Enhancing Automatic Polyp Detection Accuracy Using Fusion Techniques

In this paper, we address the problem of automatic polyp detection in endoscopy videos. We propose feature fusion and multiple classifier technique using a variety of features that include wavelet features, Local Binary Patterns and Gabor features. We show that such a combination can mitigate, to a reasonable extent, the high rate of false positives common in this particular problem. Moreover, we study the effect of specular reflection removal on performance and show that, despite being a common preprocessing step, it can in some cases lead to worse results. Experiments conducted with ASU-Mayo Clinic Polyp Database confirm the validity of our scheme.