This paper evaluates a new classifier method for image classification and retrieval. The new method is called Nearest Feature Line (NFL) method and is compared to the alternative existing nearest-neighbor (NN) and nearest-center (NC) methods. The NFL method has a better classification error rate when applied to the Brodratz texture database and better retrieval efficiency when applied to the color images in the MIT VisTex database. Unlike the alternative methods, NFL requires more than one prototype for each image class. As the authors point out, NFL does better than its rivals because it uses extra information implicit in the correlations between class prototypes.

The authors use multiple class prototypes to reduce image classification error rates and to improve image retrieval efficiency but there is still a lot of room for improvement in image classification and retrieval. Higher-level features that correspond better to human perception need to be found. Learning through relevance feedback may also help. The use of text from image captions, where available, is another way to get ahead. Image retrieval and classification are fun and valuable tasks and we have a long way to go before we achieve helpful internet image searches and “Yahoo!”-type image directories. This paper provides an incremental but helpful and clearly explained step along this path.