A Feature-Level Approach for Outdoor Surface Classification

Abstract

In this study, we address the problem of outdoor surface classification using deep convolutional neural networks. A custom dataset was generated to represent various outdoor ground types under different conditions. State-of-the-art CNN models, including VGG19, ResNet50, and InceptionV3, were employed and evaluated individually on this dataset. Based on the comparative results, we proposed a feature-level fusion model that combines VGG19 and InceptionV3 to leverage their complementary strengths. Experimental results show that the proposed fusion model significantly outperforms the individual models, achieving 97% in precision, recall, F1-score, and test score. These findings demonstrate the effectiveness of the ensemble approach in improving classification performance for outdoor surface recognition tasks.