Digital Image Processing Image Processing and Restoration

The purpose of this paper is to explore techniques for image enhancement and recovery, especially as they relate to reversing or minimizing the effects of noise degradation. The specific goal of the paper is to attempt to recover readable characters from a license plate image. I start with a short background on the importance of enhancement and then cover some of the technical approaches and techniques. The baseline techniques are some simple enhancement approaches (usharp masking, Sobel edge detection, gray level slicing) and then I move one to more specific denoising filters (contraharmonic, wiener. notch). The final topic is an IEEE transactions paper on nonlinear extrapolation in frequency space that is implemented and tested. I found most of these techniques to be severely inadequate for the level of damage in the target image. Background Image enhancement is the process of making an image more pleasing to a human viewer or more suitable for a specific task. Image enhancement helps us to more quickly and accurately understand the vast amounts of data acquired everyday in real life scenarios. For instance, MRI images can be quite useful in the screening of cancer, but the cancer itself can be difficult to visually separate from living tissue. We can use image enhancement techniques to more drastically highlight the differences and make it easier for a human being to detect the cancer from the images. Another use for image enhancement might be to restore an old film that has been damaged by dust and decay. The frames of the film can be scrubbed of imperfections using certain enhancement techniques so that the audience isn’t distracted by the noise and can better enjoy the film. In our scenario, we have a license plate that has been captured by a real world system. The circumstances of the generation of the image are unknown, but it could easily be that this particular vehicle was nearby or involved with the commission of a crime, and we would like be able to track down the owner of the vehicle via the license plate. Unfortunately, the image is badly degraded and needs to be repaired so that the characters of the license plate can be read. Technical Approaches To begin this project, I attempted to use techniques developed earlier in the class to try to extract useful content from the image. These first approaches would serve as a baseline for later experiments. It quickly became apparent that this task could continue forever, so I limited the number of approaches to investigate. The first choice in my mind was unsharp masking. This technique uses a blurred image and subtracts it from the original to create a mask which is then added back to the original to make it appear sharper than before. I also wanted to observe the effect of edge detection from the Sobel filters to see if I could detect the edges of the characters in the image. In practice, these filters calculate the gradient of the image intensity at each point. They give the direction of the largest possible increase from light to dark and the rate of change in that direction. The result shows how abruptly or smoothly the image changes at that point and how likely that part of the image represents an edge. The last approach I wanted to test was some gray level slicing to see if I could choose appropriate thresholds to separate the letter information from the background color of the license plate. The next set of approaches was formulated to attack the problem more directly and formulaically. I was asked to first try to determine what kind of noise was present in the image, and then attempt to remove that noise using some denoising filters. The filters proposed were the contraharmonic, the notch, and the Wiener. To identify the noise, I selected a slice from the blurred license plate which should have been one uniform color. This slice covered the entire top of the picture. After selecting the slice, I analyzed its histogram and the FFT to make an educated guess about the kind of noise. If I had the camera in hand, I would have been able to identify the noise present through experimentation. More specifically, I would use the same camera and environmental settings to capture a small dot of light. From this small dot of light, I would obtain the FFT and thus I would be able to characterize this system of noise completely (if it is linear and space invariant) with this obtained impulse response. Test patterns would also be useful in obtaining more information on the type blur. By observing the effect on a known pattern it is likely I could devise a technique to perform the inverse of the original blur. I attempted to remove the noise with various techniques including the contraharmonic mean filter, which is a superset of the harmonic mean and arithmetic mean filters. That is, for certain input value, the contraharmonic reduces exactly to the aforementioned filters. For a Q = 0, it becomes the arithmetic mean, and for Q = -1 it becomes the harmonic. This spatial filter is especially good at removing either salt, or pepper noise, but not both at the same time.