ADVANCED EXPLAINABLE DEEP LEARNING MODELS FOR IMAGE CLASSIFICATION

• Main Author: Irawan, Mellisa
• Format: Final Project
• Language: Indonesia
• Subjects: Teknik (Rekayasa, enjinering dan kegiatan berkaitan)
• Online Access: https://digilib.itb.ac.id/gdl/view/75059
• Institution: Institut Teknologi Bandung

This research focuses on the application of DenseNet121 neural network architecture to address two distinct classification problems and deploy SHAP, GradCAM, and SmoothGrad as the explainability model. The primary purpose of this work is to create a software to classify oil palm fresh fruit bunches (FFB) based on ripeness, with the aim of enhancing the productivity of oil palm harvesting using automated drone technology. An additional case is added to test the result from the previous case. The second problem is to classify the severity of ulcerative colitis to improve the diagnostic capabilities of medical professionals. For the oil palm classification problem, the model was constructed using annotated video data from Suharjito, G. N. et. al, achieving an accuracy score of 0.9958 in classifying oil palm FFB into three categories: raw, good harvest, and bad harvest. The three explainability models successfully validate that the FFB is an important feature for the model’s classification decision. In the case of ulcerative colitis classification, the HyperKvasir dataset was utilized, consisting of 851 ulcerative colitis images categorized by Mayo Scores. The developed model achieved an F1 score of 0.9156 in distinguishing between mild and severe ulcerative colitis. GradCAM and SmoothGrad, successfully identify blood, lumps, or white lesions indicative of ulcerative colitis.