神经网络方法在变电站消防管理的应用研究

为了提升变电站消防管理科学化水平,实现视频图片火焰特征提取及识别定位。接入视频监控码流数据,构建基于卷积神经网络的火焰识别模型,进行实时识别火焰特征并预警预报。实验表明,该方法能够自动提取火焰特征,有效提高复杂背景下的火焰识别的准确率,具有良好的鲁棒性和泛化能力,在变电站消防管理中有较大应用前景。     

A novel chromatic dispersion monitoring method for 1 024 QAM fiber-optic system based on asynchronous amplitude sampling

A new chromatic dispersion (CD) monitoring method is proposed, and the CD monitoring is achieved on the whole range of 0—600 ps/nm. A dual-polarization (DP)-1 024 quadrature amplitude modulation (QAM) 400 Gbit/s CD monitoring system is built using OptiSystem14.0 software. With different optical signal to noise ratio (OSNR) values, different filter bandwidths or different duty cycles, the simulations are carried out. The simulation results show that this new CD monitoring method can be less affected by noise and has a high tolerance for different filter bandwidths. At the same time, this method can do effective CD monitoring on the signal with duty cycle greater than 80%. A high sensitivity in 0—200 ps/nm can be achieved using this CD monitoring method. The technique supports a wide range of data traffic and enhances operation flexibility of optical networks. This work has been supported by the National Natural Science Foundation of China (Nos.61274121 and 61574080). E-mail:zhub@njupt.edu.cn      

生成对抗网络磨玻璃样肺结节分割方法

早期肺癌在影像学的表现形式之一是肺结节，其中磨玻璃样肺结节(Ground-Glass Nodules, GGNs)被认为是恶变可能性最大的肺结节之一。在诊断性CT图像上GGNs存在边缘模糊、大小形状各异、不规则等影响分割精度的问题。针对GGNs形态复杂、分割精确度欠佳的问题，提出了一种生成对抗式网络(Generative Adversarial Network, GAN)——GAN-DeepLabv3+,将DeepLabv3+作为图像分割生成器网络，通过引入编解码器结构，将深度可分离卷积应用到空洞空间卷积池化金字塔和解码器模块上，通过对抗式训练，最终获得GGNs病灶分割模型。实验结果证明，在对GGNs患者的CT医学图像的分割中，所提方法Dice系数为0.952,交并比(Intersection over Union, IoU)为0.876,像素精确度(Pixel Accuracy, PA)为0.991,相较于原始DeepLabv3+和ACRU-Net等现有方法，对GGNs均有一定的提升。