Optical Neural Networks Usher in New AI Imaging Era
Convolutional Neural Networks (CNNs) have long been known for their exceptional image recognition capabilities, particularly within platforms like ChatGPT. Recently, a team of Chinese researchers from the University of Shanghai for Science and Technology has taken AI imaging to the next level by introducing CNNs into the field of optics. Led by Prof. Min Gu and Prof. Qiming Zhang from the School of Artificial Intelligence Science and Technology (SAIST) at USST, the research team has successfully developed an ultrafast convolutional optical neural network (ONN).
The team’s groundbreaking research, published in Science Advances under the title Memory-less scattering imaging with ultrafast convolutional optical neural networks, showcases the ability to efficiently and clearly image objects behind scattering media without relying on optical memory effects. Dr. Yuchao Zhang, a researcher at SAIST, led the study as the first author, with Prof. Min Gu and Prof. Qiming Zhang as corresponding authors.
The core concept behind CNNs – convolutional operations – involves extracting local features from images and building complex, abstract representations layer by layer, thus advancing image processing and pattern recognition. However, applying this concept to optics requires overcoming the challenge of converting electronic signals into optical signals. The research team addressed this challenge by designing an all-optical solution to perform convolution network operations directly in the optical domain, achieving lightning-fast optical computing speeds.
At the heart of the research lies a multi-stage convolutional ONN composed of parallel cores capable of operating at the speed of light. This setup extracts features from scattered light to facilitate rapid image reconstruction, significantly enhancing imaging speed and quality. The computational speed of the convolutional ONN reaches an impressive 1.57 Peta operations per second (POPS), supporting real-time dynamic imaging.
One of the key highlights of this research is the network’s multitasking capability. By adjusting the network structure, the convolutional ONN can perform various image processing tasks simultaneously, such as classification and reconstruction, marking a significant advancement in optical artificial intelligence. Prof. Qiming Zhang emphasized that this versatility and efficiency not only underscore the importance of convolutional networks in AI but also open new avenues for optical imaging technology.
Prof. Min Gu envisions a future where convolutional optical neural networks will play a vital role in autonomous driving, robotic vision, and medical imaging. This pioneering research not only marks a successful integration of CNNs into the optical field but also propels AI imaging technology to new heights.
By seamlessly combining convolutional neural networks with optical technology, the research team has set a new standard for AI imaging, promising significant advancements in various industries. The introduction of convolutional optical neural networks represents a milestone in the field of artificial intelligence, paving the way for transformative applications in real-world scenarios.
