Search

Raw scanned 3D point clouds are usually irregularly distributed due to the essential shortcomings of laser sensors, which therefore poses a great challenge for high-quality 3D surface reconstruction. This paper tackles this problem by proposing a local hierarchical clustering (LHC) method to improve the consistency of point distribution. Specifically, LHC consists of two steps: 1) adaptive octree-based decomposition of 3D space, and 2) hierarchical clustering. The former aims at reducing the computational complexity and the latter transforms the non-uniform point set into uniform one. Experimental results on real-world scanned point clouds validate the effectiveness of our method from both qualitative and quantitative aspects.

Stereo matching is a hot topic in computer vision, while stereo matching in large textureless regions and slanted planes are still challenging problems. We propose a novel stereo matching algorithm to handle the problems. We novelly utilizes minimum spanning tree (MST) to construct a new superpixel-based neighboring system. The neighboring system is used to improve the matching performance in textureless regions. Then we apply the new neighboring system to the stereo matching problem, which uses the superpixel as the matching primitive. The use of the new neighboring system is efficient and effective. We compare our method with 4 popular methods. Experiments on Middlebury dataset show that our method can achieve good matching results. Especially, our method obtains more accurate disparity in textureless regions while maintaining a comparable performance of matching in slanted planes.

Generally, the process of monocular depth map generation consists of two stages: structure from motion and multi-view stereo. The multi-view stereo relies on the accuracy of the estimated camera pose and the photo-consistency assumption. However, the current methods cannot tackle the multi-view matching problem well because of the dependence on the accurate camera pose as well as the matching uncertainty. In this paper, to handle these issues, a new sparse-to-dense diffusion framework is put forward. First, the scene information is reconstructed from SFM (Structure From Motion) and the sparse point cloud is available instead of the camera pose. Secondly, the sparse depth point is re-projected to every frame as the depth label. Finally, the depth label is spread to the remaining pixels through a diffusion process. In addition, the edge detectors are used to make the propagation better-regulated. Experimentally, the results show that the proposed framework can robustly generate the depth map from monocular videos.

The 3D target tracking based on the infrared image is a challenge problem in computer vision. With the development and progress of the target tracking in recent years, key point-based target tracking is widely applied in many fields, such as augmented reality, object retrieval, human computer interaction and medical imaging. Generally, the aim of target tracking is to locate the targets in the sub-sequent frames. Because the tracking is a fundamental research topic, it should be self-adaptive under some special situations,, such as illumination variation, occlusion, background clutters and so on. Thereby it is hard to fit all scenarios by single tracking approach. And in a particle filter framework, computational cost grows linearly with the number of particles sampled. Hence, it is important to decrease computational cost but also get good performance. In the process of the tracking 3D target, traditional tracking algorithms, such as template matching using fast normalized cross correlation or HOG (Histogram of Oriented Gradients), perform not so well. The main reasons are that they are not scale and rotation invariant matching. Mean shift tracking algorithm takes advantage of a probability density to search for the position of the target, which fails in abrupt changes of luminance. Furthermore, the feature selection is important in obtaining high accuracy. So the computational cost may become a bottleneck in tracking problem. This paper proposes an effective tracking framework based on FAST BRIEF and Particle Filter (FBPF). Our approach can be divided into four steps as follows:(1) An advanced fast algorithm which includes decision trees is adopted to extract key points. (2) The BRIEF descriptor is applied to speed up the feature matching.(3)Calculate the affine transformation parameters. (4)Particle filter is adopted to predict the best position of target. The key point of traditional FAST algorithm is defined as point which has different gray value with enough surrounding constant pixels. Applied in the gray images, it means there are enough and constant pixels with more or less gray value than itself. The circle area is usually selected when taking any pixel in the image and its round area. In order to decrease the cost of calculation, we employ an advanced FAST based on decision tree, which learns the key points with the maximum entropy. Based on the binary descriptor BRIEF, the original transformation matrix can be calculated efficiently. Then, in order to estimate the best position of the target, particle filter is applied. We take the parameters as particles. It provides a convenient framework for estimating and propagating the posterior probability density function of state variables regardless of the underlying distribution through a sequence of prediction and update steps. Experimental results demonstrate that FBPF achieves state-of-the-art performance on the challenge sequences. In this paper, the main contribution is the proposal of a new tracking framework which based on FAST BRIEF and Particle Filter (FBPF). First, we employ an advanced FAST using the decision tree and BRIEF is used to establish feature for decreasing time consumption. Second, we calculate the original affine matrix base on the feature matching. Finally, the particle filter is utilized to predict the best position.

