Design of an <scp>IIoT</scp> Intelligent Image Analysis System Integrating <scp>AI</scp> and Edge Computing

ABSTRACT This study proposes an IIoT intelligent image analysis system that integrates AI with edge computing to mitigate the bandwidth, latency, and privacy limitations of cloud‐centric industrial video processing. The proposed framework deploys edge gateways between front‐end cameras and the IoT platform to perform denoising, timestamp alignment, local inference, and event pushing, thereby enabling edge–cloud collaboration for visual monitoring and production‐data fusion. On the algorithmic side, YOLOv4‐tiny is adopted as the base detector and optimized through a structured pruning‐and‐fine‐tuning strategy in which channel importance is quantified by the magnitude of batch‐normalization scaling factors, and pruning is progressively applied to the backbone and FPN while preserving the detection heads. Compared with the baseline model, the final compressed detector reduces the parameter count by 26.4%, FLOPs by 24.7%, and storage size by 27.7%; it reaches 23.2 FPS with a model size of 16.2 MB, while the mAP is recovered to 71.89%. For industrial scene understanding, a two‐level retrieval mechanism combining product quantization and inverted indexing achieves 95.5% classification accuracy on 2000 samples, improving upon the 85% baseline. Experiments on anomaly‐detection benchmarks further show strong generalization, with I‐AUROC/I‐AP of 92.7/96.7 on MVtecAD. These results indicate that the proposed framework provides a practical basis for real‐time industrial visual inspection, anomaly warning, and data‐driven monitoring under edge deployment constraints.