融合注意力机制的淡水鱼类识别方法

Freshwater fish identification method incorporating attention mechanism

针对多数淡水鱼类识别方法特征的提取进程复杂,在自然外部场景下很难进行高效识别问题,构造了残差模型及注意力机制相融合的ResNet50-SE-Fish网络对不同发育程度的8种淡水鱼类进行识别,并测试构造的网络识别性能.研究在原残差块内添加SE模块,增加所关注特征数据权重,减小外部场景噪声以及背景无关特征数据权重,并使用迁移学习来对不同发育程度幼鱼、成鱼特征数据进行识别.通过Grad-CAM方法对ResNet50-SE-Fish网络每个残差块提到的淡水鱼类特征数据可视化以解释注意力表示作用,并与7种常用网络所提到的热力图比较,以评估网络对淡水鱼类特征数据提取能力.结果表明,ResNet50-SE-Fish网络对不同发育程度淡水鱼类有很高的识别精度,验证时准确率高达95.53%,测试时准确率达90.16%,相较于AlexNet、VGG16、ResNet18、GoogleNet、VGG19、ResNet34、ResNet50,测试时准确率依次增大14.93%、6.32%、2.51%、3.55%、3.69%、2.14%、1.73%,注意力机制利用调节通道关注程度能够提高模型淡水鱼类特征的提取效果,降低数据背景干扰,为淡水鱼类识别提供参考.

Most freshwater fish species identification methods are inefficient due to their complex feature extraction process, and are difficult to accurately identify the varieties under natural background.Thus, a residual network ResNet50-SE-Fish based on attention mechanism was proposed to recognize 8 freshwater fish species at different growth stages.The species recognition effect of the improved network was analyzed and verified.In the study, the attention module SE was added into the ResNet50 network to increase the weight value of the effective feature and reduce the weight value of the invalid feature of background and noise, and the feature recognition of juvenile and adult freshwater fish at different growth stages was realized by transfer learning.Besides, in order to explain the function of attention module, the Grad-CAM method was applied to visualize the characteristics of freshwater fish at different growth stages extracted from each layer of ResNet50-SE-fish network with thermodynamic diagrams.The comparisons with thermodynamic diagrams extracted from the last convolutional layers of other 7 traditional networks were made, and then the performance of the proposed network model could be evaluated.Experimental test results showed that the network model proposed had a high identification accuracy for freshwater fish species at different growth stages, the verification accuracy was up to 95.53%,and the test accuracy was 90.16%.Compared with AlexNet, VGG16,ResNet18,GoogleNet, VGG19,ResNet34,and ResNet50,the test accuracy increased respectively by 14.93%,6.32%,2.51%,3.55%,3.69%,2.14% and 1.73%.Therefore, the attention module can effectively improve the extraction effect of freshwater fish features and reduce image background interference, which may provide reference for freshwater fish species recognition.