Abstract:Pose information plays an important role in aviation, navigation, indoor robot positioning, and other fields. An increasing number of studies have shown that polarization is the key for some of the biometric navigation and vision applications in low light. In this paper, a method based on polarization imaging was proposed to understand the object pose under the environment of low light and strong background noise. Non-polarized light was converted to polarized light by setting the polarization film in front of an ordinary light source. A polarization camera was used to obtain the images of a polarized light source. Further, computer vision technology was used to identify the polarized light source and determine the pose of the light source relative to the camera. The results show that the polarization method proposed in this paper has better performance and robustness than light intensity. At a distance of 40 m, the pose error is 2.99%. This method uses polarization imaging to determine the pose of an object relative to the light source under the environment of low light and strong background noise. Thus, this study overcomes the problem that computer vision cannot estimate pose information of objects under low light environment.
Abstract:The measurement accuracy of a dual-wavelength interference microscope (DWIM) is ensured by the two optical wavelengths, which must be calibrated to be traceable in length. A calibration method was proposed for DWIMs that can determine both the relationships between two wavelengths and between the wavelengths and the length reference. First, the relationship between the reconstructed surface and wavelengths was analyzed theoretically, where the surface jump and height place sufficient constraints on the two wavelengths. Then, based on the analysis, a dual-constraint calibration method was proposed to determine the wavelengths. Together with this method, by measuring a flat and a height standard step, two constraints on the wavelength ratio and arithmetic value were established. With the constraints, the optical wavelengths can be calculated by solving equations. Simulations demonstrate a wavelength calibration accuracy of 0.1 nm with the proposed method. A DWIM calibrated with the method achieves measurement accuracy up to 0.5 nm by measuring the depth of a micro-groove. The calibration only needs a height step standard. The proposed method provides an accurate, but simple, calibration for a DWIM and can be applied to dual-wavelength digital holography and three-wavelength interferometry as well.
Abstract:To solve the problem of high correlation of parameters that restricts the reliability of the estimated parameters in the traditional self-calibration methods for pulse-terrestrial laser scanner, a novel self-calibration method is proposed based on the least square collocation method. An error model for total station has also been proposed. Using the calibration method, the calibration network benchmark can be flexibly defined by configuring the relevant prior weight information for the estimated parameters, and all the parameters are considered as observation values in the self-calibration adjustment. Additionally, by designing appropriate calibration network configuration and adopting the robust estimate with equivalent weight matrix to perform gross error detection and variance component estimation, this method effectively reduced the correlation between the estimated parameters and improved the reliability of self-calibration. Finally, the simulation results of Leica scan station showed that the proposed method can achieve low correlation and high-precision estimation of most parameters. The only significant correlation was between the parameters of collimation error and scanner azimuth (approximately 0.8; however, this correlation was effectively reduced to 0.5) in the four-station scanning design.
Abstract:Splitters for metasurface arrays have the characteristics of high transmission, low loss, and high integration; these have become a current research hotspot for nano structure applications. This paper proposes a metasurface infrared splitter array composed of Si and SiO2. Infrared with different wavelengths was used to enter the unit, and only the radius of the cylinder was changed. The transmittance and phase of the infrared with different wavelengths were calculated based on the radius. According to the generalized Snell's law, several units with different cylindrical radii were selected to form the splitter period of the metasurface. The period had a corresponding transmission phase gradient for the infrared with different wavelengths. It was verified that the metasurface arrays had a light splitting function. Possible applications of infrared splitter arrays on the metasurface were proposed. The results reveal that the metasurface period composed of three Si nanopillars, with radii of 150, 250, and 300 nm, can separate the mixed incident infrared with wavelengths of 2.7 μm and 2.9 μm, and the transmittance can become more than 70%.
Abstract:To design a tilted beam splitter type autocollimator, a method of correcting the on-axis astigmatism introduced by the plate beam splitter is proposed herein. The reason behind the astigmatism of the points on the axis is analyzed, and the method of correcting this astigmatism using a cylindrical lens is studied. A method of calculating the correction parameters for the cylindrical lens is developed. To tackle the residual chromatic aberration problem of the polychromatic autocollimator system, the correction method of combining the cylindrical beam splitter and cylindrical lens is proposed and compared with the method of combining the plate beam splitter and double cylindrical lens. The former is found to be better through a comparative analysis. The number of components can be reduced, and is more suitable for use in a tilted beam splitter type multicolor autocollimator. The design results indicate that the RMS value of the wave aberration of the system combining the cylindrical beam splitter and cylindrical lens is 1/26 waves, and the PV value is 1/5 waves. The system shows good astigmatism and achromatic effects. The image quality is effectively restored. Detecting the wavefront of the prototype, the RMS value is 1/25 waves, and the PV value is 5/26 waves, which is less than a quarter of the wave. The system imaging is good, verifying the correctness and feasibility of the design.
