Multilevel Image Coding with Hyperfeatures

摘要

Histograms of local appearance descriptors are a popular representation for visual recognition. They are highly discriminant with good resistance to local occlusions and to geometric and photometric variations, but they are not able to exploit spatial co-occurrence statistics over scales larger than the local input patches. We present a multilevel visual representation that remedies this. The starting point is the notion that to detect object parts in images, in practice it often suffices to detect co-occurrences of more local object fragments. This can be formalized by coding image patches against a codebook of known fragments or a more general statistical model and locally histogramming the resulting labels to capture their co-occurrence statistics. Local patch descriptors are converted into somewhat less local histograms over label occurrences. The histograms are themselves local descriptor vectors so the process can be iterated to code ever larger assemblies of object parts and increasingly abstract or ‘semantic’ image properties. We call these higher-level descriptors “hyperfeatures”. We formulate the hyperfeature model and study its performance under several different image coding methods including k-means based Vector Quantization, Gaussian Mixtures, and combinations of these with Latent Dirichlet Allocation. We find that the resulting high-level features provide improved performance in several object image and texture image classification tasks.