针对目前石化危险品装车过程中海量监控视频图像人为处理效率低下、模糊图像识别率低等问题, 提出一种基于生成式对抗网络(GAN)和卷积神经网络(CNN)与极限学习机(ELM)相结合的监控模糊图像智能修复及检测方法。首先, 使用深度学习网络作为目标检测框架, 利用GAN网络中生成器与判别器间的零和博弈对模糊图像进行复原, 得到清晰完整的作业图像; 其次, 利用CNN自适应学习图像特征的能力, 对修复后的图像进行自主特征提取; 最后, 将提取的图像特征输入ELM分类器中进行目标识别与分类, 判断作业过程是否存在违规行为。试验结果表明: 所提方法图像修复速度快, 视觉效果自然, 且目标识别准确率高, 具有很好的泛化能力。

In order to obtain excellent slow-light performance, we propose a photonic crystal waveguide (PCW) that introduces extrinsic defect rods in the center row of a complete square lattice rotated 45° counterclockwise and the second row adjacent to it. The continuous cavities are used as a storage of electromagnetic energy and a speed reducer of light speed, used for slow optical transmission in PCWs. Then, the plane wave expansion method (PWE) is used to study the slow light transmission characteristics of the proposed structure, and the influence of the structure parameters on the slow light performance is analyzed. Finally, the bandwidth is obtained at 23.37 nm when the normalized delay bandwidth product (NDBP) reaches 0.40. In addition, considering the effect of material properties on slow light performance, NDBP is further optimized to 0.44, and the bandwidth reaches 27.63 nm. A simple but universal structure is designed to provide an important theoretical basis for further improving the storage capacity with high bandwidth and high NDBP slow light.