Automatic method for detecting specular reflection and motion blur artifacts on endoscopic images using complementary binary classifiers

Computer Aided Diagnosis (CAD) tools have demonstrated high performance in the identification of gastrointestinal diseases through endoscopic images (EIs). However, such diagnostic support tools could be affected by image artifacts which may appear in real videos, making that precise artifact detection become in a crucial step for training such supporting tools, even those based on convolutional neural networks (CNN). This work presents an automatic method for detecting the two most frequent artifacts in EIs, specular reflections (SR) and motion blur (MB), as a pre-processing tool for identifying informative frames, suitable for training automatic methods used in CAD tools. The proposed method identifies artifact patterns by utilizing coherence features, between regions with low and high frequencies (brightness, contrast, Comparative Gaussian- Frame Changes- CGFC), and using them to feed two complementary binary classifiers, achieving a precision of 96 % for the identification of SR and 76 % for MB.