基于MFCC和ResNet的鱼类行为识别

Fish Behavior Recognition Based on MFCC and ResNet

针对在极暗或无光条件下,采用计算机视觉手段进行鱼类行为识别效果不好的问题,本文提出了利用声音信号识别鱼类行为的方法;通过观察和试验发现鱼类的摄食、游泳等行为具有声音差异小、特征学习难等特点,基于上述发现,提出采用具有较强特征表达能力、能区别细微特征的MFCC(Mel-frequency cepstral coefficient,MFCC)特征系数表示鱼类行为声音信号特征&为有效学习不同鱼类行为的细粒度声音特征,采用残差网络(Residual Neural Network,ResNet)进行低维细节特征与高维语义特征融合,以便更好地保证特征完整性、提高识别效果。为验证所提出方法的有效性,设计了3组对比试验,用大连海洋大学鱼类行为学实验室采集的数据验证了算法的有效性,试验结果表明,鱼类行为识别的正确率、召回率和F1值均达到99%。研究表明,基于MFCC和ResNet的鱼类行为识别方法可以有效识别鱼类的游泳、摄貪等行为,为鱼类行为识别研究提供了新思路和新方法。

Aiming at the recognition deficiency of fish behavior by computer vision under extremely dark or no light conditions,a method of using sound signals to recognize fish behavior is proposed.Equal behavior has the characteristics of small sound difference and difficult feature learning.Based on the above findings,it is proposed to use the MFCC(Mel-frequency cepstral coefficient,MFCC)feature coefficient which has strong feature expression ability and can distinguish subtle features to represent the sound signal characteristics of fish behavior.In order to effectively learn the fine-grained sound features of different fish behaviors,ResNet(Residual Neural Network,Residual Network)is used to fuse low-dimensional detail features and high-dimensional semantic features to better ensure feature integrity and improve recognition effects.To verify the effectiveness of the proposed method,three groups of comparative experiments were designed,and the effectiveness of the algorithm was verified with the data collected by the fish behavior laboratory of Dalian Ocean University.The test results show that the accuracy and recall rate of fish behavior recognition and F1 values all reached 99%.This study showed that the fish behavior recognition method based on MFCC and ResNet can effectively identify the swimming,feeding and other behaviors of fish,providing new ideas and methods for fish behavior recognition research.