Spatially varying temporal noise can occur in imaging sensors from nonuniform responsivity of the detectors in the focal plane detector array. Furthermore, some pixels can have extreme responsivities or simply be unresponsive altogether. Such "bad pixels" must be detected and addressed to provide the best possible imagery from a given sensor. Restoration in the presence of bad pixels is a particularly important problem in infrared imaging systems, but it is also an issue with many other camera systems. Bad pixels are traditionally treated by some form of replacement. Rather than performing a simple pixel replacement followed by traditional image restoration, we believe that a superior method of performing restoration in the presence of bad pixels is to perform the restoration and bad pixel replacement jointly. This way, we are able to exploit knowledge of each pixel's characteristics in the restoration process. When a simple pixel replacement method is used, knowledge of which pixel was originally bad is often lost and not exploited in any subsequent image restoration.
In this thesis we propose and compare two methods for scene-based bad pixel detection. We also adapt the FIR adaptive Wiener Filter (AWF) to perform image restoration in the presence of bad pixels on other forms of spatially varying noise. The AWF estimates each pixel using a weighted sum of neighboring pixel values. The weights are determined based on spatially varying autocorrelation models which may use specific knowledge of each pixel's noise characteristics. We propose a fast version of the AWF that is able to handle the bad pixels by using pre-computed weights in a table look-up process. We also propose a modified version of the Non-Local Means (NLM) filter that is robust in handling bad pixels, as it provides background noise reduction. We quantitatively and subjectively compare the performance of the AWF and NLM methods along with several other benchmark restoration methods using simulated images with spatially varying noise and bad pixels. We use both simulated and real infrared imagery to test the proposed algorithms. A computational complexity analysis is also provided to provide further insight into the comparison between the AWF and NLM methods.