| Dermatologists base the diagnosis of skin disease on the visual assessment of the skin. This fact shows that correct diagnosis is highly dependent on the observer's experience and on his or her visual perception. Moreover, the human vision system lacks accuracy, reproducibility, and quantification in the way it gathers information from an image. So, there is a great need for computer-aided diagnosis.;We propose a content-based image retrieval (CBIR) system to aid in the diagnosis of skin disease. First, after examining the skin images, pre-processing will be performed. Second, we examine the visual features for skin disease classified in the database and select color, texture and shape for characterization of a certain skin disease. Third, feature extraction techniques for each visual feature are investigated respectively. Fourth, similarity measures based on the extracted features will be discussed. Last, after discussing single feature performance, a distance metric combination scheme will be explored.;The experimental data set is divided into two parts: developmental data set used as an image library and an unlabeled independent test data set. Two sets of experiments are performed: the input image of the skin image retrieval algorithm is either from developmental data set or independent test data set.;The results are top five candidates of the input query image, that is, five labeled images from image library. Results are laid out separately for developmental data set and independent test data set. Two evaluation systems, both the standard precision vs. recall method, and the self-developed scoring method are carried out. The evaluation results obtained by both methods are given for each class of disease.;Among all visual features, we found the color feature played a dominating role in distinguishing different types of skin disease. Among all classes of images, the class with best feature consistency gained the best retrieval accuracy based on the evaluation result. For future research we recommend further work in image collection protocol, color balancing, combining the feature metrics, improving texture characterization and incorporating semantic assistance in the retrieved process. |
