Investigation of the design and use of artificial neural networks in the CBIR

Abstract- Artificial neural networks (ANNs) are used for content based image retrieval (CBIR). Training of a neural network requires that the user specifies the network structure and sets the learning parameters. In this study, the optimum design of ANN's for retrieval of images is investigated. We use test image datasets in a series of experiments that evaluate the effects on network performance (measured in terms of the Mean Square Error and number of iteration (time required for training)) of different choices of network size and structure, network parameters, training samples size. We use a test image database of 1000 images including 10 classes. Each class has 100 images. The backpropagation algorithm, also called the generalized delta rule, was used for neural network training. An activation function was hyperbolic tangent. Experiments show that with used category of images optimal number of neurons in hidden layer is half of number of images in used training set. We intend to design small CBIR system for educational purposes and potentially for mobility environment