Research On Algorithms Of Nonuniformity Correction And Enhancement For Infrared Images

Due to the immature manufacturing process of infrared focal plane array,the responsivity of individual infrared detector varies from detector to detector,resulting in the problem of nonuniformity.As a result,there is inherent fixed-pattern-noise,which drifts slowly over time.In addition,the temperatures of objects in the scene of a single frame usually occupies small sub-ranges in the wide dynamic range of the infrared detector,which leads to low contrast and blur boundary between objects in the raw infrared image.Therefore,nonuniformity correction and image enhancement are indispensable pre-processing for infrared imaging system.This dissertation focuses on the research of nonuniformity correction and image enhancement,and its main works and contributions are as follows.1.In staring infrared focal plane array,traditional neural networks based nonuniformity correction methods generally assume that the fixed-pattern-noise is independent and identically distributed.However,in low-cost uncooled infrared detectors,the stripe noise brought by nonuniformity is significant and violates the assumption,leading to the difficulty of preserving edges and suppressing stripes at the same time.To address this problem,a new nonuniformity correction method based on adaptive sparse representation is proposed.Based on the sparse representation theory,the atoms in the over-complete dictionary learned from a small sized corpus of clean infrared image are able to reconstruct the scene radiance sparsely.Therefore,within an adaptive error tolerance,the image reconstructed by sparse representation preserves edges and removes the fixed-pattern-noise as redundant.Experimental results demonstrate that,compared to traditional neural networks based methods,the proposed method reduces the root mean square error by 1.1652 to 1.9107(a decrease of 17.92% to 26.37%),which preserves image edges and removes the fixed-pattern-noise effectively.2.As for the correction of nonuniformity in infrared line scanner,if the methods for staring focal plane array are adopted directly,these algorithms still require hundreds of frames to compute the correction coefficients,which converge slowly.Traditional correction methods for infrared line scanner utilize the characteristic of scanning imaging to update the correction coefficients column by column,and accomplish correction in a single frame.However,the thermal radiance of scene captured in a single frame generally lacks diversity,which leads to the problem of local optima for nonuniformity correction.For the applications where image sequences are available without the requirement of singleframe-correction,a registration based nonuniformity correction method is proposed for the infrared line scanner.In the proposed method,interframe motion is utilized to improve the quality of nonuniformity correction,combined with the characteristic of scanning imaging to accelerate convergence.Experimental results demonstrate that,compared to traditional correction methods for line scanner,the proposed method reduces the metric of roughness by 0.0072 to 0.0306(a decrease of 8.15% to 27.39%),which avoids the error caused by the lack of diversity of thermal radiance and obtains better correction.3.Methods based on the occurrence probability histogram are widely used for infrared image enhancement.However,in the applications such as security monitoring and military reconnaissance,the objects are generally far from the infrared camera and occupy a small fraction of pixels in image.In this case,the probabilities of target-related intensity levels are low,while the probabilities of objects-related intensity levels are high,leading to the over-enhancement on background and under-enhancement on targets.To address this problem,a saliency weight based global mapping enhancement method is proposed.Based on the characteristic that objects have larger saliency values than background,the proposed method obtains a new weight computed from visual saliency,where objects-related intensity levels have larger saliency weight than background-related intensity levels.Thus,the contrast between objects and background can be intensively improved by the saliency weight.Experimental results demonstrate that,compared to traditional probability histogram based methods,the proposed method reduces the metric of linear index of fuzziness by 0.0404 to 0.1740(a decrease of 15.33% to 40.48%),which effectively suppress the problem of over-enhancement on background and under-enhancement on objects when objects merely occupy a small fraction of pixels in the infrared image.