Object detection is one of the most important researches in computer vision. Recently, category-independent objectness in RGB images has been a hot field for its generalization ability and efficiency as a pre-filtering procedure of the object detection. Many traditional applications have been transferred from the RGB images to the depth images since the economical depth sensors, such as Kinect, were popularized. The depth data represents the distance information. Because of the special characteristic, the methods of objectness evaluation in RGB images are often invalid in depth images. In this study, we propose mEdgeboxes to evaluate the objectness in depth image. Aside from detecting the edge from the raw depth information, we extract another edge map from the orientation information based on the normal vector. Two kinds of the edge map are integrated and are fed to Edgeboxes1 in order to produce the object proposals. The experimental results on two challenging datasets demonstrate that the detection rate of the proposed objectness estimation method can achieve over 90% with 1000 windows. It is worth noting that our approach generally outperforms the state-of-the-art methods on the detection rate.

Blurred face recognition is still remaining as a challenge task, but with wide applications. Image blur can largely affect recognition performance. The local phase quantization (LPQ) was proposed to extract the blur-invariant texture information. It was used for blurred face recognition and achieved good performance. However, LPQ considers only the phase blur-invariant texture information, which is not sufficient. In addition, LPQ is extracted holistically, which cannot fully explore its discriminative power on local spatial properties. In this paper, we propose a novel method for blurred face recognition. The texture and structure blur-invariant features are extracted and fused to generate a more complete description on blurred image. For texture blur-invariant feature, LPQ is extracted in a densely sampled way and vector of locally aggregated descriptors (VLAD) is employed to enhance its performance. For structure blur-invariant feature, the histogram of oriented gradient (HOG) is used. To further enhance its blur invariance, we improve HOG by eliminating weak gradient magnitude which is more sensitive to image blur than the strong gradient. The improved HOG is then fused with the original HOG by canonical correlation analysis (CCA). At last, we fuse them together by CCA to form the final blur-invariant representation of the face image. The experiments are performed on three face datasets. The results demonstrate that our improvements and our proposition can have a good performance in blurred face recognition.

Image matching is a fundamental task in computer vision. It is used to establish correspondence between two images taken at different viewpoint or different time from the same scene. However, its large computational complexity has been a challenge to most embedded systems. This paper proposes a single FPGA-based image matching system, which consists of SIFT feature detection, BRIEF descriptor extraction and BRIEF matching. It optimizes the FPGA architecture for the SIFT feature detection to reduce the FPGA resources utilization. Moreover, we implement BRIEF description and matching on FPGA also. The proposed system can implement image matching at 30fps (frame per second) for 1280x720 images. Its processing speed can meet the demand of most real-life computer vision applications.

Aiming at the problem of tracking 3D target in forward-looking infrared (FLIR) image, this paper proposes a high-accuracy robust tracking algorithm based on SIFT and particle filter. The main contribution of this paper is the proposal of a new method of estimating the affine transformation matrix parameters based on Monte Carlo methods of particle filter. At first, we extract SIFT features on infrared image, and calculate the initial affine transformation matrix with optimal candidate key points. Then we take affine transformation parameters as particles, and use SIR (Sequential Importance Resampling) particle filter to estimate the best position, thus implementing our algorithm. The experiments demonstrate that our algorithm proves to be robust with high accuracy.

