Developing a new explainable artificial intelligence model

This thesis work addresses the development and implementation of four new methods based on Class Activation Maps (CAMs), a visual approach in the field of Explainable Artificial Intelligence (XAI). These methods, designed based on Convolutional Neural Networks (CNNs), are named CodCAM, HayCAM, HayCAM+, and HayCAMJ, respectively. CodCAM is a method that generates CAMs using four different visual XAI techniques (GradCAM, GradCAM++, LayerCAM, EigenCAM). The first study using CodCAM examines the classification of thermal images belonging to newborn babies. The results of this study show that the CNN model focuses on anatomical regions of babies (such as arm, armpit, head, neck, foot, and body). These results demonstrate that the CNN model not only captures the overall image but also recognizes the anatomical structures of babies and uses this information for accurate classification. HayCAM reduces the filters in the final layer of the CNN using Principal Component Analysis (PCA) during CAMs generation, resulting in focused CAMs on the image. When HayCAM is used with a classification model to distinguish people wearing masks from those who aren't, the generated CAMs show a concentration around the mouth region. These generated CAMs are used to measure object detection performance, and Intersection over Union (IoU) values are calculated. These values are found to be 0.1922 for GradCAM, 0.2472 for GradCAM++, 0.3386 for EigenCAM, and 0.3487 for the proposed HayCAM. HayCAM+ is an enhanced version of HayCAM that aims to automatically calculate the essential filters used in the dimension reduction process to obtain CAMs. With HayCAM+, the IoU value increases by approximately 2.5% from 0.3487 to 0.3740. HayCAMJ, on the other hand, aims to create CAMs using only a single filter. This method offers a simple yet effective approach to explaining object classification results obtained with a single filter. The developed CodCAM, HayCAM, HayCAM+, and HayCAMJ methods are compared using Resnet18, Resnet34, Resnet50, and Mobilenet CNN models on the Pascal VOC datasets. The highest IoU results for Pascal VOC 2007 are 0.2422 for HayCAM (Resnet34), 0.3303 for HayCAM+ (Resnet18), 0.3277 for HayCAMJ (Resnet18), and 0.4079 for CodCAM (Resnet34). Similarly, for Pascal VOC 2012, the results are 0.2484 for HayCAM (Resnet34), 0.3101 for HayCAM+ (Resnet18), 0.3118 for HayCAMJ (Resnet34), and 0.3674 for CodCAM (Resnet34). These developed methods contribute to visualizing the decisions made by the CNN models, enhancing object detection performance using the visualized outputs.