2011 Vol. 4, No. 5
Abstract
PDF 722KB
2011, 4(5): 431-440.
Abstract:
Developing trend of image processing intelligence is introduced in this paper, then the theory research, effective algorithms of intelligent image processing and several relative issues for parallel disposal hardware systems are summarized and reviewed. It points out that image processing intelligence is very significant for the Automatic Target Recognition(ATR) in military affairs, optical imaging instruments, internet video communication, or various high-tech evolutions, and it will be the development trend for image processing fields in the future. Finally, it suggests that the research in recent should be focused on the Automatic Target Acquiring(ATA) in complex backgrounds.
Developing trend of image processing intelligence is introduced in this paper, then the theory research, effective algorithms of intelligent image processing and several relative issues for parallel disposal hardware systems are summarized and reviewed. It points out that image processing intelligence is very significant for the Automatic Target Recognition(ATR) in military affairs, optical imaging instruments, internet video communication, or various high-tech evolutions, and it will be the development trend for image processing fields in the future. Finally, it suggests that the research in recent should be focused on the Automatic Target Acquiring(ATA) in complex backgrounds.
2011, 4(5): 441-447.
Abstract:
Traditional Shannon sampling method leads to a large amount of image data, and massive data processing brings a great pressure to bear on the post-processing of image information. Compressed Sensing(CS) theory which can overcome the problem mentioned above is researched in this paper. It can reconstruct a large amount data by sampling small quantity data, and breakthroughs the restriction of Shannon sampling theory. This paper reviews the theory and key technique of CS, and introduces the application and development of CS in imaging system, image fusion, target recognition and tracking. It points out that the CS theory is an effective data processing, and more extensive applications will be come true with the development of the theory.
Traditional Shannon sampling method leads to a large amount of image data, and massive data processing brings a great pressure to bear on the post-processing of image information. Compressed Sensing(CS) theory which can overcome the problem mentioned above is researched in this paper. It can reconstruct a large amount data by sampling small quantity data, and breakthroughs the restriction of Shannon sampling theory. This paper reviews the theory and key technique of CS, and introduces the application and development of CS in imaging system, image fusion, target recognition and tracking. It points out that the CS theory is an effective data processing, and more extensive applications will be come true with the development of the theory.
2011, 4(5): 448-452.
Abstract:
Technological characteristics of opto-electrical payloads for an Unmanned Aerial Vehicle(UAV) are introduced and the configurations of the compact image processor on the UAV are also described. By taking a Digital Signal Processor(DSP) chip as core processor and using Field Programming Gate Array(FPGA) and Micro Controlled Unit(MCU) to acquire and process target data, a compact image processor system is designed. The requirements of volume, weight, power consumption in the design are fully considered as features of airborne equipment. This system has been used in multiplicated opto-electrical payloads. It works stably and reliably and meets task requirements of trapping and locating for UAVs.
Technological characteristics of opto-electrical payloads for an Unmanned Aerial Vehicle(UAV) are introduced and the configurations of the compact image processor on the UAV are also described. By taking a Digital Signal Processor(DSP) chip as core processor and using Field Programming Gate Array(FPGA) and Micro Controlled Unit(MCU) to acquire and process target data, a compact image processor system is designed. The requirements of volume, weight, power consumption in the design are fully considered as features of airborne equipment. This system has been used in multiplicated opto-electrical payloads. It works stably and reliably and meets task requirements of trapping and locating for UAVs.
2011, 4(5): 453-460.
Abstract:
Two kinds of high-speed industrial digital cameras with a resolution of 1 2801 024 and a high frame rate of 300-500 frame/s were designed based on two different models of high frame rate CMOS imaging sensors from two different CMOS sensor manufacturers. The CMOS image sensors are MI-MV13 and LUPA 1300-2. Some experiments of imaging performance evaluation by a laboratorial test were presented. Spectral response and quantum efficiency(QE), gain, photoelectric conversion nonlinearity, dark-current, readout noise, full well electron quantity, dynamic range, etc. were tested and analyzed. Experimental results show that the LUPA 1300-2 is superior to MI-MV13 at quantum efficiency testing, and their peak value QEs are 50% and 12%, respectively, which is basically the same as the reference data of manufacturers. The test results prove that proposed test method is correct and the evalution for two high-speed CMOS image sensors is objective and credible. The high-speed CMOS cameras designed can satisfy the system performance requirements for a high-frame rate.