In planetary or lunar landing missions, hazard avoidance is critical for landing safety. Therefore, it is very important to correctly detect hazards and effectively find a safe landing area during the last stage of descent. In this paper, we propose a passive sensing based HDA (hazard detection and avoidance) approach via descent images to lower the landing risk. In hazard detection stage, a statistical probability model on the basis of the hazard similarity is adopted to evaluate the image and detect hazardous areas, so that a binary hazard image can be generated. Afterwards, a safety coefficient, which jointly utilized the proportion of hazards in the local region and the inside hazard distribution, is proposed to find potential regions with less hazards in the binary hazard image. By using the safety coefficient in a coarse-to-fine procedure and combining it with the local ISD (intensity standard deviation) measure, the safe landing area is determined. The algorithm is evaluated and verified with many simulated descent downward looking images rendered from lunar orbital satellite images.

The automatic observation of the field crop attracts more and more attention recently. The use of image processing technology instead of the existing manual observation method can observe timely and manage consistently. It is the basis that extracting the wheat from the field wheat images. In order to improve accuracy of the wheat segmentation, a novel two-stage wheat image segmentation method is proposed. Training stage adjusts several key thresholds which will be used in segmentation stage to achieve the best segmentation results, and counts these thresholds. Segmentation stage compares the different values of color index to determine which class of each pixel is. To verify the superiority of the proposed algorithm, we compared our method with other crop segmentation methods. Experiment results shows that the proposed method has the best performance.

Images are usually degraded by haze and fog due to atmospheric absorption and scattering. Especially for remote sensing, degraded images which suffer from haze will highly probably lose detailed information. Nowadays researches on haze removal mainly focus on visible images. However, as infrared imaging is more and more widely used in many fields, a comprehensive study of the effects of fog on infrared images is urgent. In this paper, we firstly introduce that the obvious attenuation effects of fog on infrared images by analyzing the mid-infrared images captured under different weather conditions. Therefore, we propose a unified dehazing approach for infrared images. The proposed approach mainly includes three steps. The first step is to obtain the local transmission of the original hazed images by using the statistical prior knowledge known as “dark channel” which is similar to visible images. With local transmission, the refined transmission map is estimated in a soft matting framework. Due to atmospheric scattering model, dehazed infrared images can be recovered finally. From experimental results, it is quantitatively demonstrated that the proposed approach can significantly improve the quality of hazed images by using principle component analysis technique.

In this paper, we propose a specularity-invariant crop extraction method using probabilistic super-pixel markov random field (MRF). Our method is based on the underlying rule that intensity change gradually between highlight areas and its neighboring non-highlight areas. This prior knowledge is embedded into the MRF-MAP framework by modeling the local and mutual evidences of nodes. The marginal probability of each node in the label field is then iteratively computed by Belief Propagation algorithm which leads to the final solution. Comparing experimental results show that our method outperforms the other commonly used extraction methods in yielding highest performance with the lowest standard deviation.

We propose a novel robust texture descriptor, the Gradient Orientation and Modulus Matrix (GOMM), which is based on the fact that human perception of an image pattern depends not only on its intensity, but also on changes in intensity and regularity (such as the gradient modulus and gradient orientation of the image). A GOMM is constructed in three steps. First, the gradient orientation of each pixel is mapped onto N intervals and the gradient modulus is partitioned into M levels. Next, a block is constructed from the gradient modulus of pixels whose gradient orientations are mapped onto the same interval. Then, each component of the GOMM is given by the sum of the ratios between two terms, namely, the differential gradient modulus grading between the elements in the above-mentioned block, and the distance between the elements. Finally, a six-dimension vector is calculated from each GOMM. By rearranging feature vectors from each GOMM, we can concatenate the vectors to construct a uniform GOMM feature for a given image, irrespective of the angle of the image. Experimental results on the KTH-TIPS2 (The Royal Institute of Technology - Textures under varying Illumination, Pose and Scale) image database show that the GOMM significantly outperforms the other classical descriptors.

