| Abstract [eng] |
It is possible to generate images from text using GAN, but how to decide whether the GAN-generated image is realistic. Usually the results are compared visually, but if there are hundreds or thousands of such results and they all need to be evaluated. The problem here is how to automatically assess whether the generated image is realistic. To avoid having to evaluate the generated images visually, as different people may evaluate the same image differently and the results of visual evaluation at different times of the day may depend on the emotional state or attitude, e.g, If the mood is better, all the generated images may appear to be a much better representation of the objects than they were yesterday, or if two people are evaluating the same generated image at the same time, one person may find it a realistic/accurate representation of the object, while the other person may find that the generated image lacks details of the object, which may lead to a description of the image as not accurate. Also, visual assessment can be a time and resource consuming process if a large number of generated images are assessed. For this reason, this thesis will attempt to develop an additional process that can view all available images according to some given criteria and automatically evaluate each of them according to the same criteria, thus automating the evaluation of the quality of the generated images without human intervention. |
