Abstract
Image classification and interpretability is a basic procedure in modern AI systems as well as being employed in the fields of security, instrumentation, and image classification. Explainable Artificial Intelligence also plays a crucial role in providing machine learning techniques with a white-box output that allows the practitioner to follow or interpret the rules. Therefore, we present CACAO-X which is a framework for ML forecasts that utilizes an algorithm which converts images into time-series and employs CNN, LSTM, and MLP on the original images and converted data in a mixed input approach achieving accurate forecasts on multiple applications. Furthermore, Shapley values are employed on the proposed ML models to extract the local explainable features of such images. These features are again used with the same conversion algorithm to generate time-series and application specific landmarks using the Logic Learning Machine (LLM) to determine the explainable rules. In addition to being generated automatically, explicative labelling is adopted in the dataset which give insight to the operator to which landmarks, positions and angles contribute to the classification of each record. The system is robust as it achieves benchmark results in the gender classification application while remaining competitive in all others where it is used as a reference model for global image explainability. Additionally, the ML inference for every application is performed according to the edge-computing paradigm, by deploying the models on a limited resources Raspberry Pi platform through SSH access. Moreover, a novel callback function has been developed which is able to stop the training of the CACAO-NET model at a stable state based on the level of oscillation around the most recent mean accuracy.