are extracted multiple times at multiple scales,which leads to redundant use of information.At the same time,the ineffective model of global long-range dependency leads to non-optimal discriminative feature representations.To solve the above problems,this paper proposes a multi-scale global channel network.This method uses the rich convolution layers of VGG16 to capture multi-scale features.The global module models the long-range context dependencies,and combines the local features with their corresponding global dependencies.The channel module adaptively recalibrates the channel response with less parameters,thus guiding the network to ignore irrelevant information and emphasize the correlation between related features.Through the ablation experiments on BSDS500 dataset and NYUD dataset,this method achieves ODS Fscore of 0.815 and 0.741 respectively,which are 0.9% and 1.2% higher than other existing algorithms.It is proved that this method can obtain clearer edges than other algorithms under the balance of parameters scale and precision,and is obviously better than other comparison algorithms.Aiming at the edge detection algorithm based on deep convolution network adopts the same single depth supervision for multi-scale middle layer output,ignoring the difference of middle layer feature scale,and the gradient vanishing/exploding and network degradation problems in deep convolution network,this paper introduces residual learning,constructs the depth supervision network based on residual network to improve the convergence effect of the network,and improves the training efficiency Stability.In this paper,the residual network is divided into different stages.The feature information of different scales in each stage is used to enrich multi-scale features.The intermediate output of each stage is supervised by relaxed label,and the output of each stage is sampled by sub-pixel convolution.The obtained high-resolution feature map is used for edge detection.Compared with the classical Canny operator and the current mainstream edge detection methods based on deep learning,such as HED,RCF,etc.,the edge obtained by this method has better continuity and accuracy,and can detect more fine edges.The ODS F-score on BSDS500 dataset is improved by 20.3%,2.6% and 0.8%respectively,and the training time is shorter.Using multi-scale features is of great significance to improve the edge detection of targets at different scales.In order to extract edge features at different scales,this paper proposes to use edge tags at specific scales to monitor the output of each convolution layer.In addition,a multi-scale enhancement module is introduced to enrich the multi-scale representation of shallow network.The module generates multi-scale features by expansion convolution to obtain a compact network with fewer parameters.In this paper,the proposed method is evaluated on BSDS500 dataset,and achieves ODS F-score of 0.818,which is 1.2% higher than the current algorithms.