基于可见光谱和卷积神经网络的贝类识别方法

目前卷积神经网络（CNN）在物体种类识别方面取得突破性进展。贝类作为农业经济的重要组成部分，种类繁多，特点复杂，大多贝类存在着相似度高，各类样本分布不均衡情况，以致CNN对贝类分类的准确率偏低。针对这一情况，提出了基于可见光谱和CNN的贝类识别方法，旨在提取更有效的贝类特征，从而提高贝类分类的准确率。首先，提出了一种包含输出熵度量和正交性度量的滤波器信息度量与特征选择方法，重新初始化修剪掉的滤波器并使其正交，捕获网络激活空间中的不同方向，使神经网络模型学习到更多有用的贝类特征信息，提升模型分类准确率；其次，提出了一种包含正则化项和焦点损失项的贝类分类目标函数，通过控制各类别样本对总损失的共享权重，来减少易分类样本的权重，以使模型注意力向预测不准的样本倾斜，均衡样本分布和样本分类难度，进一步提高贝类分类的准确率。贝类图像数据集由74类贝类组成，共11 803张图像。获取原始数据集后，对数据集图像进行水平翻转、垂直翻转、随机旋转、在0, 30°]范围内旋转、在0, 20%]范围内缩放和移动等数据增强操作，将图像数量从11 803张增加到119 964张。整个图像数据集按8∶1∶1的比例随机分为训练集95 947张图片、验证集11 996张图片和测试集12 021张图片。在建立贝类图像数据集的基础上进行了实验验证，达到了93.38%的分类准确率，将基准网络（Resnest）的准确率提高了1.18%，相较网络SN_Net和MutualNet，准确率分别提升了4.34%和0.85% ，并且训练时长为22 320 s，将基准网络（Resnest）的训练时长缩短了960 s，训练时长分别比SN_Net和MutualNet短3 180和2 460 s。实验结果证明了该方法的有效性。

Recognition of Shellfish Based on Visible Spectrum and Convolutional Neural Network

At present, Convolutional Neural Network (CNN) has made a breakthrough in species recognition. As an important part of the agricultural economy, shellfish has a wide variety of species with complex characteristics. Some of the shellfish are highly similar and the distribution of various samples is unbalanced, which causes a low accuracy of CNN classification. In view of this situation, a shellfish recognition method based on visible spectrum and CNN is proposed in this paper, which aims to extract more effective shellfish features to improve the classification accuracy of shellfish. Firstly, a filter information measurement and feature selection method including output entropy measurement and orthogonality measurement is proposed, which reinitializes the pruned filter and makes it orthogonality, captures different directions in the network activation space, so that the neural network model can learn more useful shellfish feature information and improve the classification accuracy of the model; secondly, a shellfish classification objective function including regularization term and focal loss term is proposed, which reduces the weight of easily classified samples by controlling the shared weight of each sample to the total loss, it tilts the attention of the model to the samples with inaccurate prediction, so as to balance the distribution of samples and the difficulty of sample classification, and improve the accuracy of shellfish classification. The shellfish image dataset in this paper consists of 74 shellfish species with 11 803 pictures in total. After obtaining the original dataset, data augmentation which consists of horizontal flipping, vertical flipping, random rotation, rotation within the range of 0, 30°], scaling and moving within the range of 0, 20%] and moving is performed on the images of the dataset, increasing the number of images from 11 803 to 119 964. The whole image dataset is randomly divided into training set with 95 947 pictures, validation set with 11 996 pictures and test set with 12 021 pictures in an 8∶1∶1 ratio. In this paper, based on the establishment of the shellfish image dataset, the experimental verification has reached the classification accuracy of 93.38%, which increases the accuracy of the benchmark network (Resnest) by 1.18%. Compared with SN_Net, and MutualNet, the accuracy of the proposed method is increased by 4.34% and 0.85%, respectively. And the training time is 22 320 seconds, which shortens the training time of the benchmark network (Resnest) by 960 seconds, the training time of the proposed method is 3 180 seconds and 2 460 seconds shorter than SN_Net and MutualNet, respectively. The experiments results demonstrate the effectiveness of the proposed method.