Aiming at the problem of in-harbor ship detection in forward-looking infrared image, this paper proposes a method for ship segmentation and false alarm suppressing based on k-means clustering segmentation. We obtain the simulated model images from visible satellite images and perspective relations. And the harbor area is determined by matching with HOG features. Then we segment the ship out of the harbor area. In order to suppress the false alarm, we apply k-means clustering segmentation to get the ship and the sea area simultaneously. By calculating the external convex polygon, we get rid of the false alarm targets. Experimental results suggest that our method has high detection accuracies and low false alarm rate.

This paper proposes a new approach to localize buildings from forward looking infrared (FLIR) images. The proposed approach can localize not only large buildings, but also small buildings. Furthermore, the proposed approach is also robust with those FLIR images degraded by clouds. This breakthrough is due to the following improvements: (1) the Histogram of Oriented Gradients approach is improved to match FLIR images with our templates; (2) a new kind of feature image is presented to reduce the difference between template and target; (3) we project 3D building models into images, with different colors on different sides, distinguishing those sides apart; (4) we generate templates which contain all buildings in the visual field. As a result, the FLIR images can be matched with the big templates at a high correct rate, and then target buildings can be localized. The experimental results show the superior performance of the proposed approach.

Frost is a kind of ground coagulation phenomena, and if the temperature of dew point is below 0C o , the water vapor condenses as solid, which is called frost. The frost phenomena observing is an important step in daily ground observation work, and the results is one of 36 critical data in meteorological observation field. This work is usually accomplished by manual. In this paper, we propose an effective method for frost observation based on image processing. The changing of frost formation process is well simulated by using the curve fitting of gray correlation coefficient between certain lengths of frames, while the characteristic of frost surface texture is also well described by texture analysis based on texture descriptor. The experiment results show that our method can get high detection accuracy in the different kinds of continuous changing environment.

In this paper, a novel automatic DSM and remote sensing images registration scheme using template matching technique is developed. Due to the heterogeneity of DSM and remote sensing images, the emphases of our scheme are to describe the common feature between DSM and remote sensing images, and to generate a suitable template for template matching. Based on the sparse representation theory, we present a new feature descriptor, which can highlight the similarities of DSM and remote sensing images, and can be used to form a new kind of feature image. Meanwhile we present a criterion to choose the proper region from the feature image as the template which will ensure perfect template matching performance. The experiment results show that our scheme is efficient to fulfill the task of registration.

Rock is one kind of dangerous hazards for lunar landing, a small rock about 0.5-1.0 meter tall can damage the lander and make the whole mission failed1. In this paper, we provide an improved method for rock detection based on shadows and texture for safe lunar landing, which can detect and model the rocks in the lunar surface image precisely. Although traditional rock detection algorithms can get a good result in the experimental environment full of rocks, it does not work well in the real, uneven environment with all kinds of hazards. Our algorithm can adapt to various environment because of the precisely extracting shadows of rocks and texture analyses after rock modeling.

This paper firstly improves D.L.Plillips's representation about image restoration and then points out that image restoration is just a 'partial ill-posed' problem rather than a 'total ill-posed' problem---amplitude restoration is ill-posed but phase restoration is well-posed. Basing on the viewpoints, the paper proposes a restoration method, which cuts down phase-pollution caused by traditional regularization methods, that amplitude restoration is realized by regularization and phase restoration is achieved by algebraic method. Experimental results indicate that the proposed method performs well. It can efficiently restore image phase and elaborately preserve image details.

