基于G-RepVGG和鱼类运动行为的水质监测方法

水质恶化会直接造成水产养殖产量下降,严重时会导致水产动物大量死亡,给养殖企业造成严重经济损失。因此对水产养殖中水质参数进行实时监测具有重要意义。本文以斑石鲷为研究对象,提出了一种基于鱼类行为的水质监测方法。该方法通过摄像机拍摄到的图像数据就可以非侵入地完成水质参数的实时监测,避免了安装复杂设备、对鱼类行为进行量化等繁琐过程。为了增加推理速度和降低模型参数量,通过将RepVGG block与GhostNet相结合构建了G-RepVGG模型,使该模型更适用于移动设备的部署。提出了计算量较少、推理速度快、更适合水质快速监测的Cheap Ghost操作和计算量大、精确率高、更适合水质的精确监测Expensive Ghoost操作。由于多分支网络适合进行训练但是在推理速度上低于单分支网络,因此通过模型重参数化首先将卷积层以及批归一化（Batch normalization, BN）层合并,随后再将3路卷积合并为1路,大大降低模型参数量、提高了模型推理速度,使模型更加适用于移动设备的推理。结果表明：使用Cheap Ghost操作的G-RepVGG在测试集中准确率达到96.21%,图像处理速度达到442.27f/s,使用Expensive Ghost操作的G-RepVGG模型在测试集中准确率达到97.63%,图像处理速度达到349.42f/s,从而在保证较高精度的前提下依旧具有较高的推理速度,在多个数据集中测试具有较好的鲁棒性。

Water Quality Monitoring Based on Fish Movement Behavior and G-RepVGG

Waterin aquaculture is a necessary place for aquatic animals to survive and live. The deterioration of water quality will directly lead to the decline of aquaculture production, and in severe cases, it will lead to the death of a large number of aquatic organisms and cause serious economic losses to aquaculture enterprises.Therefore, the real-time monitoring of water quality parameters in aquaculture is of great significance.A method for water quality monitoring based on fish behavior was proposed with Oplegnathus punctatus as research object.The method can non-invasively complete the real-time monitoring of water quality parameters through the image data captured by the camera, avoiding the tedious installation of complex equipment and the quantification of fish behavior.To increase the inference speed and reduce the amount of model parameters, this method combined RepVGG block with GhostNet.Aiming at the problems of rapid water quality monitoring and accurate water quality monitoring, the Cheap Ghost operation and the Expensive Ghost operation were proposed.Finally, the three branches were merged through model reparameterization, which greatly reduced the amount of model parameters and improved the model inference speed.The results showed that the G-RepVGG operated by Cheap Ghost achieved an accuracy of 96.21% in the test set and can infer 442.27 images per second. The G-RepVGG model operated with Expensive Ghost achieved 97.63% accuracy in the test set and can infer 349.42 images per second. Therefore, it still had a high inference speed under the premise of ensuring high accuracy, and had better robustness in testing in multiple data sets. The research result can quickly and accurately monitor water quality, detect water quality deterioration in time, and reduce losses caused by water quality deterioration, providing ideas and methods for water quality monitoring.