Small Sample Image Recognition Based on CNN and RBFNN

Abstract

Identification of dangerous goods based on images plays a key role in the security inspection of various situations such as airports, subways, public places etc. This paper discusses the issue in a from-simple-to-complex manner. Firstly, we classify different kinds of knives given an image including a single object without complex background in the framework of TensorFlow. Then, according to the color and shape features of a single image, where Fourier transform and Roberts operator is used to judge of the complex scene which doesn’t contain knives from an image with natural background. Finally, convolution neural network (CNN) and radial basis function neural network (RBFNN) are used to construct identification models for images of objects in six categories. The obtained accuracy of the true and predicted values of the CNN and RBFNN are 66.67% for training on CNN and 76.67% on RBFNN, for testing 50% on CNN and 44.44% on RBFNN respectively. The results showed that the constructed of identification model is able to perform recognition for small-scale image database and reduce the false alarm rate. Furthermore, our method is robust in dealing with the small sample, with high classification accuracy and low cost. The models have few layers and nodes.