Image Edge Detection Algorithm Based On Deep Learning

Image edge detection is a basic problem in the field of computer vision and image processing.Many deeper calculations and operations are based on the results of edge detection in image processing.Therefore,how to perceive and extract image edges accurately and clearly has become the key work of image processing.Because of the rapid developments of computer vision led by deep learning.But edge detection algorithm based on convolution neural network has become a research hotspot in image processing.It has achieved good results Edge detection in clear scenes.However,inaccurate edge extraction will occur when processing blurred images.The problem of incomplete edge extraction will occur in the processing of weak edge images.Based on the improved HED network,this thesis makes a new research on image edge detection algorithm.The main work of this thesis is as follows:(1)An edge detection algorithm based on HED network is proposed.By analyzing the detection results of HED network in different scenes,an improved deep learning network algorithm is proposed to solve the problem of incomplete edge of traditional HED network in weak image edge.Modify the HED network structure.Change the two pooled layers into convolution layers with the same number of convolution cores and the same length.Reduce the number of pooling layers.Improved resolution of output image of each convolution layer.Modified the deconvolution layer in the side output layer.The network model is optimized.Avoid secondary classification and other problems.Adjust the activation function of the network.The nonlinear function Re LU is introduced as the activation function of the deep learning network.Design experiments to verify the improved network model and analyze the experimental results.The experimental results verify that the performance of the modified network structure is improved,and the performance of the modified network structure is improved.(2)A deep learning network model integrating soft attention mechanism is proposed.Aiming at the problems of large amount of data computation and limited computing resources in the improved HED network.Weight distribution of output data through soft attention mechanism.Adjust the calculation order within the network.Prioritize the processing of computing resources on the external edge.Improve network model efficiency.Experiment on the above network model.Validate on BSDS500 and NYUDv2 datasets.Set up evaluation indicators to analyze the experimental results from three perspectives: network response rate,detection accuracy and visual image analysis.After experimental verification,the improved deep learning network model has improved in relevant indicators compared with the original network model.