Using AI to improve image quality on metalens cameras

Machine Learning

Researchers have used deep learning techniques to improve camera image quality using a metalens integrated directly onto a CMOS imaging chip (left). The metalens uses an array of 1000 nm tall cylindrical silicon nitride nanoposts to manipulate light (right). (Image: Ji Chen, Southeast University)

Researchers have leveraged deep learning techniques to improve the image quality of metalens cameras. This new approach uses artificial intelligence (AI) to transform low-quality images into high-quality ones, which allows these cameras to be used for a variety of imaging tasks, including complex microscopy applications and mobile devices.

Metalenses are ultra-thin optical devices (only a few millimeters thick) that use nanostructures to manipulate light. Their small size has the potential to enable very compact and lightweight cameras without traditional optical lenses, but achieving the required image quality with these optical components has been challenging.

“Our technique enables metalens-based devices to overcome limitations in image quality,” said research team leader Dr Ji Chen from Southeast University in China. “This advancement will play an important role in the future development of highly portable consumer imaging electronics, as well as for specialised imaging applications such as microscopy.”

The researchers describe how they used a type of machine learning called a multiscale convolutional neural network to improve resolution, contrast and distortion in images from a miniature camera, approximately 3 cm x 3 cm x 0.5 cm, that they created by integrating a metalens directly onto a CMOS imaging chip.

“A camera with a built-in metalens can be directly integrated into the imaging module of a smartphone, replacing the traditional refractive bulk lens,” Chen said. “It can also be used in devices such as drones, as the small and lightweight camera ensures image quality without compromising the drone's maneuverability.”

Below is Tech Briefs' exclusive interview with Chen, edited for length and clarity.

Technical Overview: What was the biggest technical challenge you faced in leveraging deep learning techniques to improve the image quality of your metalens camera?

Chen: The biggest challenge was that we are not computer science researchers and are not familiar with deep learning algorithms. As a result, we went through a long period of learning and experimentation to determine which neural network to use. Even now, we cannot guarantee that the neural network we use is the best one. Therefore, we would appreciate it if computer science experts could review our work and suggest more suitable neural network algorithms.

Technical Overview: Can you give us a quick rundown of how it all works?

Chen: Generate a large number of pairs of high-quality and low-quality image data, and train the neural network to learn the features of these high-quality and low-quality images. The trained neural network has the ability to process low-quality images into high-quality images. Therefore, when you use the device, you just need to take a low-quality image and send it to the neural network for processing, and you will get a high-quality image result immediately.

Technical OverviewAn article I read said, “Researchers are now designing metalenses with complex capabilities such as color and wide-angle imaging, and developing neural network methods to improve the image quality of these advanced metalenses.” How is that progressing? Any updates?

ChenRecently, we have used an achromatic metalens to achieve color super-resolution imaging using deep learning. A schematic diagram is shown below. The method used in this study is similar to the method used in our 2010 paper. Optical LetterBut now color images can be processed in a more complex way, which requires the metalens to operate over different wavelength ranges. Office Lady The paper pointed out that it only works at a single wavelength. We designed an achromatic metalens structure that works across the entire visible light spectrum from 400 to 700 nm, but of course the lens structure is more complicated. As a result, the neural network structure is also more complicated, and we had to process the data for the RGB channels separately. In the end, we were able to get very good results.

Figure 1. Overall process of deep learning-based super-resolution method for color image capture in a metalens integrated camera. (Image: Researchers)
Figure 2. The metalens integrated camera and its structure. (a) Photograph of the metalens integrated camera. The round metalens region is visible in the center. (b) The structure of the metalens integrated camera and color imaging of distant objects. The enlarged image shows the image details on the CMOS sensor. Due to the limited pixel size of the sensor, it is difficult to capture the image details clearly. (c) Two types of metalens nanostructures: hollow GaN nanopillars (left) and solid GaN nanopillars. (d) SEM image of a GaN metalens. (e) Focusing performance of an achromatic metalens. Different wavelengths are focused at approximately the same position. (Image: Researchers)

Technical Overview: From there, what's the next step? Any plans for more work?

ChenWe then perform intelligent content analysis and recognition on the enriched images, such as identifying whether the image contains an object and determining the object's orientation and location. This is important for communication, positioning and detection purposes. We are currently conducting research in this area. Furthermore, we aim to apply this technique to microscopic images to identify the content of the enriched microscopic images, such as recognizing pathological cells, which will also play an important role in medical diagnostics.

Technical OverviewThe article also states that “commercial applications of this technology would require new assembly techniques to incorporate the metalens into smartphone imaging modules, as well as image-enhancing software designed specifically for mobile phones.” Are there any plans for this?

Chen: We are currently working on a project in collaboration with a smartphone lens manufacturer in China. The main purpose of the project is to replace one or two refractive lenses in a smartphone lens module with metalenses and comprehensively analyze the imaging performance of the new lens module. This project is currently underway.

Technical Overview: What advice (broadly speaking) would you give to researchers looking to bring their ideas to fruition?

ChenMy personal advice:

1) Define a clear goal: Start with a clear vision and specific goals. Know what you want to achieve and break it down into manageable milestones.

2) Conduct thorough research: Before you begin your research, make sure you have an in-depth understanding of the field. This includes reviewing the existing literature, identifying gaps, and understanding current trends and technologies.

3) Build a strong network: Collaborate with others in your field. Networking can provide you with new perspectives, resources, and potential collaborators who can help advance your research.

4) Stay persistent and resilient: Research can be difficult and often involves setbacks. Stay persistent, learn from your mistakes and remain adaptable to overcome obstacles.

5) Focus on practical applications: Think about how your research can be applied to the real world. Practical applications can increase the impact of your research and attract more interest and support.

6) Keep learning and adapting: Your field of study is constantly evolving. Stay up to date with the latest developments, continually learn new skills, and adapt your approach based on new information.

Technical Overview: Is there anything else you'd like to add that I didn't mention?

ChenThe method of using neural networks to improve the image quality of metalens cameras includes not only the approach proposed in this paper to process the imaging results, but also the employment of neural networks to intelligently design the microstructure of metalens, which can transform the originally regular and periodic structures into non-periodic and arbitrarily shaped complex structures to improve the performance of metalens and enrich its functionality.

Source link

Leave a Reply

Your email address will not be published. Required fields are marked *