Image Information Extraction and Optimization Algorithm for Visual Communication Integrating Object Detection and Deep Learning

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 an innovative solution that integrates object detection and deep learning. This scheme constructs a two-stage processing method, uses the Darknet-19 backbone network to achieve multi-scale feature fusion, and combines depth separable convolution and channel rearrangement mechanisms to achieve network lightweighting, significantly improving computational efficiency while ensuring accuracy. The experiment showed that this method achieved 97% pixel level accuracy and 5 mm positioning error in water level detection tasks, which was 12% more accurate than the suboptimal model. The character recognition rate for encoding recognition tasks reached 97%, with a false detection rate of only 5%. In multi-tasking scenarios, the interference level was reduced to 0.12, the video memory usage increased by 62%, and the edge device latency remained stable at 25 ms. In terms of lighting robustness, the multi-scale photometric transformation achieved a lighting invariance index of 0.93 and improved the signal-to-noise ratio by 8.7 dB. In lightweight deployment verification, the computational density was 1.26 GMACs/mm², and the 72-hour failure rate was less than 0.1%. This study provides a solution for industrial vision systems that balances accuracy and efficiency, and has important application value in fields such as smart security and intelligent manufacturing. Future work will further optimize deployment adaptability through adaptive calibration techniques and dynamic pruning strategies.