基于前向神经网络的图像检索相关反馈算法设计

A Forward Neural Network Based Relevance Feedback Algorithm Design in Image Retrieval

相关反馈技术是近年来较为重要的图像检索方法 .该文从机器学习的角度出发 ,提出了一种基于神经网络的相关反馈算法 .在检索过程中 ,用户可以标记出与查询图像相似的正例样本反馈给系统 ,然后由系统构造出前向神经网络并再次进行检索 ,以改进查询结果 .该文对用于训练前向神经网络的构造性算法在图像检索中的几何意义进行了深入的研究 ,并在此基础上给出平移算法、反例标记学习算法、球半径系数自适应算法等 ,从而使基于神经网络自学习的相关反馈算法更加完善 .实验表明 ,改进后的算法在图像检索中具有更好的性能和更强的推广能力 .

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.