To maintain image quality under changing weather conditions, complex deep networks have been widely used for image enhancement. Owing to the restricted computing power of Internet of Things enabled smart… Click to show full abstract
To maintain image quality under changing weather conditions, complex deep networks have been widely used for image enhancement. Owing to the restricted computing power of Internet of Things enabled smart cameras, existing lightweight solutions using deep neural networks are prone to degrading image quality. Thus, our study incorporates recursive and attention-based enhancement into a lightweight residual network that uses high-efficiency convolution operations as a basic building block. With this modular design on a recurrent neural network, the resulting neural network can be easily optimized according to the hardware specifications of a smart camera. In addition, to adapt to the various weather conditions, we design a sequential environment-aware model deployer that detects continuous weather changes so that the most appropriate image-enhanced model can be deployed. Comparison of experimental results for the proposed approach with our previous study (also the latest study) and other recent research indicate that the performance of image enhancement in rainy, hazy, and snowy scenes can be slightly improved, and the operation speed (seconds per frame) can be increased by 60%. The detection error rate for the model deployer dropped from 1% to 0.03% (improved by 97%). The image-enhanced results achieved a 4% improvement in accuracy, demonstrating the method's effectiveness and benefits for realizing edge intelligence.
               
Click one of the above tabs to view related content.