Two kinds of high-speed industrial digital cameras with a resolution of 1 2801 024 and a high frame rate of 300-500 frame/s were designed based on two different models of high frame rate CMOS imaging sensors from two different CMOS sensor manufacturers. The CMOS image sensors are MI-MV13 and LUPA 1300-2. Some experiments of imaging performance evaluation by a laboratorial test were presented. Spectral response and quantum efficiency(QE), gain, photoelectric conversion nonlinearity, dark-current, readout noise, full well electron quantity, dynamic range, etc. were tested and analyzed. Experimental results show that the LUPA 1300-2 is superior to MI-MV13 at quantum efficiency testing, and their peak value QEs are 50% and 12%, respectively, which is basically the same as the reference data of manufacturers. The test results prove that proposed test method is correct and the evalution for two high-speed CMOS image sensors is objective and credible. The high-speed CMOS cameras designed can satisfy the system performance requirements for a high-frame rate.
2011, 4(5): 461-467.
Abstract:
As the real time properties of image compression can be improved by utilizing a DSP underlying structure, this paper optimizes the Mpeg-4 video encoder on a TMS320C6416 platform by the structure mentioned above. Optimizing methods include complying the assembly functions according to the characteristics of eight parallel execution functional units, optimizing the code and data memory space occupied by the encoder according to the features of DSP two-level cache, and adapting two buffers in memory by using the EDMA cascade characteristics to realize video data coding and transmission at the same time, which improves coding efficiency effectively. The experimental results show that the new encoder can realize encoding of gray scale video of 512 pixel512 pixel at 30 frame/s in real time.
As the real time properties of image compression can be improved by utilizing a DSP underlying structure, this paper optimizes the Mpeg-4 video encoder on a TMS320C6416 platform by the structure mentioned above. Optimizing methods include complying the assembly functions according to the characteristics of eight parallel execution functional units, optimizing the code and data memory space occupied by the encoder according to the features of DSP two-level cache, and adapting two buffers in memory by using the EDMA cascade characteristics to realize video data coding and transmission at the same time, which improves coding efficiency effectively. The experimental results show that the new encoder can realize encoding of gray scale video of 512 pixel512 pixel at 30 frame/s in real time.
2011, 4(5): 468-473.
Abstract:
This article proposes a construction method of vector model based on high frequency information of infrared and visible images, which can be applied to the image alignment between different wave bands(visible and infrared) to the same scene. First, the high frequency characteristic, that is a common characteristic of infrared and visibility images, is analyzed, and it is picked up from a model image. A vector model is constructed by manual interposing. Then, the vector model is performed a 3D transform by the real-time attitude information, and is projected into a 2D graph again, which can be used to carry out the characteristic searching, and achieve the image matching. Finally, matching experiments are performed, and the vector model after transform can complete the image target automatic recognition for different sources. The theoretical analysis and experiment results prove the efficiency and feasibility of this method, and show it popular to all kinds of image alignments that come from different image sources.
This article proposes a construction method of vector model based on high frequency information of infrared and visible images, which can be applied to the image alignment between different wave bands(visible and infrared) to the same scene. First, the high frequency characteristic, that is a common characteristic of infrared and visibility images, is analyzed, and it is picked up from a model image. A vector model is constructed by manual interposing. Then, the vector model is performed a 3D transform by the real-time attitude information, and is projected into a 2D graph again, which can be used to carry out the characteristic searching, and achieve the image matching. Finally, matching experiments are performed, and the vector model after transform can complete the image target automatic recognition for different sources. The theoretical analysis and experiment results prove the efficiency and feasibility of this method, and show it popular to all kinds of image alignments that come from different image sources.
2011, 4(5): 474-479.
