面向嵌入式系统的图像实时去雾方法

Real-time defogging method for embedded systems

近些年来,深度学习方法快速兴起,卷积神经网络技术被应用于图像去雾领域,去雾效果得到了进一步提升。但也面临着计算量大,无法在嵌入式设备中达到实时性的问题。为增强嵌入式设备的去雾效率,通过优化网络结构和剪枝算法压缩网络模型,设计出一种面向嵌入式系统的图像实时去雾网络,最后将模型部署在嵌入式系统中。实验结果表明,该方法在不降低去雾效果的前提下,模型大小压缩83%,处理速度提高70%,达到40帧/s,在嵌入式系统中实现了实时去雾。与目标检测网络YOLO-FASTEST联调,在雾天情况下的平均检测精度提高近3%,检测速度达到近5帧/s。

In recent years,deep learning methods have emerged rapidly,and convolutional neural network technology has been applied to the field of image defogging,achieving good defogging effects.However,it is also faced with the problem that it cannot achieve real-time performance in embedded devices due to the large amount of computation.To enhance the defogging efficiency of embedded system,an image fast defogging network for embedded system is designed.The network model is compressed by optimizing the network structure and pruning algorithm.Finally,the model is deployed in embedded system.The experimental results show that the model size is reduced by 83%and the processing speed is increased by 70%to 40 frames/s without reducing the dehazing effect.The method realizes realtime dehazing in embedded devices.Combined with the YOLO-FASTEST,a target detection network,the average detection accuracy improved by nearly 3 percent on foggy days.The detection speed reached nearly 5 frames/s.