基于运动特征提取和2D卷积的鱼类摄食行为识别研究

为了促进渔业装备智能化，近年来基于视频流的鱼类摄食行为识别研究受到了广泛关注。针对基于视频流的传统识别方法模型过于复杂，难以在边缘计算设备部署的问题，提出了一种轻量级的2D卷积运动特征提取网络Motion-EfficientNetV2，该网络以视频流为输入，能够有效识别鱼类摄食行为。提出的模型以EfficientNetV2为主干网络，基于TEA和ECANet构建了运动特征提取模块Motion，并将该模块嵌入到EfficientNetV2的每个Fused-MBConv模块中，使改进后的EfficientNetV2具有运动特征提取能力。同时使用ECANet对EfficientNetV2网络中的MBConv进行改进，增强其通道特征提取能力。在此基础上利用空洞卷积扩大感受野，提高大范围特征提取能力。试验结果表明，Motion-EfficientNetV2的参数量和浮点运算量分别为9.3×106和1.31×1010，优于EfficientNetV2。在TSN-ResNet50、TSN-EfficientNetV2、C3D以及R3D模型上进行对比试验，本文模型在降低参数量和浮点运算量的同时，使识别准确率提高到93.97%。该研究对于渔业装备智能化升级和科学养殖具有推动作用。

Recognition of Feeding Behavior of Fish Based on Motion Feature Extraction and 2D Convolution

In order to promote the intelligence of fishery equipment, video streaming-based fish feeding behaviour recognition has received extensive attention in recent years. The model of traditional recognition methods based on video streaming is too complex to be realized on edge computing devices. To address this problem, a lightweight 2D-convolutional motion feature extraction network, Motion-EfficientNetV2, was proposed which can effectively recognize fish feeding behaviour by using video streams as input. The proposed model used EfficientNetV2 as the backbone network, constructed the motion feature extraction module Motion based on TEA and ECANet, and embeded the Motion module into each Fused-MBConv module of EfficientNetV2, in order to give EfficientNetV2 the ability to extract motion features. The MBConv in the EfficientNetV2 network was also improved by using ECANet to enhance its channel feature extraction capability. Null convolution was used in Motion-EfficientNetV2 to expand the receptive field and improve the wide-range feature extraction capability. The experimental results showed that after introducing the designed Motion module and a series of improvements, the number of parameters and FLOPs of Motion-EfficientNetV2 was 9×106 and 1.31×1010, respectively, which were reduced compared with EfficientNetV2. Comparison experiments using the same dataset in the algorithmic models of TSN-ResNet50, TSN-EfficientNetV2, C3D, and R3D, respectively, showed that the present algorithm achieved an accuracy of 93.97% while the number of parameters and FLOPs were lower than the rest of the models. Therefore, the model proposed can effectively identify fish feeding behavior and guide aquaculturists to develop fish feeding strategies.