ISSN 2079-3537      

 
 
 
                                                                                                                                                                                                                                                                                                                                                                                                                                                                             
Scientific Visualization
Issue Year: 2016
Quarter: 4
Volume: 8
Number: 4
Pages: 67 - 79
Article Name: GENERATION OF SYNTHETIC INFRARED IMAGES AND THEIR VISUAL QUALITY ESTIMATION USING DEEP CONVOLUTIONAL NEURAL NETWORKS
Authors: V.V. Kniaz (Russian Federation), V.S. Gorbatsevich (Russian Federation), V.A. Mizginov (Russian Federation)
The paper is recommended by program committee of 26th International Conference on Computer Graphics and Vision GraphiCon’2016.
Address: V.V. Kniaz
vl.kniaz@gosniias.ru
Federal State Unitary Enterprise «State Research Institute of Aviation Systems» GosNIIAS, Moscow, Russiam Federation.

V.S. Gorbatsevich
gvs@gosniias.ru
Federal State Unitary Enterprise «State Research Institute of Aviation Systems» GosNIIAS, Moscow, Russiam Federation.

V.A. Mizginov
kevin5garnett-kg@yandex.ru
Federal State Unitary Enterprise «State Research Institute of Aviation Systems» GosNIIAS, Moscow, Russiam Federation.
Abstract: In recent years deep convolutional neural networks became a powerful tool for object recognition. Accuracy of deep convolutional neural networks is comparable with the accuracy of a human operator. Convolutional networks also provide a flexible framework for hyperspectral image processing. The simultaneous usage of images that were captured using sensors of different wavelengths increases the probability of the correct classification.
To achieve a high accuracy of classification during the learning process it is required to have a large number of images in the training dataset. Training images should be captured under various conditions. A number of large image datasets (Pascal VOC, MS COCO) for training of image classification algorithms are available in the open access nowadays. However, training of image classification algorithm for classification of infrared images is more complex as large datasets of infrared images are not available in the open access.
This paper is focused on the development of a method for augmentation of images captured in the visible wavelength with synthetic infrared images. The method is based on the rendering of textured 3D-models of objects of interest and generation of pseudo-infrared background textures from a source image captured in the visible wavelength. An algorithm for registration of real images and 3D-models is presented. A dedicated algorithm for object tracking in complex environment is discussed. The proposed method was used to generate a large learning dataset for detection of foreign objects on a runway. To evaluate the quality of synthetic images a special quality measure is proposed. The measure is based on the accuracy of a deep neural network. The paper is concluded with the discussion of the accuracy of deep convolutional neural network classifier trained on the dataset that were generated using the proposed method.
Language: Russian


Open Article   Download ZIP archive