We are living in an era where people believe in capturing every moment rather than living it. Here are some interesting photography facts - it is estimated that over 250 billion photographs have been uploaded on facebook since its launch. Every year around 300 billion photo graphs are clicked. People are fond of clicking photographs and sharing them on social media. Facebook and Instagram are the perfect examples. Needless to say everybody wants to click good pictures, but everybody is not a professional photographer with professional camera in hand. So how to do professional photography with simple digital camera (smart phone camera) and with limited photography skills. This is where the computational photography comes to the rescue.

Computational Photography (source: cs.brown.edu)

Computational photography is using digital computational techniques rather than optical processes to obtain quality pictures. Almost all digital cameras including the cell phone cameras employ some sort of computational photography by using various image processing algorithms. Then there are techniques available to give a retouch to the clicked image like improving brightness, tuning contrast etc.

Researchers at MIT (USA) had in the past developed a technique where a low resolution version of the clicked image could be sent to the cloud for processing, for image enhancement. The cloud sent back, what MIT called a 'transform recipe' that could be used to enhance or retouch the clicked image. The idea of processing the low resolution version of the image was to save bandwidth and other resources. This project got Google interested in it.

MIT and Google have now together developed a technique using machine learning where image can be enhanced or retouched in real time while the user is clicking the shot. The idea is that instead of sending the low resolution image to the cloud for processing, the processing could be done on device itself. This made the system more efficient and processing could be done in real time.

Since researchers have used machine learning algorithm, they trained the system on dataset created by Durand' group and Adobe System, the creator of Photoshop. The dataset includes 5000 images each retouched by 5 different photographers. The system was also trained on 1000s of images enhanced using various image processing algorithms.

The challenge is this technique is that processing is done on low resolution version of the image to save time and resources, but low resolution image is devoid of lot of information. Thus the obtained enhanced image could be different from desired image. To overcome this challenge researchers used two techniques - a) the output of the processing of low-resolution image are a set of mathematical formulae, b) these formulae are applied to each pixel of the final image using some well-defined procedure.

This real time technique for image enhancement using machine learning is ideal for portable devices like smartphones as the technique process low resolution version of the image to save both time and power. Researchers at MIT have claimed to compare their results with other image processing or enhancement algorithm that uses full resolution image for processing and obtaining enhanced image. The full res processing algorithms take 12GB memory space to execute the operation while the proposed algorithm by MIT take about 100MB to execute. Also it is at least 10 times faster making it ideal for smart phone applications.

By - Sagar Juneja, Associate Prof, ECE & Nidhi, Assistant Prof, CSE, Chitkara University, H.P.

References:-

1. http://www.csail.mit.edu/automatic_image_retouching_on_your_phone
2. http://www.zdnet.com/article/google-mits-ai-instantly-fixes-your-smartphone-snaps-as-you-shoot/
3. http://gadgets.ndtv.com/mobiles/news/google-mit-adobe-photo-editing-ai-machine-learning-1733136
4. http://www.irishexaminer.com/breakingnews/world/mit-and-googles-new-algorithms-can-retouch-your-photos-before-you-take-them-800682.html