Deep neural networks provide accurate results for most of the applications. However, they need a big dataset to train properly. Providing a big dataset is an important challenge in most applications. Image augmentation refers to some of the techniques that increase the number of image data. Common operations for image augmentation are related to changes in illumination, rotation, contrast, size, viewing angle, etc. Recently Generative Adversarial Networks (GAN) have been employed for image generation. However the same as image augmentation methods, the GAN approaches can only generate images that are similar to the original images. Therefore they also cannot generate new classes of the data. Texture images have more challenges than general images and the generation of textures is more complicated than other types of images. In this paper, a gradient-based deep neural network method is proposed, that generates a new class of the texture. By using different kernels of pre-trained deep networks, it is possible to generate new classes of textures rapidly. After generating new textures for each class, the number of textures is increased by image augmentation. During this process, some techniques are proposed to remove incomplete and similar created textures automatically. The proposed method is faster than some well-known generative networks by around 4 to 10 times. Furthermore, the quality of generated textures is better than those networks. In terms of some image quality metrics the proposed method can generate better textures than some of the GANs and parametric models. The proposed method can provide a big texture dataset to train deep networks. By using the proposed method, a new big texture dataset has been made artificially. This dataset is around 2GB and includes 30000 textures with a size of 150×150 in 600 classes. It has been uploaded to the Kaggle site and Google Drive. This dataset is called BigTex. Compared to other texture datasets the proposed dataset is the greatest one and can be used as a big texture dataset to train more powerful deep neural networks and avoid overfitting
