Adaptive color discrimination for image classification

Semantic understanding of images remains an important research challenge in machine intelligence and statistical learning. It mainly includes two steps: feature extraction and classification. This study mainly focuses scene image recognition, where color information plays an important role. The conventional color representation of images mainly includes color distribution (histogram) and its statistical information based on uniformed quantization color bin. However, for a specific recognition application such several scene types, some quantized colors maybe never appear in any scene image, and at the same time the detail variation in other quantized colors include much discriminative features. Therefore, this study proposes to characterize the color information of scene images using a leaning strategy for producing adaptive color level, and extract the histogram of the learned color levels for image representation. With the proposed strategy, the compact (learned) color levels can represent the image in our application more faithful than the uniform quantized conventional RGB color. Experimental results show that the recognition rate with our proposed methods can be greatly improved compared to the conventional color histogram.