Comparison of Classical and Reinvented Convolutional Neural Networks for Surface Type Classification

Abstract

This study compares two convolutional neural network (CNN) architectures for surface type classification. The first model employs a classical architecture achieving 96.62% accuracy during validation but struggles with recognizing complex surfaces like hills and potholes. The second model incorporates advanced features, including parallel processing paths and multi-level normalization, boosting accuracy to 99%. The training process utilized a dataset with augmented images of surfaces such as clay, hills, potholes, roads, and water-covered concrete. Metrics such as accuracy, recall, and F1-score were analyzed to evaluate performance. The modified CNN demonstrated superior capabilities in feature extraction and classification, particularly for heterogeneous terrains. Experimental results suggest that this enhanced architecture significantly reduces errors, improving adaptability to real-world conditions. Such improvements make it ideal for applications in robotics and autonomous systems. Future research will focus on expanding the dataset, further refining network architecture, and optimizing computational efficiency for deployment in field robotics.