Abstract:
As for low contrasts and fuzzy edges of infrared images, an infrared image enhancement method combined with adaptive platform histogram equalization and Laplace transformation is pressented to not only improve the image contrasts but also to sharp the image edges. The algorithm uses double Digital Signal Processors(DSP) to complete parallel processing. Among them, a DSP obtains the image contrast enhancement by using the adaptive platform histogram equalization, and the other one performs Laplace transform algorithm to obtain the edge image from an original image. Finally, the multiplied fusion of two images is finished by coefficient superposition. Experimental results show that this algorithm has an excellent enhancement effect and real-time performance, which satisfies the requirement of 50 Hz processing frequency and improves the image contrast and sharps the image edge.
As for low contrasts and fuzzy edges of infrared images, an infrared image enhancement method combined with adaptive platform histogram equalization and Laplace transformation is pressented to not only improve the image contrasts but also to sharp the image edges. The algorithm uses double Digital Signal Processors(DSP) to complete parallel processing. Among them, a DSP obtains the image contrast enhancement by using the adaptive platform histogram equalization, and the other one performs Laplace transform algorithm to obtain the edge image from an original image. Finally, the multiplied fusion of two images is finished by coefficient superposition. Experimental results show that this algorithm has an excellent enhancement effect and real-time performance, which satisfies the requirement of 50 Hz processing frequency and improves the image contrast and sharps the image edge.
2011, 4(5): 480-488.
Abstract:
As variety illumination, deformation and rotation of targets and many other complicated conditions are always hard to be cracked in target tracking, this paper researches and improves particle filtering and Scale Invariant Feature Transform(SIFT) algorithms. By combined with the two algorithms, it proposes a multi-pattern tracking technique, which adopts a particle filter to predict target position and then chooses SIFT character matching to get the accurate position of an object. This target tracking is quite robust and has applied to an image processor in a artillery wheel type scout car. Experiments has been performed in many aspects and results prove it is accetable and effective.
As variety illumination, deformation and rotation of targets and many other complicated conditions are always hard to be cracked in target tracking, this paper researches and improves particle filtering and Scale Invariant Feature Transform(SIFT) algorithms. By combined with the two algorithms, it proposes a multi-pattern tracking technique, which adopts a particle filter to predict target position and then chooses SIFT character matching to get the accurate position of an object. This target tracking is quite robust and has applied to an image processor in a artillery wheel type scout car. Experiments has been performed in many aspects and results prove it is accetable and effective.
2011, 4(5): 489-496.
Abstract:
A fusion algorithm based on the Nonsubsampled Contourlet Transform(NSCT) is proposed to fuse infrared and visible images. A weighted averaging' scheme based on the physical features of infrared and visible images and a selection principle based on the local energy matching are presented for the low frequency subband coefficients and the high frequency subband coefficients, respectively. Experimental results show that the NSCT has a fast computing ability and can provide more image information due to the concentration energy by image processing. As compared to the image fusion algorithm based on pixels, this algorithm has higher fusion performance, and is a Multiscale Geometric Analysis(MGA) tool more suitable for image fusion.
A fusion algorithm based on the Nonsubsampled Contourlet Transform(NSCT) is proposed to fuse infrared and visible images. A weighted averaging' scheme based on the physical features of infrared and visible images and a selection principle based on the local energy matching are presented for the low frequency subband coefficients and the high frequency subband coefficients, respectively. Experimental results show that the NSCT has a fast computing ability and can provide more image information due to the concentration energy by image processing. As compared to the image fusion algorithm based on pixels, this algorithm has higher fusion performance, and is a Multiscale Geometric Analysis(MGA) tool more suitable for image fusion.
2011, 4(5): 497-502.
Abstract:
In order to study moment invariants of color images, the quaternion is used to process color images and to implement the parallel processing of R, G and B components. Traditional complex moments for graylevel images are introduced to the quaternion, and the quaternion moments for describing a color image are presented. Then, the quaternion affine moment invariants are derived. Experimental results show that the stability of this method is superior to that of L.V.Gool's color affine moment invariants, and the value of /u have be improved by two orders of magnitude. The proposed quaternion moment invariants could be a useful tool in color pattern recognition and tracking.
In order to study moment invariants of color images, the quaternion is used to process color images and to implement the parallel processing of R, G and B components. Traditional complex moments for graylevel images are introduced to the quaternion, and the quaternion moments for describing a color image are presented. Then, the quaternion affine moment invariants are derived. Experimental results show that the stability of this method is superior to that of L.V.Gool's color affine moment invariants, and the value of /u have be improved by two orders of magnitude. The proposed quaternion moment invariants could be a useful tool in color pattern recognition and tracking.
