Improving the Accuracy of Deep Learning-Based Object Detectors on Raspberry PI

Abstract

This study focuses on the implementation and performance evaluation of the YOLOv8 object detection algorithm on a resource-limited device, Raspberry Pi 5, to detect abnormalities in melon leaves. A secondary dataset containing images of melon leaves captured using a Raspberry Pi camera was utilized in two forms: original and augmented. The model was trained and evaluated using key performance metrics such as mean Average Precision (mAP), precision, recall, and F1 score. Two backbone architectures, Original and GhostNet, were compared to assess both detection accuracy and computational efficiency. The results revealed that data augmentation significantly improved detection performance, with mAP@0.5 for the Original Backbone increasing from 0.52 to 0.80, and for GhostNet from 0.49 to 0.78. After hyperparameter tuning, the best results achieved were mAP\@0.5 of 0.81 for the Original Backbone and 0.80 for GhostNet. Although the Original Backbone attained slightly higher accuracy, the GhostNet model demonstrated superior computational efficiency with faster inference speed (3.97 FPS vs. 2.75 FPS), lower latency, and smaller model size (9.23 MB vs. 20 MB). Based on these findings, the GhostNet backbone with augmented data and optimized hyperparameters provides the best trade-off between accuracy and efficiency, making it the most suitable option for real-time deployment on lowpowered devices such as Raspberry Pi 5, and offering a promising solution for fieldbased monitoring of melon leaf abnormalities in resource-constrained agricultural environments. Keywords: Deep learning, YOLOv8, GhostNet. Melon leaf, Raspberry Pi.