Abstract:To improve point measurement accuracy in ultra-high-precision and large-scale 3D control network layouts, a high-precision 3D control network is established using a laser tracker measurement system. The precision of point intersection is affected by the spatial geometric distribution of the measuring points. A weighted geometric precision factor model based on the laser tracker measurement system is constructed. The precision of multi-directional and multi-distance spatial point intersection is compared and analyzed from the distribution and number of measurement points, from the horizontal plane and elevation .The experimental results show that the shorter the measuring point distance, the greater the deviation between the zenith distance, and the higher the precision of point intersection when measuring at a three-point intersection . When adding a measuring point inside a known tetrahedron, the shorter the observation distance of the measuring point, the greater the zenith distance, and the higher the precision of point intersection; When the number of known points is 6, the point intersection accuracy can meet the measurement requirements, and the precision tends to be stable. With an increase in the number of measuring points, the precision improves slowly. This method can select favorable positions of points, effectively improve the accuracy of point rendezvous, and has high application value in practical engineering applications.
Keywords:laser tracker;multipoint rendezvous;precision factor;spatial geometric distribution
Abstract:To solve the problem of unsuitability of the traditional laser stripe centerline extraction method for translucent objects, this study examines the extraction algorithm of laser stripe center on translucent strip surface to realize accurate measurement of the contour of the translucent strip. First, a comparative analysis of the laser stripes imaging mechanism of opaque and translucent glue strip was performed. Then, the one-dimensional gray-scale sequence of the translucent glue strip laser stripe was analyzed, and the difference in the peak distribution of the direct reflection component and the subsurface scattering component was elucidated, based on which the algorithm for extracting the laser stripe center of the translucent glue surface was proposed. The one-dimensional gray-scale sequence of the laser stripe was scanned and then the direct reflection component was extracted using mathematical morphology based on the value and distribution of the gray value peak. Finally, the gray barycenter method was used to windowing the direct reflection components in the gray sequence to extract the sub-pixel center position coordinates; the laser stripe center could be obtained via continuous scanning. Experimental results show that the average error of the laser stripe center extraction method reaches 3.197 pixels, and the average error after quadratic curve fitting is 0.714 pixels, which is better than the gray barycenter method without pretreatment. The proposed method can effectively extract the laser stripe center of translucent glue strip surface, which can help in laying the foundation for the contour measurement of translucent glue strip in the industrial field.
Abstract:A non-contact acoustic resonance method for landmine detection was studied by a laser self-mixing interference vibration measurement technique. Based on the principle of the laser self-mixing vibration measurement technique, taking a type 69 anti-tank plastic landmine, a type 72 anti-tank metal landmine, and a type 58 anti-personnel rubber landmine and brick as testing targets, an experimental system was built to measure their acoustic characteristics. Experimental tests were conducted for a comparative analysis. An experimental system for detecting buried landmines was designed and built, and the laser self-mixing vibration measurement technique was used to study the acoustic vibration characteristics of different types of buried landmines and those of the same landmine under different buried depths and soil conditions. The experimental results show that landmines exhibit unique multi-modal vibration characteristics under external excitations. Buried landmines exhibit different acoustic vibration characteristics under different buried conditions. The vibration intensity increases with the increase of buried depth and decrease of soil porosity, thus is not sensitive to soil humidity. These results indicate that the laser self-mixing vibration measurement technique can be used for the non-contact detection of acoustic vibration characteristics of landmines. The acoustic vibration characteristics of buried landmines are different from other objects such as bricks and are related to the type of landmine and the buried conditions.