2011, 4(5): 503-508.
Abstract:
Basic principles of dim target detection based on the wavelet transform is analyzed. According to the multi-scale and multi-direction in the wavelet coefficients and combining with the selectable diffusion direction characters, a new method is designed to denoise and detect small targets by the diffusion filter in different directions and different scales of wavelet coefficients, respectively. Combined a anisotropic filter, the algorithm is used to a diffusion filtering experiment for wavelet coefficients. Results show that the algorithm can detect the dim target with a contrast of 2%, and has better performance in the detection of dim target, robustness to Gaussian noise and uneven backgrounds.
Basic principles of dim target detection based on the wavelet transform is analyzed. According to the multi-scale and multi-direction in the wavelet coefficients and combining with the selectable diffusion direction characters, a new method is designed to denoise and detect small targets by the diffusion filter in different directions and different scales of wavelet coefficients, respectively. Combined a anisotropic filter, the algorithm is used to a diffusion filtering experiment for wavelet coefficients. Results show that the algorithm can detect the dim target with a contrast of 2%, and has better performance in the detection of dim target, robustness to Gaussian noise and uneven backgrounds.
2011, 4(5): 509-513.
Abstract:
As traditional methods for calculating Moir fringes can't satisfy the demands of precision and calculating speed, a fast method that can figure out the angle of the Moir fringes accurately is put forward in this paper. According to the principle of the Root Mean Squared Error(RMSE) that the closer a series of data is, the smaller the value of the RMSE will be, the angle of the Moir fringe is come out primarily. Then Mean-Shift algorithm is used to calculate the angle of Moir fringe accurately in the neighborhood of the angle . Experimental result shows that when the contrast level is up to 5.4%, the precision of angle calculation achieves 29according to the evaluation index of angle detection error. The computing speed of the algorithm has been increased largely as compared with those of the traditional methods, and it can achieve 15 ms, and satisfies the demand of real time.
As traditional methods for calculating Moir fringes can't satisfy the demands of precision and calculating speed, a fast method that can figure out the angle of the Moir fringes accurately is put forward in this paper. According to the principle of the Root Mean Squared Error(RMSE) that the closer a series of data is, the smaller the value of the RMSE will be, the angle of the Moir fringe is come out primarily. Then Mean-Shift algorithm is used to calculate the angle of Moir fringe accurately in the neighborhood of the angle . Experimental result shows that when the contrast level is up to 5.4%, the precision of angle calculation achieves 29according to the evaluation index of angle detection error. The computing speed of the algorithm has been increased largely as compared with those of the traditional methods, and it can achieve 15 ms, and satisfies the demand of real time.
2011, 4(5): 514-518.
Abstract:
To modify the image degradation caused by atmospheric turbulence, this paper proposes a new algorithm based on the iterative blind deconvolution with weighted prediction to solve the problems on instability convergence and huge complexity from traditional blind deconvolution methods. By optimizing an existing iterative blind deconvolution L-R algorithm, the proposed algorithm uses the weighting to obtain prediced values at the end of every iterative step, it then calculates the acceleration operators according to the predicted values to improve its convergence speed. Experiments show that the algorithm is capable of restoring the turbulence degraded image and the convergence have speeded about 43.8 times as compared with that of L-R algorithm. The algorithm's fast convergence shows its great practical value.
To modify the image degradation caused by atmospheric turbulence, this paper proposes a new algorithm based on the iterative blind deconvolution with weighted prediction to solve the problems on instability convergence and huge complexity from traditional blind deconvolution methods. By optimizing an existing iterative blind deconvolution L-R algorithm, the proposed algorithm uses the weighting to obtain prediced values at the end of every iterative step, it then calculates the acceleration operators according to the predicted values to improve its convergence speed. Experiments show that the algorithm is capable of restoring the turbulence degraded image and the convergence have speeded about 43.8 times as compared with that of L-R algorithm. The algorithm's fast convergence shows its great practical value.
2011, 4(5): 519-524.
