科研成果详情

Because the background information of the multi-label natural scene image is complex and contains many kinds of things at the same time, it is a big challenge to improve recognition accuracy. At present, most methods usually give the same importance to each label for independent prediction, ignoring the correlation between labels. In this paper, an algorithm named Label Correlation based K-Nearest Neighbor (LC-KNN) method is proposed through analyzing and weighting the essential correlation between natural instances in the images. Considering that the label correlation of similar samples is also similar, this method is based on the local weighted method to mark the maximum cross-correlation label with high probability and the minimum cross-correlation label with low probability. It firstly finds out the neighbor samples of each sample to construct a multi-label count vector for the test sample according to the multi-label information of the neighbor samples, and then calculates the weight between the related labels based on naive Bayes model, and finally obtains the statistical correlation between the features and the labels to build classifier, which is more in line with the inherent law of the combination of things in natural scenes. The experimental results on the natural scene dataset show that the LC-KNN algorithm is significantly better than mainstream multi-label learning algorithms such as RELIAB, ML-KNN, Rank-SVM, and BoosTexter in tasks of multi-label natural scene recognition.