Efficient YOLOv11-Based Oil Palm Fresh Fruit Bunch Detection Using Structured Detection-Path Pruning and INT8 Quantization

Abstract

This study proposes an efficient you only look once (YOLO)v11s-based framework for real-time oil palm fresh fruit bunches (FFB) detection, addressing the high computational cost that typically limits YOLO-based detectors on edge devices. The framework integrates structured detection-path pruning (SDPP) and 8-bit Integer (INT8) post-training quantization. SDPP analyzes the object-scale distribution of FFBs in UAV imagery to identify the detection paths that contribute most effectively to the target object characteristics. Based on this analysis, redundant detection branches with limited contribution to detection performance are removed, reducing the computational burden while preserving accuracy. The pruned model is subsequently quantized to INT8 precision to further improve inference efficiency on resource-constrained devices. Experimental results show that SDPP reduces computational complexity by 53.5% while maintaining an mAP@0.5 of 0.987. With INT8 quantization, the optimized model achieves 30.42 FPS on the NVIDIA Jetson Orin Nano, demonstrating its potential for real-time FFB detection on edge devices.

Access to Document

Other files and links

• Link to publication in Scopus
• Open Access Version Available

Fingerprint