Author: Zixuan Zhang, Chengxuan Cai
In Hollywood crimes movies we often see detectives solving crimes with the help from one of their computer geeks who can reveal hidden information from blurred, low-quality images. This project is an effort to achieve the same task, but on one specific type of image - license plates. License plate enhancement is a detailed application of a broader field called Single Image Super Resolution (SISR).
The project is inspired by several state-of-the-art SRSR models such as:
The dataset used in this project is called the Chinese City Parking Dataset, a large-scale collection of plate images in various conditions.
Read my post on Medium for further understanding
Preprocessing
Training & Evaluation
Before training the model it is important to preprocess the raw dataset using the preprocess.py script
Our plate enhancer model is trained in an adversarial fashion(GAN), meaning the generator is trained to create realistic reconstruction of images that can fool the discriminator, which is a binary classifier. Why GANs? Well, according to several papers, GAN network tend to create more realistic image reconstruction comparing to model solely trained in the supervised fashion. For instance, models that minimize Mean Square Error tend to have over-smoothing artifacts. Therefore, there are two models - the generator(reconstructor) and the discriminator(classifier).
The generator is trained to minimize a novel hybrid loss function, namely the perceptual loss defined in the SRGAN paper
I'd like to thank Olaoluwa Adigun for his amazing suggestions during the span of this project!
This project won the Best Deep Learning Design Award in USC EE599-Deep Learning. Here's the link
to our amazing rojects done by my classmates!
Also, this project stands on the shoulder of many other SISR projects:
此处可能存在不合适展示的内容,页面不予展示。您可通过相关编辑功能自查并修改。
如您确认内容无涉及 不当用语 / 纯广告导流 / 暴力 / 低俗色情 / 侵权 / 盗版 / 虚假 / 无价值内容或违法国家有关法律法规的内容,可点击提交进行申诉,我们将尽快为您处理。