Image scene classification based on discriminative clustering and weighted fusion representation

In this paper, a new discriminative clustering method was proposed. For each traversing sliding block in the training samples, the deep features were extracted and their correlation distance matrices were obtained. The cluster was obtained by traversing the correlation distance matrix in turn from the minimum value, and the average value of each cluster was taken as the clustering center to construct the visual sub-codebook of each image category. The visual sub-codebooks of all the image categories were concatenated to construct the universal visual codebook. After the image segmentation, the local description of the image was obtained based on the universal visual codebook. The integrated image representation has richer semantics information than the single local or global description of an image. To obtain a more accurate image representation, in this paper, we weighted and fused the local information and the global information to form the final image representation. In the experiment, classifications were performed on two common scene datasets. The results showed that the image fusion representation of local and global information achieved better performance of classification.