摘要
养殖环境中饲料投放、水流变化等刺激源导致鱼类声音分辨难,使行为识别准确率不高,为解决上述问题,提出基于Mel声谱图(Mel spectrogram)与改进SEResNet的鱼类行为识别模型TAP-SEResNet。首先针对鱼类行为声音频率波动大、特征差异小,造成特征提取难的问题,采用高分辨率、特征表示较好的Mel声谱图以捕捉鱼类声音的频谱特征。其次针对鱼类声音特征关键信息易丢失的难题,提出在SEResNet模型中融合时序聚合池化层(Temporal Aggregated Pooling,TAP),提取池化区域的最大值和平均值,保留鱼类行为更多细粒度声音特征,提高识别准确率。为验证所提模型的有效性,分别设计了消融试验和模型性能对比试验,试验结果显示:TAP-SEResNet相比SEResNet在不降低检测速度的条件下准确率提升了3.23%;相比PANNS-CNN14、ECAPA-TDNN及MFCC+ResNet等先进声音识别模型,TAP-SEResNet在准确率上分别提升了5.32%、2.80%和1.64%。所提模型有助于养殖过程中对鱼类行为实现精准监测,对精准养殖具有重要的推动作用。
In order to solve the problem that the sound discrimination of fish is difficult and the behavior recognition accuracy is not high due to the stimulus sources such as feed release and water flow change in the breeding environment,a fish behavior recognition model based on Mel spectrogram and improved SEResNet was proposed.Firstly,in view of the difficulty of feature extraction due to the large frequency fluctuation and small feature difference of fish behavior sounds,a high-resolution Mel spectrogram with good feature representation is adopted to capture the spectral features of fish sounds and enhance the recognition ability of fine-grained sound information of fish.Secondly,to solve the problem that key information of fish sound features is easy to be lost,it is proposed to integrate the Temporal Aggregated Pooling layer in the SEResNet model,extract the maximum value and average value of the pooled region,and retain more fine-grained sound features of fish behaviors to improve the recognition accuracy.To verify the effectiveness of the proposed model,the ablation experiment and the model performance comparison experiment were designed respectively.The test results showed that the accuracy of TAP-SEResNet was improved by 3.23%compared with SEResNet without reducing the detection speed.Compared with advanced voice recognition models such as PANNS-CNN14,ECAPA-TDNN and MFCC+ResNet,TAP-SEResNet has improved its accuracy by 5.32%,2.80%and 1.64%,respectively.The results show that the proposed model can effectively solve the problem of low accuracy of fish behavior recognition in aquaculture environment,help to realize accurate monitoring of fish behavior in aquaculture process,and play an important role in promoting precision aquaculture.
作者
杨雨欣
于红
杨宗轶
涂万
张鑫
林远山
YANG Yuxin;YU Hong;YANG Zongyi;TU Wan;ZHANG Xin;LIN Yuanshan(College of Information Engineering,Dalian Ocean University,Dalian 116023,Liaoning,China;Dalian Key Laboratory of Smart Fisheries,Dalian 116023,Liaoning,China;Key Laboratory of Environment Controlled Aquaculture(Dalian Ocean University),Dalian 116023,Liaoning,China;Liaoning Provincial Key of Marine Information Technology,Dalian 116023,Liaoning,China)
出处
《渔业现代化》
CSCD
北大核心
2024年第1期56-63,共8页
Fishery Modernization
基金
辽宁省教育厅重点科研项目“面向鱼类行为分析的声音与视觉特征融合关键技术研究(LJKZ0729)”
国家自然科学基金项目“水下实时背景下鱼类精准识别新方法研究:融合VSM和DELM(31972846)”
设施渔业教育部重点实验室开放课题“基于鱼类骨架和轨迹特征的异常行为识别方法研究(202313)”。
