Exploring space-frequency co-occurrences via local quantized patterns for texture representation

Local Binary Pattern (LBP) has shown its power in texture classification and face recognition. However, the LBP operator is performed in the original image space, and it lacks deeper pixel interactions to capture a richer description. In this paper, we propose to explore space-frequency co-occurrences via local quantized patterns for texture representation. The proposed method proceeds in two channels. In each channel, the multi-resolution spatial maps are first obtained by specific spatial filtering, and local frequency features (spectral maps) are subsequently extracted by applying the local Fourier transform to the spatial map. Two types of quantization via global thresholding are employed to quantize the spatial and spectral maps into three and two levels, respectively. The quantized patterns are then jointly encoded to construct a space-frequency co-occurrence histogram. Finally, the two-channel histograms are combined to characterize the texture. Without resort to the texton-based representation, our method directly encodes the joint information in the space and frequency domains while preserving the robustness to image rotation, illumination, scale and viewpoint changes. Extensive experiments have been conducted on three well-known texture databases, and our method achieves the best classification results compared with state-of-the-art approaches investigated.