融合图像处理技术的红树林鸟类鸣声识别算法  

A mangrove birdsong recognition algorithm integrating image processing technology

作　　者：陈炀 周雁 王庆娟 张馨元 谌业恒 Chen Yang;Zhou Yan;Wang Qingjuan;Zhang Xinyuan;Chen Yeheng(Beijing Institute of Technology(Zhuhai),Zhuhai 519088,China;Xuteli School,Beijing Institute of Technology,Beijing 102401,China)

机构地区：[1]北京理工大学(珠海),珠海519088 [2]北京理工大学徐特立学院,北京102401

出　　处：《现代计算机》2024年第21期38-42,48,共6页Modern Computer

基　　金：广东省普通高校特色创新项目(2022KTSCX202);广东省科技创新专项资金(“攀登计划”专项资金,pdjh2022b0711);大学生创新创业训练计划项目(BIT2023XX568)。

摘　　要：鸟类的物种监测一直是生物多样性研究的重要内容。然而,目前的鸟类鸣声识别算法大多未考虑环境干扰和算法速度。提出一种基于深度学习、融合图像处理技术对鸟类鸣声音频进行识别的算法。算法对鸟鸣声进行降噪后采用音频与图像处理技术相结合的方法提取鸟类鸣声特征;利用卷积神经网络ResNet进行建模,并将最优模型转换为TensorRT模型,提升推理速度。经实验测试,该算法对鸟类鸣声的分类识别具有良好效果,识别速度也有明显的提高。Birds species monitoring is an important part of biodiversity research.However,most of the current birdsong recog-nition algorithms do not consider environmental disturbances and algorithm speed.The paper proposes an algorithm for audio recog-nition of birdsong based on deep learning and fused image processing technology.After denoising the birdsong,the algorithm com-bines audio and image processing technology to extract birdsong features;and with the convolutional neural network ResNet,it mod-els and selects the optimal model,and then converts it into TensorRT model to improve the inference speed.In the experimental tests,the algorithm has shown good performance in classifying and recognizing birdsongs.The optimized model also exhibits a no-ticeable improvement in recognition speed.

关 键 词：鸟类鸣声识别 Mel谱图 卷积神经网络 TensorRT 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]