Intelligent Security Image Classification on Small Sample Learning

Abstract

Image classification is an important task in the field of the intelligent security and deep learning methods represented by convolutional neural networks have achieved many great results in this field. Image classification based on deep learning usually performs well on large-scale datasets, but its performance is often greatly limited by the size of the data. When the dataset is not sufficient, the traditional deep learning method cannot perform well on the small-scale datasets and this situation often occurs in practical applications. To address the drawback, we propose a deep learning framework based on the combination of SoftMax classifier and Bayes learning for small-sample image classification. Within this framework, we utilize transfer learning to solve the problem of too few data, and it can also reduce model training time and space costs. At the same time, we make use of the combination of the above two classifiers to improve the effectiveness and accuracy of the model on different datasets. We empirically find that the model has higher classification accuracy and less training time than the general deep learning model on the datasets. The experiment results demonstrate that the proposed method generally has better classification accuracy on the small-scale datasets, compared with mainstream methods.