基于生成对抗网络的动态插图图像细节增强方法

Dynamic illustration image detail enhancement method based on generative adversarial network

通过对动态插图图像细节增强处理,提高插图识别检测能力,提出基于生成对抗网络的动态插图图像细节增强方法。构建动态插图图像的多视图立体成像模型,采用生成对抗网络实现对插图图像细节特征提取的自适应训练,采用伪三维卷积的轻量级网络模型,提取动态插图图像细节特征值,采用深度图作为中间变量连接动态插图图像细节分量,通过区块模糊匹配实现对图像增强过程中的套索分割,通过密集点云重建,根据对动态插图图像深度图的三维重建结果,采用生成对抗网络实现对图像细节补偿和增强处理;在D-MVSNet和3D视图软件中实现图像细节增强的交互设计。测试结果表明,该方法进行动态插图图像细节增强处理,能较为完整地重建动态插图的细节点云,提高动态插图的辨识度,其重建精度可达到0.98,人机交互性较好。

This paper proposes a dynamic illustration image detail enhancement method based on generative adversarial network to improve image recognition and detection ability.The multi-view stereo imaging model of dynamic illustration images was constructed,and the adaptive training of detail feature extraction was realized by using the generative adjunct network.The lightweight pseudo-three-dimensional convolution network model was used to extract detail feature values of dynamic illustration images,and the depth map was used as the intermediate variable to connect the detail components of dynamic illustration images.The lasso segmentation in the process of image enhancement was realized by block fuzzy matching,and the image detail compensation and enhancement were realized by generative adversarial network based on the 3D reconstruction results of dynamic illustration image depth map through dense point cloud reconstruction.The interactive design of image detail enhancement is realized in D-MVSNet and 3D view software.The test results show that this method can reconstruct the detail point cloud of dynamic illustration completely and improve the recognition of dynamic illustration.The reconstruction accuracy can reach 0.98,and the human-machine interaction is good.