This video introduces the concept of compressed sensing, a powerful algorithm that can reconstruct a complete, high-fidelity image from a small, randomly sampled subset of its pixels. The technique is demonstrated on a custom-built X-ray backscatter imaging system, a device where acquiring each pixel is a time-consuming process. The video details the construction of the scanner, which uses a gimbal-mounted X-ray tube, analog potentiometers for position sensing, a sensitive photon-counting detector with a photomultiplier tube, and a Teensy microcontroller for system control and data acquisition.
The core principle of the algorithm is explained: it solves the reconstruction problem by assuming that natural images are "sparse" in the frequency domain. It finds the image that fits the measured data points while being composed of the fewest number of unique spatial frequencies. This works because the vast majority of all possible images are random noise, whereas natural, real-world images occupy a tiny, structured subset of this possibility space. By minimizing frequency components, the algorithm effectively chooses the most "natural" and least random solution.
Applying Compressed Sensing to a Custom X-Ray Backscatter Imager
00:00:01 Introducing Compressed Sensing: The video explains compressed sensing, a technique that can reconstruct an image from a small fraction (e.g., 10%) of its pixels. This is especially useful for imaging methods where data acquisition is slow.
00:01:05 X-Ray Backscatter Explained: This imaging method works like "X-ray goggles," scanning a thin beam of X-rays across an object and collecting the reflection. It doesn't require placing a detector behind the object. Materials with low atomic numbers (like carbon and hydrogen) appear bright, while dense, high-atomic-number materials are dark.
00:02:03 The Need for Speed: The X-ray backscatter signal is extremely weak, requiring a photon-counting detector and long dwell times (8 milliseconds per pixel). A full high-resolution scan (800x800 pixels) would take approximately 85 minutes. Compressed sensing drastically reduces this time to 10-20 minutes by only sampling a fraction of the pixels.
00:04:05 Random Sampling is Crucial: The algorithm requires random, non-uniform sampling of the image. This allows it to accurately determine all frequency components and avoids blind spots that would occur with regular, grid-like sampling. The scanner's slight mechanical imperfections naturally contribute to this randomness.
00:05:49 Scanner Hardware and Control: The system is built on a motorized gimbal mount. Position is measured with analog potentiometers, and a Teensy microcontroller coordinates the motors, reads the sensors, and communicates with a PC.
00:09:22 Photon-Counting Detector: The detector consists of a phosphor screen (from an X-ray film cassette) that converts X-rays to light, a photomultiplier tube (PMT) to detect the faint light flashes, and a transimpedance amplifier. The Teensy's built-in analog comparator and hardware counter efficiently count the photon pulses without using CPU time.
00:12:41 The Core Algorithm: Compressed sensing reconstructs the image by finding a solution that both fits the measured data points and uses the fewest number of unique frequencies. This is based on the principle that natural images are inherently "sparse" in the frequency domain.
00:15:34 Why It Works - Natural Images vs. Noise: The video explains that virtually all "natural" images (photos, scans, etc.) share the property of being compressible and having minimal frequency components. In contrast, a random noise image contains all frequencies. The algorithm is biased to find solutions in the tiny sliver of "natural images" out of the astronomically large space of all possible images, thus avoiding a random, noisy reconstruction.
00:20:13 Open Source Project: All the project files, including the code and processing scripts for data capture and reconstruction, are available on GitHub for others to learn from and experiment with.