Unified multi-task YOLO-Lite framework for road defect detection and image-level classification with edge-optimized quantization and explainable visual inference

Abstract

Automated road defect assessment requires models that can accurately localize pavement damage, identify the overall defect category, and operate with limited computational resources. However, most existing approaches address object-level detection and image-level classification using separate networks, resulting in redundant computation and limited interaction between local and global representations. This study proposes YOLO-Lite , a unified lightweight multi-task framework that performs road defect detection and image-level classification within a single end-to-end architecture. The framework employs a shared depthwise-separable convolutional backbone with residual refinement, together with dedicated detection and classification heads. A composite multi-task objective jointly optimizes objectness prediction, bounding-box localization, instance-level defect classification, and image-level categorization, enabling complementary learning between the two tasks. On the RDD test set, YOLO-Lite achieves an mAP@0.50 of 0.907 and an image-level classification accuracy of 0.921 while requiring only 1.2 million parameters and 3.5 GFLOPs. Compared with recent lightweight detectors, YOLO-Lite achieves a superior balance between accuracy and computational efficiency. Extensive ablation, robustness, cross-dataset, and statistical analyses further validate its effectiveness and generalization capability. Quantitative and qualitative Grad-CAM analyses confirm that the learned activation regions correspond closely to annotated pavement defects. INT8 quantization further reduces model size and improves inference speed with only a marginal performance reduction. The reported GPU-based measurements indicate strong preliminary suitability for future edge deployment, although validation on physical embedded hardware remains future work.