增强细节信息特征提取的鱼类个体识别算法

Fish individual recognition algorithm of feature extraction of enhanced detailed information

在鱼类个体识别的实际应用场景中,由于水下环境噪声大、鱼体角度倾斜以及类内特征差异不明显,导致卷积神经网络特征提取能力低下,影响识别准确性。针对该问题,提出一种增强细节信息特征提取的鱼类个体识别算法(FishNetv1)。改进YOLOv5网络并建立损失函数,优化鱼类个体目标的检测结果。主干网络在MobileNet‐v1的基础上完成优化,改进深度卷积层,更新ReLU激活函数,使用Leaky ReLU保留负值特征信息,实现特征信息的获取。在网络结构末端全连接层前增加特征加权层,去除卷积神经网络中常用的池化层,完成图像细节信息的增强和特征提取。实验结果表明,所设计模型在DLOUFish数据集上的平均准确率为92.46%,最高准确率达到95.69%。

In practical application scenarios for fish individual recognition,due to high underwater environmental noise,tilted fish angles,and insignificant intra class feature differences,the feature extraction ability of convolutional neural networks is low,which affects the recognition accuracy.On this basis,The YOLOv5 network is improved and a loss function is established to optimize the detection results of fish individuals.Based on MobileNet‐v1,the main network is optimized,deep convolution layers are improved,ReLU activation functions are updated,and Leaky ReLU is used to retain negative feature information for feature acquisition.A feature weighting layer is added before the fully connected layer at the end of the network structure to enhance the fine details in the images and extract features by removing the pooling layer commonly used in convolutional neural networks.The experimental results show that the proposed model achieves an average accuracy of 92.46%and a maximum accuracy of 95.69%on the DLOUFish dataset.