Performance Evaluation of Handcrafted Feature Extraction Techniques using Bayes Net Classifier for Skin Disease Image Classification
Ayyappan G., Senthil K
Abstract
The medical field requires increased reliance on automated skin disease diagnosis through image classification because both dermatological disorders are more prevalent and patients require accessible clinical help. This research examines how Bayes Net performs as a classification system by using six different handcraft descriptor techniques when analyzing high-resolution dermatological images. The analysis incorporates six particular feature descriptors which are PHOGF (Pyramid Histogram of Oriented Gradients), SCHF (Simple Color Histogram Filter), Gabor Filter, JPEG Coefficient Filter (JPEGCF), FCTH (Fuzzy Color and Texture Histogram Filter), and FOHF (Fuzzy Opponent Histogram Filter). The research utilized publicly available images from a ten-disordered skin diseases collection for the experiments. The evaluation of classifiers used Accuracy, Precision, Recall together with ROC (Receiver Operating Characteristic) and PRC (Precision-Recall Curve) and computational time measurement metrics. A combination of Gabor Filter with Bayes Net produced the most effective result with an accuracy rate of 97.59% alongside 0.98 precision and recall and 0.99 ROC and PRC scores at a computation time of 0.02 seconds. PHOGF+Bayes Net generated 95.71% accuracy while SCHF+Bayes Net produced 95.49% accuracy but the extended computational time reached 33.63 seconds. The combination of JPEGCF with Bayes Net allowed an efficient execution time to produce 92.83% accuracy results. This research confirms that the combination between custom-designed features and the probabilistic Bayes Net classifier leads to highly precise solutions that maintain rapid computational effectiveness for skin disease diagnosis systems.