Research On Ear Detection Methods Based On Convolutional Neural Network

The past few decades have witnessed the booming of biometrics.Biometrics generally include measurable,unique,and permanent biological characteristics such as iris,fingerprint,face,and DNA,and biometrics technologies generally refer to that using biometrics for identification.The human ear is a relatively new feature in the field of biometrics.The structure of the human ear is not only unique,but also permanent,and the appearance of the human ear does not change easily in human life.If we want to obtain a certain reliable biological characteristic through the image and carry on the identification of the identity,the object detection is an essential means.Since human ear has these very good characteristics above,automatic human ear detection has recently attracted great attention in the biometric community.Because the human ear is a very effective feature for identification,when the face of the person in the surveillance video is obstructed,the human ear can be used as a supplementary feature to assist recognition.Since the convolutional neural network model win the championship in the Image Net 2012 competition,convolutional neural network in recent years has achieved very good results in object detection.This paper attempts use convolutional neural network for human ear detection,and propose a method of ear detection using the combination of convolutional neural network and direct regression.This article mainly carries out the following three aspects of research:(1)In terms of human ear classifiers,this paper utilize the idea of migration learning to fine-tune our designed classification network using existing ear images,and then obtain human ear classifiers.The human ear classifier is mainly used to extract human ear features for subsequent human ear detectors.At the same time,this paper compare the accuracy of the above network models with different depths on the classification test set.Experiments show that the increase in depth can bring about an increase in accuracy.At the same time,the human ear classifier designed in this paper is compared with the currently popular convolutional neural network and traditional methods.In the human ear classification task,the network model of this article is superior to the other method.(2)In terms of human ear detector,this paper proposes the Conv Tran layer for human ear detection.The last Softmax layer of the trained human ear classifier network is replaced with a Conv Tran layer to achieve the task of ear detection.The human ear detector and the human ear classifier use the same backbone network and share network weights,this paper uses this method to accelerate network training.Conv Tran uses the idea of bounding box regression and the anchor generated by k-means to implement the human ear detection task.At the same time,in order to improve the accuracy of human ear detection,the non-maximum suppression(NMS)method is also used in this paper.The proposed method for human ear detection is compared with other traditional feature-based human ear detection methods and convolutional neural network based human ear detection methods.The experimental results show that the method of this paper has good performance in human ear detection,and Average Precision(AP)is higher than other methods.(4)In terms of human ear position alignment,when the human ear detector is used on the image sequence,the jitter of the box is found to be severe.For this reason,this paper propose a location-alignment model for further refinement of the box position.The method firstly extends the detected ear box on the original map by 25%,then the extended image is fed to the convolutional neural network to obtain a more accurate human ear box.In the prediction,the model also uses the original image information to further improve the accuracy of the prediction results.This paper compares the original detected box with the box that has been further refined by location-alignment.The average IOU of the box obtained by the latter is higher than the former,making the detected box more accurate and more stable in the image sequence.