UROP Proceedings 2022-23

School of Engineering Department of Computer Science and Engineering 99 Generative AI Supervisor: CHEN, Qifeng / CSE Student: ZHANG, Muzi / QFIN Course: UROP1100, Summer DragGAN is a new kind of GAN that allows user to modify multiple features such as shape, body pose, emotion of the images in a generative approach, users can achieve this by adding one more multiple pairs of points and DragGAN will perform a series of iterations that “drag” the handle points to the corresponding target points without modifying pixels outside a region specified by the user. With DragGAN, the position, shape, expression, and body pose of a person or animal in a casual photograph may be changed by social media users; professional movie previsualization and media editing may call for quickly sketching out scenes with specific layouts; and car designers may wish to interactively change the shape of their creations. Also, in order to satisfy these applications, DragGAN can do image process with these advantages: • 1) Flexibility: it can control different spatial attributes including position, pose, shape, expression, and layout of the generated objects or animals • 2) Precision: it can control the spatial attributes with high precision • 3) Generality: it is applicable to different object categories but not limited to acertain category. While previous works only satisfy one or two of these properties, DragGAN target to achieve them all. Deep Learning for Medical Image Analysis Supervisor: CHEN, Hao / CSE Student: DENG, Yufan / SENG Course: UROP1100, Fall UROP2100, Spring Jointly grading Diabetic Retinopathy (DR) and Diabetic Macular Edema (DME) provides more information from the cross-task feature and can improve performance, but the features of the two tasks have discrepancies, and direct usage of the cross-task feature is not appropriate. In this project, we proposed three normalization methods—lastnorm, allnorm, adanorm—to align the cross-task feature to improve the performance of the dual-stream disentangled learning architecture (DETACH) network. Experiment results show that the proposed normalization methods generally both perform better than the model without added normalization, and we also gave a partial explanation by doing salience analysis on classifiers.