Low-rank Shared Dictionary Learning with Incoherence Constraint for Endoscopic Gastrointestinal Image Classification

Endoscope has been widely used in clinical examination of gastrointestinal diseases. Many automatic endoscopic image classification algorithms based on dictionary learning are proposed to assist doctors in diagnosing diseases, where the learning method of shared dictionary and class-specific dictionaries enables training dictionary to be more discriminative. Nevertheless, in the process of dictionary learning, the appearance of common features in class-specific dictionaries may cause low classification accuracy. To remedy this deficiency, herein we introduce a coherence constraint between low-rank shared dictionary and class-specific dictionaries. The proposed dictionary learning method is applied to the classification system of endoscopic gastrointestinal images, including normal, polyp and ulcer images, whose experimental results prove that it has promising classification performance.