The improved scene-based adaptive nonuniformity correction (NUC) algorithms using a neural network (NNT) approach for infrared image sequences are presented and analyzed. The retina-like neural networks using steepest...The improved scene-based adaptive nonuniformity correction (NUC) algorithms using a neural network (NNT) approach for infrared image sequences are presented and analyzed. The retina-like neural networks using steepest descent model was the first proposed infrared focal plane arrays (IRFPA) nonuniformity compensation method,which can perform parameter estimation of the sensors over time on a frame by frame basis. To increase the strength and the robustness of the NNT algorithm and to avoid the presence of ghosting artifacts,some optimization techniques,including momentum term,regularization factor and adaptive learning rate,were executed in the parameter learning process. In this paper,the local median filtering result of AX^U_ ij (n) is proposed as an alternative value of desired network output of neuron X_ ij (n),denoted as T_ ij (n),which is the local spatial average of AX^U_ ij (n) in traditional NNT methods. Noticeably,the NUC algorithm is inter-frame adaptive in nature and does not rely on any statistical assumptions on the scene data in the image sequence. Applications of this algorithm to the simulated video sequences and real infrared data taken with PV320 show that the correction results of image sequence are better than that of using original NNT approach,especially for the short-time image sequences (several hundred frames) subjected to the dense impulse noises with a number of dead or saturated pixels.展开更多
For infrared focal plane graded during signal acquisition array sensors, imagery is departicularly nonuniformity. In this paper, an adaptive nonuniformity correction technique is proposed which simultaneously estimate...For infrared focal plane graded during signal acquisition array sensors, imagery is departicularly nonuniformity. In this paper, an adaptive nonuniformity correction technique is proposed which simultaneously estimates detector-level and readout- channel-level correction parameters using neural network approaches. Firstly, an improved neural network framework is designed to compute the desired output. Secondly, an adaptive learning rate rule is used in the gain and offset parameter estimation process. Experimental results show the proposed algorithm can achieve a faster convergence speed and better stability, remove nonuniformity and track parameters drift effectively, and present a good adaptability to scene changes and nonuniformity conditions.展开更多
We present a robust and fiducial-marker-free algorithm that can identify and correct stick-slip distortion caused by nonuniform rotation(or beam scanning)in distally scanned catheters for endoscopic optical coherence ...We present a robust and fiducial-marker-free algorithm that can identify and correct stick-slip distortion caused by nonuniform rotation(or beam scanning)in distally scanned catheters for endoscopic optical coherence tomography(OCT)images.This algorithm employs spatial fre-quency analysis to select and remove distortions.We demonstrate the feasibility of this algorithm on images acquired from ex vivo rat colon with a distally scanned DC motor-based endoscope.The proposed algorithm can be applied to general endoscopic OCT images for correcting non-uniform rotation distortion.展开更多
The striping pattern nonuniformity of the infrared line scanner （IRLS） severely limits the system performance. An adaptive nonuniformity correction （NUC） algorithm for IRLS using neural network is proposed. It use...The striping pattern nonuniformity of the infrared line scanner （IRLS） severely limits the system performance. An adaptive nonuniformity correction （NUC） algorithm for IRLS using neural network is proposed. It uses a one-dimensional median filter to generate ideal output of network and can complete NUC by a single frame with a high correction level. Applications to both simulated and real infrared images show that the algorithm can obtain a satisfactory result with low complexity, no need of scene diversity or global motion between consecutive frames. It has the potential to realize real-time hardware-based applications.展开更多
文摘The improved scene-based adaptive nonuniformity correction (NUC) algorithms using a neural network (NNT) approach for infrared image sequences are presented and analyzed. The retina-like neural networks using steepest descent model was the first proposed infrared focal plane arrays (IRFPA) nonuniformity compensation method,which can perform parameter estimation of the sensors over time on a frame by frame basis. To increase the strength and the robustness of the NNT algorithm and to avoid the presence of ghosting artifacts,some optimization techniques,including momentum term,regularization factor and adaptive learning rate,were executed in the parameter learning process. In this paper,the local median filtering result of AX^U_ ij (n) is proposed as an alternative value of desired network output of neuron X_ ij (n),denoted as T_ ij (n),which is the local spatial average of AX^U_ ij (n) in traditional NNT methods. Noticeably,the NUC algorithm is inter-frame adaptive in nature and does not rely on any statistical assumptions on the scene data in the image sequence. Applications of this algorithm to the simulated video sequences and real infrared data taken with PV320 show that the correction results of image sequence are better than that of using original NNT approach,especially for the short-time image sequences (several hundred frames) subjected to the dense impulse noises with a number of dead or saturated pixels.
基金supported by the National Natural Science Foundation of China (61101199)the Natural Science Foundation of Jiangsu Province (K2011699)the Colleges and Universities Innovation Projects (CX08B 045Z)
文摘For infrared focal plane graded during signal acquisition array sensors, imagery is departicularly nonuniformity. In this paper, an adaptive nonuniformity correction technique is proposed which simultaneously estimates detector-level and readout- channel-level correction parameters using neural network approaches. Firstly, an improved neural network framework is designed to compute the desired output. Secondly, an adaptive learning rate rule is used in the gain and offset parameter estimation process. Experimental results show the proposed algorithm can achieve a faster convergence speed and better stability, remove nonuniformity and track parameters drift effectively, and present a good adaptability to scene changes and nonuniformity conditions.
基金supported in part by the National Institutes of Health under the grant No.R01 HL121788The Wallace H.Coulter FoundationMaryland Innovation Initiative(MII)Fund from TEDCO
文摘We present a robust and fiducial-marker-free algorithm that can identify and correct stick-slip distortion caused by nonuniform rotation(or beam scanning)in distally scanned catheters for endoscopic optical coherence tomography(OCT)images.This algorithm employs spatial fre-quency analysis to select and remove distortions.We demonstrate the feasibility of this algorithm on images acquired from ex vivo rat colon with a distally scanned DC motor-based endoscope.The proposed algorithm can be applied to general endoscopic OCT images for correcting non-uniform rotation distortion.
基金This work was supported by the Pre-Research Foundation of National Defense under Grant No. 30404.
文摘The striping pattern nonuniformity of the infrared line scanner （IRLS） severely limits the system performance. An adaptive nonuniformity correction （NUC） algorithm for IRLS using neural network is proposed. It uses a one-dimensional median filter to generate ideal output of network and can complete NUC by a single frame with a high correction level. Applications to both simulated and real infrared images show that the algorithm can obtain a satisfactory result with low complexity, no need of scene diversity or global motion between consecutive frames. It has the potential to realize real-time hardware-based applications.
