基于深度卷积神经网络的蝇类面部识别  

Fly facial recognition based on deep convolutional neural network

作　　者：陈彦彤 陈伟楠 张献中 李雨阳 王俊生[1] CHEN Yan-tong;CHEN Wei-nan;ZHANG Xian-zhong;LI Yu-yang;WANG Jun-sheng(College of information science and technology, Dalian Maritime University, Dalian 116026, China)

机构地区：[1]大连海事大学信息科学技术学院,辽宁大连116026

出　　处：《光学精密工程》2020年第7期1558-1567,共10页Optics and Precision Engineering

基　　金：国家自然基金资助项目(No.61901081);中央高校基本科研业务费专项资助资助(No.3132020199)。

摘　　要：针对蝇类昆虫物种繁多、特征复杂等因素,导致蝇类识别准确率低、耗时较长等问题。本文借鉴深度学习方法中的人脸识别算法,提出一种基于深度卷积神经网络的蝇类面部识别方法。首先,在图像对齐过程中,使用多任务卷积神经网络并进行优化即应用深度可分离卷积减少计算参数,缩短图像预处理时间。其次,应用轮廓特征粗提取和具体部位特征细提取相结合的方式提取更加丰富的特征信息即使用卷积池化粗提取出图像的轮廓特征值;同时,使用Inception-ResNet网络、Reduction网络细提取出具体部位特征值。最终在网络训练时,结合上述方法使得提取到的特征信息更加精确全面。实验表明,所提方法的准确率达到94.03%,相较于其他网络训练方法,该方法在保证较高准确率的情况下提升计算效率。Given the large number of species of flies and their individual complex characteristics,recognizing a particular type of fly has proved to be time consuming and,for the most part,inaccurate.In this paper,a method for the facial recognition of a fly using deep learning technologies was proposed,specifically a Convolutional Neural Network(CNN),and its related face recognition algorithms.Initially,a multi-task convolutional neural network was utilized and optimized for the image alignment process.Thus,depth-wise separable convolutions were applied to reduce the number of calculation parameters as well as the image preprocessing time.Next,we combined the rough extraction of contour features and fine extraction of specific parts to derive more abundant feature information.The convolution and pooling layers were harnessed to elicit contour eigenvalues of the image,while Inception-ResNet and Reduction networks were administered simultaneously to obtain eigenvalues of specific parts.Finally,the above methods were coalesced to enhance the accuracy and comprehensibility of the resultant feature information for network training.Experimental results show that the mean average precision of the proposed method is 94.03%.When compared with other network training methods,this method not only improves the computational efficiency but also ensures high accuracy.

关 键 词：蝇类面部识别 深度卷积神经网络 多任务卷积神经网络 Inception-ResNet网络 Reduction网络 

分 类 号：TP306[自动化与计算机技术—计算机系统结构]