针对雾霾天气下视频监控图像出现的细节缺失、色彩暗淡和亮度降低等问题,目前现有的图像去雾算法在视频监控场景中往往难以同时满足去雾效果和实时处理的要求.为了恢复出质量更高的无雾图像,文章在传统AOD-Net算法中引入Squeeze and Excitation机制,以自适应的方式分配通道权重,同时引入金字塔池化模块,扩大网络感受野,最终采用复合损失函数,以均衡考虑图像的边缘特征及纹理细节.同时,此系统以Zynq作为实现平台,使用Vivado HLS进行接口为AXI4-Stream的新型AOD-Net算法IP核的开发,使用PL端作为算法的实现单元,PS端作为控制核心,充分发挥异构SoC的架构优势.实验结果表明:基于Zynq平台下的新型AOD-Net算法,图像去雾效果显著,信噪比极值优化了2.45 dB,结构匹配度提升至91.2%,降低了雾霾天气对视频监控图像的影响.
Design of Image Defogging System Based on Zynq New AOD-Net Algorithm
In view of the problems of detail loss,color dimming and brightness reduction in video surveillance images in haze weather,the existing image de-fogging algorithms are often difficult to meet the requirements of de-fogging effect and realtime pro-cessing in video surveillance scenes.In order to restore the fog-free image with higher quality,the Squeeze and Excitation mecha-nism is introduced into the traditional AOD-Net algorithm,and the channel weights are allocated in an adaptive way.At the same time,the pyramid pool module is introduced to expand the network sensitivity field.Finally,the composite loss function is adopted to equalize the image edge features and texture details.At the same time,the system uses Zynq as the implementation platform and Vivado HLS to develop the IP core of the new AOD-Net algorithm whose interface is AXI4-Stream.The PL end is used as the imple-mentation unit of the algorithm,and the PS end is used as the control core,giving full play to the advantages of the heterogeneous SoC architecture.The experimental results show that the new AOD-Net algorithm based on Zynq platform has significant image de-fogging effect.The SNR extreme value is optimized by 2.45 dB,and the structural matching degree is increased to 91.2%,which reduces the influence of haze weather on video surveillance images.
new AOD-Net algorithmimage defoggingZynqVivado HLSpyramidization
国家科技期刊平台
NSTL
万方数据
