<p>Tobacco quality inspection plays a vital role in ensuring standardized processing, reducing economic losses, and improving industrial automation. However, traditional inspection methods often suffer from inefficiency, high labor costs, and limited real-time capabilities. To address these challenges, this paper proposes an Internet of Things (IoT)-based data acquisition and edge computing framework enhanced with deep learning models for tobacco quality inspection. The system integrates heterogeneous sensing devices for multi-source data collection, including moisture, color, and texture features, while leveraging edge computing nodes to conduct real-time preprocessing, feature extraction, and anomaly detection. Specifically, a convolutional neural network (CNN) is employed to extract spatial texture and color features, while a long short-term memory (LSTM) network captures temporal dependencies in moisture and process variations. A lightweight data transmission protocol and optimized scheduling algorithm are designed to balance computational efficiency and energy consumption. Experimental results demonstrate that the proposed hybrid edge model achieves an accuracy of 96.3% in tobacco quality classification, while reducing average processing latency by 38% compared with cloud-only architectures. In addition, complexity profiling shows that the deployed model requires 12.8M parameters, 3.42 GFLOPs, 486 MB peak runtime memory, and 25.0 ms end-to-end latency at <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(512\times 512\)</EquationSource> </InlineEquation> input resolution on a Jetson Xavier NX platform. The study provides a practical solution for real-time and scalable tobacco quality monitoring, offering theoretical insights and engineering value for smart agriculture and industrial IoT applications.</p>