Research On Circle Detection Method Based On Convolution Neural Network

Vision inspection is a kind of inspection technology based on machine vision,which has the characteristics of non-contact,non-damage,fast and convenient,so it is widely used.Circle detection is a basic operation in vision detection,which plays an important role in industrial detection,assistant driving,biometrics and other fields.Circle detection is to detect the circular object in the image and estimate the center and radius of the circular object in the image.The existing circle detection algorithms mainly rely on the edge map obtained by the edge detector for calculation.This kind of edge map not only contains a large number of invalid edges,but also mixes the effective arc edges,which is not conducive to multi circle detection.Inspired by the success of convolution neural network in other fields,this paper uses convolution neural network technology to design and implement a circle detection method with high detection accuracy and strong robustness.A strategy of "coarse detection before fine detection" is adopted,which is divided into three modules: coarse detection module,fine detection module and circle parameter estimation module.The coarse detection module uses the target detection technology to divide the multi circle detection task into several single circle detection tasks.The fine detection module uses semantic segmentation technology to segment the edge of a specific circle(excluding the edge of background and texture).The circle parameter estimation module calculates the circle parameters(center and radius)by using the edge map.For the research of coarse detection module and fine detection module,this paper collects coin data set to train and test the two modules.In order to enhance the robustness of the proposed method,the coin dataset contains three detection difficulties: occlusion interference,texture interference and shadow interference.For the coarse detection module,this paper analyzes the principle of three target detection algorithms and compares them with experiments,and takes SSD algorithm with better performance as the coarse detection module.For the fine detection module,this paper improves the stacked hourglass network as the fine detection module,and compares the network performance before and after the improvement through experiments.At the same time,four groups of parameters about the stacked hourglass network are set,and the best one is selected through experiments.In order to verify the performance of the circle detection method,after obtaining the prediction results of the network,a circle parameter estimation module is designed according to the geometric characteristics of the circle.Finally,the proposed method and three comparison algorithms are tested on the coin test set,and the accuracy,recall and error of the experimental results are counted.The experimental results show that the proposed method has good accuracy and robustness in the test set.