Abstract: This work introduces an explainable deep learning system that is designed for road scene feature classification based on a convolutional deep learning network VGG16 in combination with SHAP for explanation. The system processes the input road images by resizing the images to 160 x 160 x 3, normalized the pixel values and made predictions from six possible classes: HV, LMV, Pedestrian, Road Damages, Speed Bump and UnsurfacedRoad. A filtering mechanism based on confidence is included, whereby if the predictions are below a threshold the results will be labeled Plain Road to avoid uncertain outcomes. In order to produce constant explanations, a balancing of data set using TensorFlow's ImageDataGenerator is created from test images and stored for SHAP initialization of background data set. SHAP is then used with an Image masker with inpaint_telea in order to generate pixel-wise attributions to indicate the most influential areas in the input image. The whole system is implemented using Streamlit interface including image uploading, real-time prediction, class probability visualization, and explanation heatmaps rendered using Matplotlib. All technical components such as preprocessing, caching, background generation and explainer setup are directly derived from the project's code which has been implemented. The resulting framework helps to add transparency in road feature classification with the combination of deep learning performance and interpretable visual outputs to help users understand what the model is reasoning about.

Keywords: RoadFeatureClassification DeepLearning VGG16 SHAP ExplainableAI Streamlit ImageProcessing ComputerVision RoadSafety FeatureVisualization.

Downloads: | DOI: 10.17148/IARJSET.2025.121223