Explainable deep learning for car image classification via concept bottleneck models

This study presents an innovative deep learning application of the Concept Bottleneck Model (CBM) architecture for the classification of car images, with a particular focus on interpretability and transparency in the field of artificial intelligence. The main objective of the research is not only to develop a system that can classify vehicle types with high accuracy, but also to construct a model capable of explaining its decision-making process through human-understandable visual concepts. In this context, a two-stage CBM architecture was implemented. In the first stage, a ResNet-18 based concept predictor model was trained to estimate 20 binary visual attributes such as "2-door," "sunroof," or "boxy" from car images. In the second stage, a fully connected neural network was trained to predict the vehicle class labels using the predicted conceptual outputs as input. The model was designed to distinguish between 8 car types: Cabrio, Coupe, Crossover, Hatchback, Pickup, Sedan, StationWagon, and SUV. For the training and evaluation of the model, a custom dataset composed of labeled vehicle images and manually annotated visual concepts was used. Experimental results demonstrated that the model achieved a classification accuracy of 82% and was able to provide conceptual explanations for each prediction. Unlike traditional end-to-end "black box" models, the proposed approach significantly enhances model interpretability by enabling decision processes to be traced through internal concepts. In this way, it contributes to critical engineering processes such as human-centered validation, model debugging, and system reliability analysis. Future work will focus on integrating uncertainty estimation and expanding the dataset to improve the model's generalization capability.