Abstract:
As for the image stabilization of frame and frame under a complicated movement background, we present a motion estimation algorithm with motion vector modified by gray projection. First, we separately calculate grayscale projection sequences on the row and column of a current frame and a reference frame. Second, the projection sequences of current frame are divided into blocks. Each piece of block respectively performs a cross-relation calculation with the projection curves of rows and columns of the reference frame. Then, we can obtain the motion vector sets of rows and columns based on local projection. Finally, by using the confidence degree of block area as calculation parameters of weight coefficient, we calculate the weight of each piece of pixel displacement by the parameters to calculate the weighted motion vector in one direction. The experimental results show that this method can make the impact of moving target only on the one of the several local blocks and the other blocks are not affected. It ensures the best possible image stabilization accuracy. The Root-Mean-Square Error(RMSE) values obtained by image stabilization and reference images have decreased significantly, which shows the image stabilization by this method is more consistent with that of the reference image.
As for the image stabilization of frame and frame under a complicated movement background, we present a motion estimation algorithm with motion vector modified by gray projection. First, we separately calculate grayscale projection sequences on the row and column of a current frame and a reference frame. Second, the projection sequences of current frame are divided into blocks. Each piece of block respectively performs a cross-relation calculation with the projection curves of rows and columns of the reference frame. Then, we can obtain the motion vector sets of rows and columns based on local projection. Finally, by using the confidence degree of block area as calculation parameters of weight coefficient, we calculate the weight of each piece of pixel displacement by the parameters to calculate the weighted motion vector in one direction. The experimental results show that this method can make the impact of moving target only on the one of the several local blocks and the other blocks are not affected. It ensures the best possible image stabilization accuracy. The Root-Mean-Square Error(RMSE) values obtained by image stabilization and reference images have decreased significantly, which shows the image stabilization by this method is more consistent with that of the reference image.
2011, 4(5): 525-530.
Abstract:
In order to recognize plane surface scenes and buildings in a collection image, the relationship of the building shape in a real world and the pixel numbers of building imaging in the collection image is analyzed by a geometry way. Firstly, the distortion of the plane surface imaging is computed, then the building shape imaging is analyzed. Finally, the relationship of the real world scene and the imaging scene is obtained. Experimental results indicate that the plane surface scene and building imaging information can be deduced by camera states and parameters, and the information can do great help for the recognition and tracking of targets. This method is suitable for three-dimensional automatic target acquisition and tracking, like typical buildings, isolated buildings, and buildings with obvious shape features in building groups.
In order to recognize plane surface scenes and buildings in a collection image, the relationship of the building shape in a real world and the pixel numbers of building imaging in the collection image is analyzed by a geometry way. Firstly, the distortion of the plane surface imaging is computed, then the building shape imaging is analyzed. Finally, the relationship of the real world scene and the imaging scene is obtained. Experimental results indicate that the plane surface scene and building imaging information can be deduced by camera states and parameters, and the information can do great help for the recognition and tracking of targets. This method is suitable for three-dimensional automatic target acquisition and tracking, like typical buildings, isolated buildings, and buildings with obvious shape features in building groups.
2011, 4(5): 531-536.
Abstract:
The local variance distribution of a gray level image is taken as an important characteristic to express image structural information, and the Singular Value Decomposition(SVD) is performed on a local variance distribution matrix. The angle between the singular vectors of the reference image and distorted image is used to measure the structural similarity of the two images, and then the image quality assessment is achieved. Experimental result shows that the local variance distribution can emphasize the structural information. It is better consistent with human visual perception characteristics and the assessment results are superior to those from Mean Square Error(MSE), Peak Signal to Noise(PSNR), Structure Similarity(SSIM) and SVD methods based on pixel value distribution.
The local variance distribution of a gray level image is taken as an important characteristic to express image structural information, and the Singular Value Decomposition(SVD) is performed on a local variance distribution matrix. The angle between the singular vectors of the reference image and distorted image is used to measure the structural similarity of the two images, and then the image quality assessment is achieved. Experimental result shows that the local variance distribution can emphasize the structural information. It is better consistent with human visual perception characteristics and the assessment results are superior to those from Mean Square Error(MSE), Peak Signal to Noise(PSNR), Structure Similarity(SSIM) and SVD methods based on pixel value distribution.