This paper analyzes the difference between the imaging mechan ism of the infrared images an d that of the visible light images, and find that it is important to extract the stable and reliable common feature for object recognition. Then we propose a target recognition algorithm based on histograms of oriented gradients (HOG) which evaluates normalized local histograms of image gradient orientations in a dense grid. Last we adopt linear SVM trained for a binary object/non-object classifier a nd detect the object in the fo rward-looking infrared (FLIR) images. The experiment results suggest that the proposed approach has high rates of detection. Furthermore, we study how to select positive and negative samples for a better performance. Keywords: HOG, FLIR image, Linear SVM, Target recognition 1. INTRODUCTION In automatic target recognition system with forward-looking infrared (FLIR) images, the imaging mechanism is different from that of the visible light because the infrared images depend only on the temperature of the object itself and the radiation of heat. Owing to the solar radiation and the air temp erature as well as the constant heat exchange between the target and its surrounding environment, there are great differences between the grayscale characteristics of the same object at different moments. Moreover, the scale and angle deviation and a variety interference of noise in distanced FLIR images also complicate the target recognition process. So the template matching algorithm based on normalized cross-correlation is not good for automatic target recognition of FLIR images. Also the Hausdorff distance measures algorithm

Color image segmentation draws a lot of attention recently. In order to improve efficiency of spectral clustering in color image segmentation, a novel two-stage color image segmentation method is proposed. In the first stage, we use vector gradient approach to detect color image gradient information, and watershed transformation to get the pre-segmentation result. In the second stage, Nystrom extension based spectral clustering is used to get the final result. To verify the proposed algorithm, it is applied to color images from the Berkeley Segmentation Dataset. Experiments show our method can bring promising results and reduce the runtime significantly.

SIFT tracking algorithm is an excellent point-based tracking algorithm, which has high tracking performance and accuracy due to its robust capability against rotation, scale change and occlusion. However, when tracking a huge 3D target in complicated real scenarios in a forward-looking infrared (FLIR) image sequence taken from an airborne moving platform, the tracked point locating in the vertical surface usually shifts away from the correct position. In this paper, we propose a novel algorithm for 3D target tracking in FLIR image sequences. Our approach uses SIFT keypoints detected in consecutive frames for point correspondence. The candidate position of the tracked point is firstly estimated by computing the affine transformation using local corresponding SIFT keypoints. Then the correct position is located via an optimal method. Euclidean distances between a candidate point and SIFT keypoints nearby are calculated and formed into a SIFT-based distance histogram. The distance histogram is defined a cost of associating each candidate point to a correct tracked point using the constraint based on the topology of each candidate point with its surrounding SIFT keypoints. Minimization of the cost is formulated as a combinatorial optimization problem. Experiments demonstrate that the proposed algorithm efficiently improves the tracking performance and accuracy.

In this paper, the joint distribution of the Nonsubsampled Contourlet Transform coefficients is studied. It is found that the estimation of the joint distribution is implement impossible due to the complex of joint empirical distribution function and dependence of NSCT coefficient vector components. To distinguish different joint distributions of different images, the sample covariance matrix feature is proposed. The texture retrieval experiment is conducted in order to evaluate the performance of the sample covariance matrix feature. The result shows that the proposed feature is efficient in representing the texture and the difference of the joint distribution of the Nonsubsampled Contourlet Transform coefficients.

With the advent of information age, especially with the rapid development of network, "information explosion" problem has emerged. How to improve the classifier's training precision steadily with accumulation of the samples is the original idea of the incremental learning. Support Vector Machine (SVM) has been successfully applied in many pattern recognition fields. While its complex computation is the bottle-neck to deal with large-scale data. It's important to do researches on the SVM's incremental learning. This article proposes a SVM's incremental learning algorithm based on the filtering fixed partition of the data set. This article firstly presents "Two-class problem"s algorithm and then generalizes it to the "Multiclass problem" algorithm by the One-vs-One method. The experimental results on three types of data sets' classification show that the proposed incremental learning technique can greatly improve the efficiency of SVM learning. SVM Incremental learning can not only ensure the correct identification rate but also speedup the training process.

