It would be of great benefit to develop a single imaging device that can achieve all machine-learning-desired radiation patterns and that can switch its functional theme simply by training with samples of a new target group. Moreover, these imagers are primarily designed to accomplish imaging functions for a specified theme, which cannot be altered after fabrication 1, 2, 3. Specifically, almost all microwave imagers cannot produce the radiation patterns corresponding to the machine-learning optimized measurement modes in real-time and cost-efficient way. However, a gap exists between machine-learning techniques and their direct employment in physical-level microwave imaging, due to the restricted configurations of the imagers mentioned above. Recent advances in optics show that the machine learning can be utilized to conduct the measurements such that the high-quality imaging and high-accuracy object recognition with a few measurements can be realized 9, 10, 11, 12, 13. Such a challenge has been remarkably tackled in recent times by emerging techniques in machine learning 9, 10, 11, 12, 13, 14, 15, 16. To properly process a large data flux of high complexity, it is necessary to study and extract the common features across the data set. In this sense, it is typically required to solve the inverse scattering problem again and again when the scene changes, making them largely ineffective for complex in-situ sensing and monitoring. The optimization solves a time and resource consuming inverse problem for each individual scene. Recently, compressive sensing inspired computational imagers 1, 2, 3, 4, 5 have been proposed to reduce remarkably hardware cost and speed data acquisition, which are at the cost of iterative reconstruction algorithms. However, microwave imagers to date always have to compromise between speed and image quality (fidelity and compression ratio) 1, 2, 3, 4, 5, 6, 7. In this case, just a few relevant data are recorded to reconstruct the scene without losing the information of interest, which is particularly useful for microwave or millimeter-wave radars. In the past decade, a wide variety of computational imagers 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 have been introduced to perform image compression at the physical level, thereby eliminating the need for storage, transfer and processing of the full-pixel original scenes. These situations require an imaging device that can instantly reconstruct scenes in an intelligent and efficient way, i.e., rendering the important feature extraction with high speed, fidelity, and compression ratio, as it commonly happens in biological systems, such as our brain. In many cases, such as for security screening or pipeline monitoring, the flux of data is so large that full-resolution imaging is inefficient and unmanageable, and represents a huge waste of resources and energy, since only a few properties of the images are actually of interest, such as the position of an object, or its dynamic changes 8, 9, 10, 11, 12, 13. Our electronically controlled metasurface imager opens new venues for intelligent surveillance, fast data acquisition and processing, imaging at various frequencies, and beyond.Įfficient microwave imaging systems are becoming increasingly important in modern society, however, rapid processing of information and data poses new challenges for current imaging techniques 1, 2, 3, 4, 5, 6, 7. High-accuracy image coding and recognition are demonstrated in situ for various image sets, including hand-written digits and through-wall body gestures, using a single physical hardware imager, reprogrammed in real time. This imager is electronically reprogrammed in real time to access the optimized solution for an entire data set, realizing storage and transfer of full-resolution raw data in dynamically varying scenes. Here, we experimentally report a real-time digital-metasurface imager that can be trained in-situ to generate the radiation patterns required by machine-learning optimized measurement modes. Conventional microwave imagers usually require either time-consuming data acquisition, or complicated reconstruction algorithms for data post-processing, making them largely ineffective for complex in-situ sensing and monitoring.
0 kommentar(er)
