使用改进自编码器的细粒度图像分类研究

Improved Auto-encoder for Fine-grained Image Classification

针对深度连续聚类算法(Deep Continuous Clustering, DCC)特征提取能力有限,对复杂图像不能提取足够有效细节特征的不足,本文提出一个新的循环卷积自编码器(Recurrent Convolutional Auto-Encoder, R-CAE).自编码器结合门控循环网络GRU和卷积网络CNN构造编码层;同时在门控循环网络GRU部分添加空间域注意力通道,增强网络的特征学习能力.图像信息经过R-CAE自编码器编码后获取细节信息,传入经典卷积神经网络学习特征;当优化结果接近或者达到聚类阈值的时候,获得最终的聚类结果实现分类.训练过程中,模型首先预训练,确定自编码器参数;然后结合编码部分和经典网络学习训练,微调网络参数.本文通过实验证明了改进方法结合DCC在聚类实验中优于大部分经典聚类算法,在针对真实图像的细粒度分类实验中也有显著的进步.

In view of the limited feature extraction capabilities of the Deep Continuous Clustering(DCC)algorithm, this paper proposes a new Recurrent Convolutional Auto-Encoder(R-CAE),which is used to solve the problem that complex images cannot extract sufficient and effective detailed features.The autoencoder combines the gated recurrent network and the convolutional network to construct the coding layer;at the same time, the spatial domain attention channel is added to the GRU part of the gated recurrent network to enhance the feature learning ability of the network.The decoding layer of the self-encoder adopts an asymmetric CNN structure.The image information is first encoded by the R-CAE autoencoder, and the detailed information is obtained and then passed into the classic convolutional neural network to learn features;by continuously optimizing the objective function of the clustering algorithm, when the optimization result is close to or reaches the set clustering threshold, the final clustering result is obtained to realize classification.In the training process, the model adopts the pre-training method, firstly determine the autoencoder parameters;then combine the coding part and the classical network learning and training to fine-tune the network parameters.Experiments have proved that the improved method of this article combined with DCC is superior to most classic clustering algorithms in clustering experiments, and it has also made significant progress in fine-grained classification experiments for real images.