攜帶式小型害蟲檢測裝置設計與影像辨識模型優化

Abstract

粉蝨、薊馬與蕈蚊等小型害蟲不僅是多種植物病害的傳播媒介，亦對農作物生產造成重大威脅。由於這些害蟲體積微小、外觀相似使得自動化檢測極具挑戰。本研究針對溫室黏蟲紙害蟲檢測需求，開發一套可攜式害蟲檢測設備，能即時進行小型昆蟲檢測與計數。系統整合了一個高解析度 64 MP 影像模組、 LED 側光照明組件，以及Raspberry pi嵌入式邊緣運算平台，無需依賴遠端伺服器即可完成完整檢測流程。系統亦配置直觀的觸控式操作介面及雲端連線功能，以支援數據同步及長期害蟲族群分析。為提升小型害蟲檢測的精準度，本研究提出改進的目標檢測框架 CAFE-YOLO（Context-Aware Feature Enhanced YOLO），引入 Edge Boost Stem、Adaptive Downsampling 與 Context-Aware Feature Pyramid Network 三個架構模組，強化小目標特徵提取能力。CAFE-YOLO 在測試中達到 0.946 的 mAP 與 0.887 的 F1-score，並透過結構化剪枝與知識蒸餾進行模型輕量化，兼顧嵌入式裝置運算效率與檢測準確率。系統同時採用密度感知閾值調整策略（Density-Aware Thresholding, DAT），能依據局部害蟲密度動態調整檢測閾值，以提升在高密度及雜訊環境下的檢測穩健性。於四個場域進行的實地實驗顯示，本系統達成 8.08 的平均絕對誤差（MAE）、10.52 的均方根誤差（RMSE） 與 5.19 的平均絕對百分比誤差（MAPE），相較於傳統檢測框架在小型及密集害蟲檢測上有顯著提升，同時在嵌入式裝置上可於 48.3 秒內完成單張影像推論。

Whiteflies, thrips, and fungus gnats are major agricultural pests that serve as vectors for plant diseases and pose significant threats to crop productivity. Their small size, high visual similarity, and the visually complex environment of greenhouses make automated detection particularly challenging. This study presents the development of a portable, intelligent pest detection system specifically designed for real-time monitoring of small insects on sticky traps. The system integrates a high-resolution 64 MP imaging module, a custom-engineered LED side-lighting assembly to minimize glare, and an embedded edge-computing platform that enables on-device, real-time inference without reliance on remote servers. A user-friendly touchscreen interface and seamless cloud connectivity allow for centralized data synchronization and long-term pest population analysis. To achieve high detection accuracy for small pests, this research proposes a novel object detection framework, CAFE-YOLO (Context-Aware Feature Enhanced YOLO), which introduces three architectural modules: Edge Boost Stem, Adaptive Downsampling, and Context-Aware Feature Pyramid Network. CAFE-YOLO achieves a mean average precision of 0.946 and an F1-score of 0.887. The model is further optimized through pruning and knowledge distillation to meet embedded hardware constraints while maintaining detection performance. The system also incorporates a density-aware thresholding strategy, which dynamically adjusts detection thresholds based on local pest density to improve robustness in cluttered scenes. Field experiments conducted across four greenhouse sites demonstrated that the system achieves a mean absolute error of 8.08, a root mean square error of 10.52, and a mean absolute percentage error of 5.19. Compared to conventional detection frameworks, CAFE-YOLO delivered substantially improved accuracy, especially for small and densely clustered pests, while maintaining efficient on-device processing with an inference time as low as 48.3 seconds per image. This research demonstrates an integrated, lightweight, and high-precision pest detection solution. The device enhances the accuracy and efficiency of greenhouse pest monitoring, reduces manual labor, and provides timely, data-driven decision support for integrated pest management in smart agriculture.