| 平行视觉框架下深度卷积神经网络可视化分析 |
| |
| 引用本文： | 翟永杰,杨旭,王金娜,王坤峰,赵振兵. 平行视觉框架下深度卷积神经网络可视化分析[J]. 计算机工程与应用, 2020, 56(19): 139-145. DOI: 10.3778/j.issn.1002-8331.1907-0284 |
| |
| 作者姓名： | 翟永杰 杨旭 王金娜 王坤峰 赵振兵 |
| |
| 作者单位： | 1.华北电力大学 自动化系，河北 保定 0710032.中国科学院 自动化研究所 复杂系统管理与控制国家重点实验室，北京 1001903.华北电力大学 电子与通讯工程系，河北 保定 071003 |
| |
| 基金项目： | 国家重点实验室研究项目;国家自然科学基金;河北省自然科学基金 |
| |
| 摘 要： | 深度学习方法在计算机视觉领域取得了很大的发展，多种深度卷积神经网络在实际的目标检测中取得了很好的应用效果，但均存在网络可解释性较差的问题。通过将特征图反向映射到输入图像的像素空间，来对网络的特征图进行可视化分析；在平行视觉研究框架下，分别采用真实和人工绝缘子图像样本来分析网络的特征响应，最后依据可视化结果对网络参数进行调整。研究结果表明，人工图像中绝缘子的占比、角度和位置对网络的特征响应和分类正确率均有不同程度的影响，根据真实和人工绝缘子图像的特征图可视化结果来对网络的结构和参数进行调整，能够较好地提升网络的性能。
|
| 关 键 词： | 平行视觉 卷积神经网络 可视化 特征响应 |
Visual Analysis of Deep Convolutional Neural Networks in Parallel Vision Framework |
| |
| ZHAI Yongjie,YANG Xu,WANG Jinna,WANG Kunfeng,ZHAO Zhenbing. Visual Analysis of Deep Convolutional Neural Networks in Parallel Vision Framework[J]. Computer Engineering and Applications, 2020, 56(19): 139-145. DOI: 10.3778/j.issn.1002-8331.1907-0284 |
| |
| Authors: | ZHAI Yongjie YANG Xu WANG Jinna WANG Kunfeng ZHAO Zhenbing |
| |
| Affiliation: | 1.Department of Automation, North China Electric Power University, Baoding, Hebei 071003, China2.The State Key Laboratory of Management and Control for Complex Systems, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China3.Department of Electronic and Communications Engineering, North China Electric Power University, Baoding, Hebei 071003, China |
| |
| Abstract: | Deep learning methods have made great progress in the field of computer vision, various deep convolutional neural networks have achieved good application effects in actual target detection, but the interpretability of the network is poor. The feature map is reversely mapped to the pixel space of the input image and the feature map of the network is visualized for analysis; under the framework of parallel vision research, the feature response of the network is analyzed with real and artificial insulator image samples, the network parameters are adjusted according to the visualization results. The results show that the artificial samples with different proportions, angles and positions have different effects on the accuracy of the network, the characteristic response of the network is also different. The structure and parameters of the network are adjusted according to the visualization results of the feature graph, which improves the performance of the network. |
| |
| Keywords: | parallel vision convolutional neural network visualization characteristic response |
| 本文献已被 万方数据 等数据库收录！ |
| 点击此处可从《计算机工程与应用》浏览原始摘要信息 |
|
点击此处可从《计算机工程与应用》下载免费的PDF全文 |
|
