基于深度学习的小龙虾分级算法

Crawfish Classification Algorithm Based on Deep Learning

传统人工分拣小龙虾存在着分拣速度慢、效率低、人工代价高等问题。针对小龙虾人工分级存在的问题,提出基于卷积神经网络的小龙虾分级算法,该算法以预训练的EfficientNetB0网络为基础框架,为了能更好地实现知识迁移,通过两层全连接层对EfficientNetB0特征进行迁移,并利用ReLU激活函数。利用TensorFlow平台来实现算法,并利用Adam优化器对两层全连接层和线性分类层进行训练。为了模拟实际生产线,分别评估提出算法在干净图像数据集、不同类型噪声图像数据集的性能,并与经典的ResNet50、VGG16、VGG19等卷积网络进行性能对比。实验结果表明EfficientNetB0在不同小龙虾数据集上具有较好的泛化性能,在干净图像、高斯噪声图像和椒盐噪声图像上的识别率分别达到99.70%、93.75%和88.41%。

The traditional manual sorting of crayfish is characterized by slow sorting speed,low efficiency and high labor cost.According to the prob⁃lem of crawfish artificial classification,this paper proposes a crawfish classification algorithm based on convolutional neural network.The algorithm uses the pretrained EfficientNetB0 network as the basic framework.In order to achieve better knowledge transfer,we employ two Full Connection(FC)layers and utilizes the ReLU activation function to transfer the EfficientNetB0 features.The algorithm is implement⁃ed using the TensorFlow platform,and the Adam optimizer is used to train the two of FC layers and the linear classification layer.In order to simulate the practical applications,the performance of proposed algorithm on the clean images data set and different types of noisy imag⁃es data sets are evaluated respectively,by comparing with the classical convolutional network including the ResNet50,VGG16 and VGG19.The results show that EfficientNetB0 has better generalization performance on different crayfish data sets,with the recognition rate of clean images,Gaussian noisy images and Salt and Pepper noise images as 99.70%,93.75%and 88.41%,respectively.