Evidence record · UDINUS · SDGs
- Unit
- Rektorat UDINUS
- Period
- 2025
- Goals
- Record
- ea931fba-0dc0-4d8a-8749-073ef96dd5ca
- Published
- 2026-10-30
Universitas Dian Nuswantoro (UDINUS), through the participation of Dirganjaya, Sie Deta; Shidik, Guruh Fajar; Ardiena, Maulidya Ayu; Marendratama, Radhitya; and Kusuma, Edi Jaya, conducted a study on the architectural impact of transpose convolutional layers in convolutional neural network-based deep learning for pothole detection. The activity addressed the need for precise and real-time identification of road damage in intelligent transportation systems, where timely detection can improve safety and maintenance efficiency.
The study examined the replacement of static upsampling layers with ConvTranspose2d as a learnable upsampling mechanism in YOLOv5m and YOLOv8m models. The scope focused particularly on small and irregular potholes under challenging real-world visual conditions. These conditions included potholes that were partially obscured by water, surface debris, or shadows. By enabling the network to reconstruct fine spatial details adaptively, the approach was assessed through the detection performance of both models.
Experimental results demonstrated that adding transpose layers enhanced detection performance across YOLOv5m and YOLOv8m. YOLOv5m with transpose layers achieved a mAP@50 of 0.839, indicating a 1.57% improvement over the baseline. YOLOv8m showed an increase from 0.751 to 0.777, corresponding to a 3.46% relative gain. YOLOv5m maintained a more favorable balance between precision and inference latency, with precision of 0.894. YOLOv8m demonstrated higher generalization, reflected in recall increasing from 0.640 to 0.720, although its inference time also increased.
Through this activity, UDINUS conducted an architecture-focused examination of ConvTranspose2d for real-time object detection in visually complex road monitoring scenarios. The findings provide insights into architecture-aware enhancements and the potential of ConvTranspose2d for safety-critical object detection tasks.