Abstract: Deep learning has become a research hotspot in the field of image recognition.Different from traditional image recognition methods,deep learning is to automatically learn features from a large amount data and has a strong ability of feature learning and representation.However,under the condition of small samples,the traditional deep learning methods such as convolutional neural network are difficult to learn effective features,resulting in low image recognition accuracy.Thus,a new image recognition algorithm under small samples was proposed to solve the classification and recognition of SAR images.On the basis of convolutional neural network,it combines convolution operation with autoencoder to form a deep convolutional autoencoder network structure.The algorithm firstly preprocesses the image and enhances the image using 2D Gabor filter,and thentrains the model,finally,constructsthe image classification model.The proposed model can automatically learn and extract effective features from small sample images,and improve the recognition accuracy.On 10 categories of target classification of MSTAR data set,10% samples from the training data were selected as new training data,the rest were valid data,and the recognition accuracy of the test data in the convolutional neural network is 76.38%,while that in the proposed convolutional autoencoder is 88.09%.Experimental results show that the proposed algorithm is more effective than convolutional neural network in small sample image recognition.

Key words: Autoencoder, Convolutional neural network, Deep learning, Few-shot learning