Real-time identification of agricultural pests and diseases on edge devices requires models with high accuracy, low latency, and stable deployment capabilities. In this paper, an edge computing recognition system based on improved lightweight neural network is constructed, which is organized around edge-aware input coding, multi-scene image lightweight representation, detection and classification collaborative network and end-to-end scheduling link to realize target localization and category recognition of crop diseases and pests. The experiment uses a self-built agricultural disease and insect pest image dataset, which contains 18,420 labeled images covering 12 types of disease targets and 8 types of insect targets. The data collection covers the orchard and vegetable field scenes, and the unified manual labeling and category verification post-processing are completed. The results show that the mAP@0.5 of the proposed system on Jetson Orin Nano reaches 95.8%, the classification Accuracy reaches 96.4%, and the average inference delay is 27.3 ms. The system maintains a stable response to occlusion, illumination change and small target samples in complex field scenes, and has the actual deployment ability for intelligent agricultural monitoring tasks.
