基于深度学习的鱼类分类算法研究

FISH CLASSIFICATION ALGORITHM BASED ON DEPTH LEARNING

回顾近年来国内外对鱼类分类的研究进展,指出传统方法存在的缺陷。深度学习是目前图像分类的主流方法。研究基于卷积神经网络CNN(Convolutional Neural Network)的鱼类分类模型,并以该模型为基础,进一步提出利用迁移学习,以预训练网络的特征结合SVM算法(Pre CNN+SVM)的混合分类模型。实验以Fish4-Knowledge(F4 K)作为数据集,使用Tensor Flow训练网络模型。实验结果表明,利用Pre CNN+SVM算法,取得了98.6%的准确率,较传统方法有显著提高。对于小规模数据集,有效解决了需要人工提取特征的不可迁移性。

Reviewing the research progress of fish classification at home and abroad in recent years, the shortcomings of the traditional methods are pointed out. Deep learning is the mainstream method of image classification. This paper studied the fish classification model based on Convolutional Neural Network （CNN）. Based on this model, we proposed a hybrid model based on transfer learning algorithm using pre-training neural network and SVM algorithm （PreCNN ＋ SVM）. The experimental results showed that using PreCNN ＋ SVM algorithm with Fish4-Knowledge （F4K） data sets and TensorFlow model, the accuracy of 98.6% was achieved, which was significantly higher than the traditional method. For small-scale data sets, it effectively solved the immutability that need to extract features manually.