EDGE-AWARE VIDEO IMAGING SYSTEM FOR REAL-TIME TOMATO DISEASE MONITORING: A YOLOV9 PIPELINE WITH SENSOR CALIBRATION, RATE-ADAPTIVE COMPRESSION, AND ON-DEVICE INFERENCE

Venkatesh Rachakonda^1*Malarvizhi N²

• ¹Malla Reddy Engineering College, Secunderabad, India
• ²Vel Tech Rangarajan Dr Sagunthala R&D Institute of Science and Technology, Chennai, India

Word count: 204 This paper presents an end-to-end imaging system for field-scale, real-time plant-health monitoring from video. The system integrates (i) a camera module with radiometric/colour calibration and temporal exposure normalization, (ii) a rate-adaptive compressor that maintains structural similarity under bandwidth constraints, and (iii) a detection stage using a lightweight, quantization-aware YOLOv9 model. A temporal consensus module fuses short video windows to suppress frame-level false positives, while an uncertainty-aware quality gate flags low-confidence segments for re-capture. We report system-level metrics—throughput (fps), energy per processed frame on Jetson-class edge devices, SSIM/PSNR preservation after compression, and end-to-end latency—alongside vision metrics (precision, recall, F1, mAP@0.5:0.95). On a 3,750-frame, five-class tomato dataset captured in variable outdoor conditions, the pipeline sustains 30 fps with sub-65 ms median latency, preserves SSIM ≥0.96 at 40–60% bitrate reduction, and achieves mAP@0.5 of 0.95 with an F1 of 0.92. Ablations show that calibration reduces domain shift by 18% (ΔmAP@0.5:0.95), while temporal consensus lowers false positives by 34% without sacrificing recall. The system design—sensor calibration, compression control, and deployable inference—demonstrates a reproducible imaging workflow for agricultural video, and the components generalize to other outdoor monitoring tasks where robust, low-latency vision at the edge is required. These contributions fit IMA's emphasis on imaging systems, algorithms, and practical deployment.