基于同步重建与分类的深度自编码的分类网络

Image Classification Based on Deep Autoencoder for Synchronous Reconstruction and Classification

图像分类一直是计算机视觉领域中的研究热点,传统的图像分类研究多数为基于粗粒度图像分类,即不同类别的图像分类,而对于类间差异小,类内差异大的细粒度分类仍存在着挑战性。文章基于深度自编码框架提出采用一致性重建与分类损失函数来实现同步训练的深度自编码分类网络,可适用于仅用类别标签的细粒度图像分类。此外,文章还提出了一种加入自适应平均池化操作的分类网络,使得模型对不同尺寸的输入图像都适用,且该方法不仅可以实现二分类还可以实现多分类任务。在公开数据集Oxford 102 Flowers和ZJU-Leaper上分别展开多分类和二分类实验,平均分类精度分别达到了85.51%和95.43%。

Image classification has always been a research hotspot in the field of computer vision. Most of the traditional image classification research is based on coarse-grained image classification, that is, image classification of different categories. However,there are still challenges for fine-grained classification with small inter class differences and large intra class differences. Therefore,based on the deep autoencoder framework, this paper proposes a classification network using consistency reconstruction and classification loss function to realize synchronous training, which can be suitable for fine-grained image classification using only category labels.In addition, in this paper proposes a classification network with adaptive average pooling operation, which makes the model applicable to input images of different sizes, and this method can not only realize binary classification, but also realize multi classification tasks. In this paper, multi classification and binary classification experiments are carried out on the public data sets Oxford 102 flowers and ZJU Leaper, and the averageclassification accuracy is 85.51% and 95.43% respectively.