A Forward Neural Network Based Relevance Feedback Algorithm Design in Image Retrieval
Lei Zhang · Chinese Journal of Computers · 2002
By transferring the process of relevance feedback into a learning problem of neural network, this paper proposes a novel learning method based on the Constructive Learning Algorithm (CLA), which describes the samples distribution with a set of sphere neighborhoods and constructs the neural network directly. From the training data of positive and negative samples marked by users, a feed forward neural network could be constructed directly by CLA. Then this learned neural network can be used to measure the similarity between the query concept and each image in the database. Thus more images relevant to the query can be retrieved. This paper studies the geometrical representation of the CLA which is used to train the neural network in image retrieval. From the geometrical representation, some new algorithms, including sphere moving, negative sample learning, radius coefficient self adapting and sphere distance, are proposed to improve the previous retrieval result. Contrasting to the traditional re weighting method of relevance feedback, which assumes that relevant images conform to the single Gaussian distribution, CLA does not make any assumption to the training data but try to capture the distribution of the positive samples by means of sphere neighborhoods. Experiments were carried out on a large size database of 9918 images. It shows that more images relevant to the query can be found efficiently by the interactive learning and retrieval processing. The experimental result shows that the algorithms have better performance and generalization ability and are able to fulfill the user's requirement.