Enhance Low Light Images using Keras, Python and Weights & Biases

🚀Hey everyone! In this video we’ll learn about enhancing low light images using a technique called Zero-DCE, which stands for Zero-Reference Deep Curve Estimation. And we will be doing so using Python, Keras and Weights & Biases. The interesting thing about the Zero-DCE approach is that it doesn't require labelled images so we can just train the model in an unsupervised learning way using ONLY the dark, low light images and still get stunning results. And it usually works a lot better than simply adjusting brightness or exposure. Also, the resulting model is really, really small - about 350 KB (yes, that small!) - and will be able to run really fast too, allowing for real-time enhancement of images. We can use it then do things like perform low light object detection or low light semantic segmentation. We also have a great conversation with Soumik Rakshit: the person who built the implementation of Zero-DCE in Keras. We dig into that implementation and pick his brain on its weaknesses and strengths. And, of course, when we get to training the model, we'll learn about using Weights & Biases and Keras together to track our machine learning experiments, as well as using W&B Tables to explore model predictions. --- Links 📍 Google Colab to train Zero-DCE in Keras to enhance low light images: http://wandb.me/zero-dce-colab 📍Zero-DCE paper: https://arxiv.org/abs/2001.06826 📍Soumik's Zero-DCE Kaggle kernel: https://www.kaggle.com/code/soumikrakshit/tensorflow-zero-reference-deep-curve-estimation 📍Keras.io Zero-DCE page: https://keras.io/examples/vision/zero_dce/ Follow Soumik: 👉 Twitter: https://twitter.com/soumikRakshit96 👉 GitHub: https://github.com/soumik12345/ Follow Ivan: 👉 Twitter: https://twitter.com/Ivangrov 👉 YouTube: https://www.youtube.com/c/IvanGoncharovAI --- ⏳ Timestamps ⏳ 00:00 Intro 00:23 What's gonna be in the video 02:16 Colab notebook and exploring data with W&B Tables 05:36 Hitting up Soumik 06:32 Use-cases for Zero-DCE 09:24 Looking through example model predictions 11:15 Why it works well on some images, but struggles on others? 14:20 A challenge for our amazing W&B community 16:42 How Zero-DCE works explained by Soumik 22:56 Why Soumik really enjoys Keras and TensorFlow 24:35 Keras and TensorFlow work really well with Weights & Biases 25:50 Thanks for chatting and explaining stuff Soumik 26:35 Training Zero-DCE 29:43 Monitoring training 32:34 Visualizing model predictions on the test dataset 35:26 Outro --- Get started with W&B: http://wandb.me/intro Follow us: Twitter: http://twitter.com/weights_biases Linkedin: https://www.linkedin.com/company/weights-biases Thanks for watching! If you have any questions, please don't hesitate to ask! If you have any suggestions, please don't hesitate either! We love hearing from the community and look forward to seeing you join our community!