China
Africa
Explore chapters and articles related to this topic
Generative Adversarial Network
Published in S. Kanimozhi Suguna, M. Dhivya, Sara Paiva, Artificial Intelligence (AI), 2021
Image translation is the process of converting an image into another modified image or other forms of data like text, audio, etc. Image-to-image translation leads to many fascinating applications. Image translation includes the conversion of objects from one domain to objects of another domain. For example, images of a horse can be converted into images of zebra and vice versa. Many neural networks like CycleGAN [14], Pix2Pix [12], and BigGAN [18] have demonstrated this using various data sets. The conversion between day and night images, seasons, colourization of black and white images, and many more have been demonstrated using Pix2Pix. Even altering human features like hair colour and facial expression has been translated using StyleGAN [17] and DiscoGAN [13]. An example of converting label image from CMP facades data set [19] to real-world buildings using Pix2Pix [12] is provided in Figure 7.4. Even the conversion of faces into cartoon or anime characters has been demonstrated. Image super-resolution is another application where GAN shines.
Recognition of hidden distress in asphalt pavement based on convolutional neural network
Published in International Journal of Pavement Engineering, 2022
Wenchao Liu, Rong Luo, Yu Chen, Xiaohe Yu
As a branch of ML, DL (Hsieh and Tsai 2020) has flourished in many fields, such as natural language processing (NLP) and computer vision (CV), in recent years (LeCun et al.2015). With the development of DL, the research on applying the DL algorithm to GPR data interpretation has begun, and good results have been achieved. Research on the intelligent interpretation of GPR data by combining the DL algorithm in the CV field with the image data of GPR B-scan is the most popular approach. Although the interpretation of 3D data of GPR C-scan has also developed, most studies still convert C-scan data into B-scan data from multiple angles before the interpretation. The different DL algorithms adopted by these studies according to the purpose of the task can be mainly divided into the convolutional neural network (CNN) for image classification, object detection algorithms for locating target position, and image segmentation algorithms for the classification of pixels in the image. Among them, the CNN, representative of DL in the field of CV, performs well in image recognition tasks (Hussain et al.2018). CNN is an artificial neural network model inspired by the neural mechanism of the visual system. Compared to traditional digital image processing methods, CNN can automatically extract image features and is invariant to image translation, flipping, and other transformations (Singla et al.2021).