This paper presents an implementation of a real-time video surveillance system based on TMS320DM6437 DVDP (Digital Video Development Platform). We propose an improved motion detection method based on adaptive Gaussian Mixture Model. In tracking module, association matrix is built between two frames for targets association according to a kind of cost function of features. Besides, Mean Shift algorithm is used to predict targets' parameters, so target tracking can be maintained in the case of short-term loss. In the test scene, the system shows a satisfactory performance.

Correlation match is widely used in the field of target recognition for its stability and adaptability, but traditional correlation matching does not meet the demand for real time performance. For this reason, this paper presents an improved recursive pyramid correlation image matching algorithm and accomplishes the algorithm's parallel implementation on Multi-DSPs interconnected by FPGA with using several effective optimization strategies. Moreover, we design detailed experiments which give a comparison of different matching speed using optimization methods step by step. The results prove that our system is the best optimum with high efficiency compared with conventional methods.

Automatic 3-D building reconstruction is the key technology of "digital city" construction, which can be widely used in urban planning and management. In this paper, we firstly analyze the disadvantages in traditional automatic building reconstruction methods, and then according to characteristics of complex buildings, which is multi-storey and non-flattopped, a new automatic building reconstruction method is proposed, which can solve the complex building reconstruction problem effectively and conveniently. The reconstruction experiments show that the proposed method has excellent universality, besides a small calculating amount compared with traditional methods.

This paper focuses on preprocessing methods for the infrared images. Due to noises, low contrast and local difference of materials, Forward-looking Infrared targets have a great difference of local gray levels. Such different gray levels add the difficulty of targets detection and may lead a decrease in accuracy and rate of detection in Gray-based template matching. To overcome the difficulty, we present a novel preprocessing method based on non-subsampled contourlet transform (NSCT) which is good at multi-direction and multi-scale. This method firstly do Non-subsampled contourlet transform to infrared images. The level of decomposition is decided by the size of the template which is considered as the basic structure and the directions of decomposition are chosen by the imaging angles. Then we reconstruct the image according to the coefficients of the decomposition. Experiments based on template matching in forward-looking infrared images show that the proposed preprocessing method based on NSCT can significantly improve the detection rate (DR) and the stability. Experimental results also show that our proposed method has better performance than some other preprocessing methods.

2-D Convolution is a simple mathematical operation which is fundamental to many common image processing operators. Using FPGA to implement the convolver can greatly reduce the DSP's heavy burden in signal processing. But with the limit resource the FPGA can implement a convolver with small 2-D kernel. In this paper, An FIFO type line delayer is presented to serve as the data buffer for convolution to reduce the data fetching operation. A finite state machine is applied to control the reuse of multipliers and adders arrays. With these two techniques, a resource limited FPGA can be used to implement a larger kernel convolver which is commonly used in image process systems.

The paper presents a design method for high-speed image processing system based on FPGA+DSP+ASIC structure. According to the feature of the devices, the system reconfigures the data path and distributes the algorithm flexibly among DSP, FPGA, and the ASIC. Based on the above processing elements, an infrared small target detecting parallel system is constituted, which makes better use of the advantages of DSP, FPGA and ASIC while avoiding their disadvantages. Finally, the prototype test shows good performance and can meet the expected design target related to real-time processing.

In this paper, a pixel-level image fusion algorithm based on Nonsubsampled Contourlet Transform (NSCT) has been proposed. Compared with Contourlet Transform, NSCT is redundant, shift-invariant and more suitable for image fusion. Each image from different sensors could be decomposed into a low frequency image and a series of high frequency images of different directions by multi-scale NSCT. For low and high frequency images, they are fused based on local-contrast enhancement and definition respectively. Finally, fused image is reconstructed from low and high frequency fused images. Experiment demonstrates that NSCT could preserve edge significantly and the fusion rule based on region segmentation performances well in local-contrast enhancement.© (2007) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

The recognition of bridge in long-range infrared images presents a number of problems due to the complexity of background, high noise interference, small size of target and low contrast between bridge and its surrounding water area. To counter these barriers, we have developed a new knowledge-based recognition algorithm. It first detects the candidate bridge sub-regions and then focus on them. According to the degree to which they match with our pre-built framework, different credits are given, so the false objects are excluded and eventually real target is found. The experimental results demonstrate our localized method is always superior to the traditional global algorithms adopted by most former researchers.

