A Novel Method Based-on Gradient and Deep Neural Network Filters to Generate Texture Images
Subject Areas : electrical and computer engineering
1 - Dept. of Comp. Eng., Faculty of Eng., Arak University, Arak, Iran
Keywords: Image augmentation, data generation, gradient ascent, convolutional neural network. ,
Abstract :
Production of image databases is one of the necessities of machine vision. There are various methods such as rotating, changing the viewing angle, resizing, etc., to increase the image data. The disadvantage of these methods is that the generated images are very similar to the original images and it is not enough to prevent overfitting. Among all types of images, texture images have more challenges. In this research, a new texture is generated using the convolution coefficients of pre-trained deep networks. In this method, new textured images are artificially produced by applying an ascending gradient to the images resulting from convolution filters. The difference between this method and the generative methods is that there is no initial texture image to increase, but here a new class of texture image is generated from the coefficients of the pre-trained deep network. After the new texture is produced, its number is increased by image processing methods. This method is between 3 and 5 times faster than some well-known generator networks. The quality of the images is much better. With this method, a texture database example has been produced, which includes 2400 images in 80 classes, and has been uploaded to the Kaggle site.