Multi-Task Visual Information Extraction in Industrial Environments Using Darknet-19 with Depthwise Separable Convolutions

Abstract

In response to the dual challenges of insufficient generalization of traditional image processing methods and high computational complexity of deep learning models in industrial visual scenes, this study proposes a two-stage solution integrating object detection and a deep learning scheme. The scheme employs a modified Darknet-19 backbone with depthwise separable convolutions and channel rearrangement mechanisms for multi-scale feature fusion, significantly improving computational efficiency while maintaining accuracy. Experiments on a dataset of 4,000 industrial water level images and 10,000 encoding samples showed that the research method achieved 97% pixel-level accuracy and 5 mm positioning error in water level detection, outperforming suboptimal models by 12%. For encoding recognition, it reached a 97% character recognition rate with only 5% false detection rate. In multi-task scenarios, system interference was reduced to 0.12, with 62% increased video memory usage and stable 25 ms edge latency. The multi-scale photometric transformation achieved a lighting invariance index of 0.93 and improved SNR by 8.7 dB. Lightweight deployment yielded a computational density of 1.26 GMACs/mm² and a 72-hour failure rate below 0.1%. This work provides an accuracy-efficiency balanced solution for industrial vision systems, with applications in smart security and intelligent manufacturing. Future work will focus on adaptive calibration and dynamic pruning for enhanced deployment adaptability.