基于改进VGG16的射流式鱼泵内鱼类损伤的图像识别与分类

Image Recognition and Classification of Fish Damage in Jet Fish Pumps Based on Improved VGG16

鱼类损伤分类研究对捕捞作业过程中的鱼类健康状况监测具有重要意义。针对现有鱼类损伤研究中存在的未准确分类损伤类型的问题,在射流式鱼泵捕捞作业时鱼逐条通过泵喉管的特殊场景下,本文使用高速摄像机拍摄鱼通过泵的过程,建立了鱼类损伤数据集,提出了一种基于VGG16改进的S-VGG卷积神经网络分类模型。损伤分类实验表明,S-VGG模型的损伤分类准确率为96.52%,比ResNet16和GoogLeNet分别提高了0.96%和8.2%。与VGG16模型相比,本文所提出的S-VGG模型仅有9层,整体参数减少了93.75%,有效地降低了模型训练所需的计算成本。本研究采用迁移学习方法进一步优化了S-VGG模型初始权重。结果表明,经过迁移学习的S-VGG*模型准确率达到了99.70%,比未进行迁移学习的S-VGG模型提高了3.18%。本研究建立的S-VGG*模型取得了良好的鱼类损伤分类效果。

Fish damage classification researches are of a great significance for monitoring the health status of fish during fishing operations.To address the issue of inaccurately classification of damage types in existing studies,in this study,we focused on the unique scenario of fish passing through the pump throat individually during jet fish pump operations.Using a high-speed video camera to record the process,we constructed a self-built fish damage dataset and proposed an improved S-VGG convolutional neural network classification model based on VGG16.The experimental results for damage classification showed that the S-VGG model achieves an accuracy of 96.52%,representing an improvement of 0.96%and 8.2%over ResNet16 and GoogLeNet,respectively.Notably,compared to the VGG16 model,the S-VGG model proposed in this study was only 9 layers,which reduced the overall parameters by 93.75%,and effectively minimized the computational cost of model training.Furthermore,we optimized the initial weight of the S-VGG model using the transfer learning method.The results indicated that the accuracy of the S-VGG*model after transfer learning reached 99.70%,showed an improvement of 3.18%and demonstrated an excellent performance in fish damage classification.