Trade-off between speed and performance for colorectal endoscopic NBI image classification

This paper investigates a trade-off between computation time and recognition rate of local descriptor-based recognition for colorectal endoscopic NBI image classification. Recent recognition methods using descriptors have been successfully applied to medical image classification. The accuracy of these methods might depend on the quality of vector quantization (VQ) and encoding of descriptors, however an accurate quantization takes a long time. This paper reports how a simple sampling strategy affects performances with different encoding methods. First, we extract about 7.7 million local descriptors from training images of a dataset of 908 NBI endoscopic images. Second, we randomly choose a subset of between 7.7M and 19K descriptors for VQ. Third, we use three encoding methods (BoVW, VLAD, and Fisher vector) with different number of descriptors. Linear SVM is used for classification of a three-class problem. The computation time for VQ was drastically reduced by the factor of 100, while the peak performance was retained. The performance improved roughly 1% to 2% when more descriptors by over-sampling were used for encoding. Performances with descriptors extracted every pixel ("gridl") or every two pixels ("grid2") are similar, while the computation time is very different; grid2 is 5 to 30 times faster than gridl. The main finding of this work is twofold. First, recent encoding methods such as VLAD and Fisher vector are as insensitive to the quality of VQ as BoVW. Second, there is a trade-off between computation time and performance in encoding over-sampled descriptors with BoVW and Fisher vector, but not with VLAD.