The generative technique of the dynamic infrared image was discussed in this paper. Because infrared sensor differs from CCD camera in imaging mechanism, it generates the infrared image by incepting the infrared radiation of scene (including target and background). The infrared imaging sensor is affected deeply by the atmospheric radiation, the environmental radiation and the attenuation of atmospheric radiation transfers. Therefore at first in this paper the imaging influence of all kinds of the radiations was analyzed and the calculation formula of radiation was provided, in addition, the passive scene and the active scene were analyzed separately. Then the methods of calculation in the passive scene were provided, and the functions of the scene model, the atmospheric transmission model and the material physical attribute databases were explained. Secondly based on the infrared imaging model, the design idea, the achievable way and the software frame for the simulation software of the infrared image sequence were introduced in SGI workstation. Under the guidance of the idea above, in the third segment of the paper an example of simulative infrared image sequences was presented, which used the sea and sky as background and used the warship as target and used the aircraft as eye point. At last the simulation synthetically was evaluated and the betterment scheme was presented.

Image segmentation technique is the foundation of the high-level digital image processing, and it is widely applied in many areas, it is also a classic difficult problem on the domain of advanced information processing. Because of itsi» importance and difficulties, image segmentation processing motivates large numbers of researchers to work for it, and quite a number of segmentation thoughts and algorithms have been proposed over the years. But by far, it is very difficult to provide a general effective segmentation algorithm, and further more there is no objective criteria to value the performance of the image segmentation algorithms. From the theory research of the image segmentation, there are two main directions:1). To improve the classic segmentation algorithms. 2). To provide the novel thought and novel approaches. From the application of the image segmentation. it is mainly applied for ATRs(Automatic Target Recognition system) and industry testing. Because the segmentation cannot only greatly compress data and reduce the required memory space but also simplify the analysis and the following processing steps. In this paper, we propose a new application of the image segmentation-image matching in real-time systems. And the key problem is that we need provide the best segmentation methods which are suitable to the following matching processing. The rule to select the segmentation method is also provided. Using the traditional matching scheme, the experimental results show that performance of the segmentation algorithm(2-D OTSU) is unstable, and the correct matching probability (CMP) is increased rapidly when the size of real images(matching basic unit) become larger and larger, compared with the tradition methods, some of them are better than the direct gray-level image matching when the real image becomes large, but the average CMP(ACMP) is not good. To improve the ACMP, we provide a new method for image matching-combined matching. We make every three sequential images a group, and make this group as the matching basic unit.

All IRFPAs require nonuniformity correction. Although two- point or multi-point correction algorithms may correct the nonuniformity of IRFPAs they can be limited by pixel nonlinearities and instabilities. So adaptive nonuniformity correction techniques are needed. Many researchers develop the methods of real-time correction based the scene being viewed. The nonuniformity correction process is completed in IRFPA sensor named smart FPA. However, the smart IRFPA is developing. The purpose of this paper is to describe a digital signal processing electronics for the nonuniformity correction. It includes ADSP21060 digital signal processor, 8751 chip and display circuit module. The image data from IRFPA are put into the dual RAM. The ADSP21060 DSP completes the nonuniformity correction function while the 8751 chip operates control function. At the same time, the correction results will be displayed on a monitor. The neural networks algorithms and the constant-statistics algorithm are tested in our digital implementations. When the image size is 60*97, processing time per frame is 14.75 millisecond for the neural network algorithm and it is 12.65 millisecond for the constant-statistics algorithm. Measured results show that digital processing system designed by us may achieve demand of real-time nonuniformity correction based the scene for small IRFPAs.