人工智能识别与人工识别红外相机动物影像准确率分析:以上海大金山岛猕猴监测为例  

Analysis of the Accuracy of Artificial Intelligence Recognition and Artificial Recognition of Camera Traps Images:An Example of Macaca mulatta Monitoring on Dajinshan Island,Shanghai.

作　　者：李必成[1] 张晨曦 季钰翔 孙锬锋[3] 丁屹旻 张伟 谢汉宾 王军馥[1] 张云飞[1] 李雪梅 王小明[1] 杨刚[1,2] LI Bi-cheng;ZHANG Chen-xi;JI Yu-xiang;SUN Tan-feng;DING Yi-min;ZHANG Wei;XIE Han-bin;WANG Jun-fu;ZHANG Yun-fei;LI Xue-mei;WANG Xiao-ming;YANG Gang(Shanghai Science and Technology Museum/Laboratory of Ecological Security and Biodiversity Conservation of Yangtze River Delta Urban Agglomeration,Shanghai Science and Technology Museum,Shanghai 200127,China;College of Marine Science,Shanghai Ocean University,Shanghai 201306,China;School of Cyber Science and Engineering,Shanghai Jiaotong University,Shanghai 200240,China)

机构地区：[1]上海科技馆/上海科技馆长三角城市群生态安全与生物多样性保护实验室,上海200127 [2]上海海洋大学海洋学院,上海201306 [3]上海交通大学网络空间安全学院,上海200240

出　　处：《生态与农村环境学报》2023年第7期918-923,共6页Journal of Ecology and Rural Environment

基　　金：上海市自然科学基金(20ZR1437100);国家自然科学基金(31601872);上海科技馆长三角城市群生态安全与生物多样性保护实验室项目。

摘　　要：红外野生动物相机影像人工智能识别已成为生态学与人工智能交叉学科研究的热点之一。为探究人工智能识别红外相机动物影像的准确率及其影响因子,比较人工智能识别与人工识别的差异。以上海大金山岛猕猴(Macaca mulatta)监测为例,应用YOLO v3模型进行训练与测试,探讨YOLO v3模型识别大量红外相机图像的可行性。同时,对比人工智能图像识别与人工识别的准确率与识别效率,找出特定样本容量条件下识别方式的最优解。对11106张照片的识别结果表明,人工智能识别总准确率为69.0%,均值为68.2%;人工识别总准确率为99.0%,均值为99.1%。人工识别准确率显著高于人工智能识别准确率(t=-9.256,df=22,P<0.01)。简单生境背景的人工智能识别准确率显著高于复杂生境背景(Z=-2.270,P=0.023)。简单生境背景的人工识别准确率与复杂生境背景无显著差异(Z=-0.406,P=0.685)。因此,人工智能识别更适用于生境及背景简单的红外影像,但需谨慎用于识别复杂生境的背景。同时,人工智能识别可用于对大量照片的初筛。人工识别可用于识别复杂生境背景的照片和对人工智能初筛后照片的复核。对于万张级的样本量,人工智能并未显示出明显的时间优势,人工识别反而具有准确率优势。随着各类训练数据集的不断建立与开放应用,对于大型脊椎动物,特别是一些公众熟知的明星物种的人工智能识别可能会率先代替人工识别。Artificial intelligence(AI)recognition of camera traps images has become one of the hot spots in the interdisciplinary research of ecology and AI.In order to explore the accuracy and influencing factors of artificial intelligence recognition of infrared camera animal images,the differences between artificial intelligence recognition and artificial recognition were compared.Taking the monitoring of macaques(Macaca mulatta)on Dajinshan Island in Shanghai as an example,the TOLO v3 model was applied for training and testing,and the feasibility of the TOLO v3 model to recognize a large number of infrared camera images was discussed.Meanwhile,the accuracy and recognition efficiency of AI image recognition and artificial recognition are compared to find out the optimal solution of recognition method under specific sample capacity.The recognition results of 11106 photos show that the total recognition accuracy of artificial intelligence is 69.0%,and the average is 68.2%.The total accuracy of artificial recognition was 99.0%,and the average was 99.1%.The accuracy of artificial recognition was significantly higher than that of artificial intelligence(t=-9.256,df=22,P<0.01).The recognition accuracy of simple habitat background was significantly higher than that of complex habitat background(Z=-2.270,P=0.023).For artificial recognition,there was no significant difference in accuracy between simple habitat background and complex habitat background(Z=-0.406,P=0.685).AI recognition can be used for infrared images with a single habitat and background,but should be cautiously used for background recognition of complex habitats.In addition,AI recognition can be used for initial screening of large numbers of photographs.Artificial recognition can be used to identify photos with complex habitat backgrounds and to review photos after initial screening by artificial intelligence.For the sample size of ten thousand photos,artificial intelligence does not show obvious time advantage,but artificial recognition has an advantage of accurac

关 键 词：人工智能 人工识别 准确率 猕猴 大金山岛 

分 类 号：X835[环境科学与工程—环境工程] Q958.1[生物学—动物学]