<![CDATA[Road damage such as cracks, potholes, and uneven surfaces poses serious risks to transportation safety, logistics efficiency, and maintenance budgeting in Indonesia. Manual inspection is time consuming, labor intensive, and prone to error, motivating the use of reliable computer vision solutions. This study proposes MYCD, a hybrid and mobile ready architecture that combines the fast detection ability of YOLO with the dense feature reuse of DenseNet, enhanced by the Convolutional Block Attention Module (CBAM) for spatial and channel focus and Spatial Pyramid Pooling (SPP) for multi scale context understanding. The system detects and classifies the severity of road damage into minor, moderate, and severe categories using images captured by standard cameras. MYCD was trained and validated on 1,120 field images using an 80/20 split to simulate realistic deployment. Validation achieved 64 percent accuracy, with the highest per class precision of 0.72 for minor damage and mAP@0.5 = 0.677. The confusion matrix showed that most errors occurred in the moderate category because of visual similarity with minor and severe damage. Unlike earlier studies that extended YOLO with heavy backbones such as ResNet or EfficientNet, MYCD focuses on feature propagation (DenseNet), attention precision (CBAM), and multi scale fusion (SPP) optimized for real time operation on standard hardware. Efficiency profiling confirmed its deployability. After compression, the model size is 46.8 MB and it requires 3.7 GFLOPs per inference at 640×640 resolution. On a mid-range Android device (Snapdragon 778G, 8 GB RAM), MYCD runs at 19 frames per second with 1.2 GB peak memory. Compared with YOLOv8 WD (68 MB; 5.2 GFLOPs), MYCD reduces computation by 31 percent while maintaining similar accuracy. Overall, MYCD achieves a practical balance of speed, accuracy, and efficiency, providing a deployable and reproducible framework for real time road damage detection in resource limited settings.]]>
Copyrights © 2026
