Research Article
CAD-Pest: A CUDA-Accelerated Deep Learning Framework for Intelligent, Real-Time Pest Detection and Classification on the AgroPest-12 Dataset
@INPROCEEDINGS{10.4108/eai.19-12-2025.2362518, author={Mankirat Kaur and Abhinavpreet Singh Arora and Pratham Mittal}, title={CAD-Pest: A CUDA-Accelerated Deep Learning Framework for Intelligent, Real-Time Pest Detection and Classification on the AgroPest-12 Dataset}, proceedings={Proceedings of the 1st International Conference on Computational Techniques in Data Science, IICTDS 2025, 19-20 December 2025, Chandigarh, India}, publisher={EAI}, proceedings_a={IICTDS}, year={2026}, month={11}, keywords={Precision Agriculture Deep Learning CUDA TensorRT Edge AI YOLOv8 MobileNetV3 HUDA AgroPest-12 AgriPest Small Object Detection Model Quantization Real-Time Inference GPU Optimization Sustainable Agriculture}, doi={10.4108/eai.19-12-2025.2362518} }- Mankirat Kaur
Abhinavpreet Singh Arora
Pratham Mittal
Year: 2026
CAD-Pest: A CUDA-Accelerated Deep Learning Framework for Intelligent, Real-Time Pest Detection and Classification on the AgroPest-12 Dataset
IICTDS
EAI
DOI: 10.4108/eai.19-12-2025.2362518
Abstract
Agricultural pests continue to threaten global food security, damaging up to 40% of annual crop yields worldwide. The manual detection of pests is still applied, is susceptible to errors and cannot be applied on large scale, particularly in cases where the real-time monitoring is required. This paper presents CAD-Pest, a CUDA-implemented deep learning system that is capable of detecting and classifying pests and their segmentation in real-time. The architecture integrates YOLOv8 for object detection, MobileNetV3 for lightweight classification, and HUDA for segmentation within a unified GPU-optimized pipeline employing cuDNN, mixed-precision FP16, and TensorRT FP16/INT8 quantization. Experiments on the AgroPest-12 dataset demonstrate strong in-domain performance (mAP@0.5 = 1.0000, Macro-F1 = 0.8336, Dice = 0.6525). To evaluate robustness, a zero-shot cross-dataset validation on AgriPest was conducted. Using only the AgroPest-12–trained weights, CAD-Pest achieved mAP@0.5 ≈ 0.82, Macro-F1 ≈ 0.75, and Dice ≈ 0.58 on AgriPest, evidencing strong cross-domain generalization and reduced dataset bias. INT8 quantization exhibited hardware-dependent variation, degrading accuracy on Jetson Nano but improving performance-per-watt on Jetson Orin Nano. Comparative benchmarking across YOLOv8, Faster R-CNN, and EfficientDet further highlights CAD-Pest’s superior accuracy–latency trade-off for real-time pest monitoring.