From Farm to Frame: Accelerating Agricultural Vision Systems Through Preprocessing Pipeline Optimizations

Abstract

Precise, real-time plant disease detection is vital for sustainable agriculture, yet preprocessing mismatches between training and inference can severely undermine model performance. In this work, We evaluate two inference pipelines for YOLOv8m on a 14 150-image potato late-blight dataset (1 062 images held out for testing): (1) the default Ultralytics workflow (letterbox resize + per-channel normalization) and (2) a lightweight “Custom” pipeline (416×416 area resize, BGR→RGB, scale to [0,1]). Inference-only deployment of the custom pipeline more than doubles throughput (54→122 FPS) and halves latency (18.6→8.2 ms, std = 0.9 ms), but drops mAP@0.5 from 99.4 % to 76.2 %. Retraining YOLOv8m under the same custom preprocessing fully restores accuracy (mAP@0.5 = 99.4 %, mAP@[0.5:0.95] = 94.5 %) while preserving high throughput (120 FPS) and low latency (< 9 ms). We analyze resource utilization and latency stability across both pipelines. System profiling reveals modest resource use (<7 % CPU, ! 80 % GPU, 1.3 GB VRAM). These findings demonstrate that end-to-end consistency in preprocessing is essential to achieve both state-of-the-art accuracy and real-time performance in agricultural object detection.