تشخیص تصویر نمونه کوچک با استفاده از شبکه عصبی پیچشی بهبود یافته Small sample image recognition using improved Convolutional Neural Network

1. Introduction With the development of science and technology, image processing technology has improved a lot [1–6]. A lot of works has been done about image processing [7–9]. Image recognition is an important field of artificial intelligence. As the development of image processing, image recognition technology has been gradually applied in many fields [10–14].At the same time, the accuracy, reliability and real time requirements of image recognition are becoming stricter. Recent years, as the development of deep learning, CNN based on it has been used in many field about image processing. Owing to the close connection between the layers of CNN and the sufficient space information of CNN, CNN can work well on image processing and understanding task. CNN can even extract rich correlated features automatically from images. On account of above features of CNN, it has achieved excellent results in all kinds of image recognition tasks such as face recognition, eye detection and pedestrian detection [15–17]. Though CNN has achieved big success in image recognition, it still has its own limitations. The BP neural network inside CNN model is too simple, so it needs multiple iterations with a large number of training samples. In other words, it can learn the image representation well only when it has enough training data and iterate sufficient times. The BP neural network adopts the descending gradient training method, which make the model converge slowly and it easily come to the local optimization, affecting the final recognition accuracy. So we propose a new hybrid model CNN-GRNN, it can get excellent performance even with small sample. GRNN can get ideal recognition result even it do not has enough feature and it do not need iteration. Our method on image recognition consists of two parts: 1) In the training time, the CNN-GRNN model use CNN to extract the image representation, and then it let the full connected layer to do the prediction work. 2) In the testing time, CNN is responsible for the representation extraction task as in the training time, and then GRNN will classify the image using the extracted feature, which is different from the training. This model aims to establish relevance between the image representations and objective prediction result.