Abstract:Rapid detection of sulfamethazine (SM2) and ofloxacin (OFL) residues in chicken was achieved through synchronous fluorescence technology coupled with chemometric methods. First, the three-dimensional synchronous fluorescence spectra of an SM2 standard solution, an OFL standard solution, chicken extract without antibiotics, and chicken extract containing SM2 and OFL residues were analyzed. The wavelength difference (Δλ) of SM2 and OFL were determined to be 150 nm and 210 nm, respectively. The fluorescence excitation peaks of SM2 and OFL were determined to be 292.5 nm and 295 nm, respectively, for the detection of SM2 and OFL residues in chicken. Subsequently, the effects of addition amounts of β-mercaptoethanol solution and o-phthalaldehyde solution, as well as time, on the fluorescence intensities were investigated through the single factor test. A β-mercaptoethanol solution of 300 μL, an o-phthalaldehyde solution of 25 μL and a time of 44 min were the best combination of conditions for detections. Finally, the prediction models were established by the peak height and peak area algorithms respectively. The experimental results show that the prediction model based on the peak height algorithm provides a better comprehensive evaluation than the prediction model based on the peak area algorithm. For the prediction models of SM2 and OFL residues based on the peak height algorithm, the coefficients of determination for the prediction set (R2P) are 0.897 3 and 0.997 3 for SM2 and OFL, respectively. The root mean square errors for the prediction set (RMSEP) are 6.060 5 mg/kg and 0.539 2 mg/kg, for SM2 and OFL, respectively. The recovery rates of SM2 and OFL are in the range of 76.1%-115.2% and 96.7%-110.1%, respectively, and the relative standard deviations (RSDs) are in the range of 2.7%-7.0% and 2.8%-10.0%, respectively. The proposed method is fast and simple and could achieve rapid detection of SM2 and OFL residues in chicken.
Abstract:In order to solve the problem of contamination damage of large-aperture lens in final optical assembles and realize the online cleaning of lens surface, a new method of contaminant removal was proposed, which used the air knife technology to realize the online removal of particle contaminants. The flushing process of particles on the lens surface was simulated by finite element method. The flow field distribution and particle trajectory in the lens module were studied under laminar air and air knife. The results show that when the laminar wind velocity is 0.3 m/s, a stable flow field can be formed inside the lens module. When the distance between the air knife and the upper surface of the lens is 78.73 mm, the stress concentration of the air knife on the lens surface can be effectively reduced. At the same time, when the air knife has a wind speed of 30 m/s, the air knife can effectively isolate silica particles with an exit velocity less than 50 m/s and a particle diameter less than 75 μm. The installation position of the air knife on the lens surface and the distribution characteristics of the air flow velocity on the lens surface are verified by experiments. The sputtering particle contaminants of mechanical components are more difficult to remove than those of optical elements. Finally, the effective technical parameters of the air knife on the surface of the lens are obtained, which provides a new method for the on-line removal of particle contaminants of the large-aperture lens module.
Keywords:laser inertial confinement fusion;large-aperture lens;air knife;particle contaminant;flow field simulation
Abstract:To investigate generation and evolution mechanisms of “comet-tail defects” on surfaces of components after magnetorheological finishing (MRF), a self-developed MRF machine was used to analyze the effects of concentration and moisture content of MR fluid particles on comet-tail defects in MRF, using fused silica glass as a sample. Extant studies show that comet-tail defects are generated following initial defects on component surfaces. Moreover, the accumulation of particles at these initial defects is the main reason for their evolution into comet-tail defects. Increases in the concentration of polishing particles lead to evident accumulation of particles at initial defects, and more comet-tail defects emerge resultantly. However, fluidity is strengthened with increases in the moisture content, thereby resulting in a decreased number of comet-tail defects. With increases in processing times, initial craters evolve into comet-tail defects At this point, the crater depth at the head of the comet-tail initially increases and then decreases when compared to initial defects. Similarly, increases in processing times monotonically increase the crater edge angle until the defects are completely removed. Hence, the results of this study establish a theoretical foundation for restricting the formation of comet-tail defects on surfaces of components in MRF and facilitate subsequent development of MR fluids and process optimization methods to control the generation of comet-tail defects.
Abstract:Owing to the Earth's rotation, image rotation always appears on the primary focus detector during the long exposure of Alt-Az telescopes with large aperture and wide field of view. This implies that except for the central point, other stars rotate around the center of the field of view; thus the target cannot be accurately identified and observed. To eliminate the image rotation, the de-rotator tracking error of a telescope with a large aperture and wide field of view should be less than 5″. The position and velocity of the image rotation of the primary-focus detector were analyzed and verified based on the star rotation principle. Subsequently, a dual-motor driving de-rotator system was designed, where the clearance of the mechanism was removed by the recombination current command. The experimental results indicate that the image rotation characteristics are consistent with the expression of image rotation, and the measurement error is less than 2%. The dual-motor diving de-rotator system can improve the stability and accuracy of the tracking system; thus the tracking accuracy of the de-rotator mechanism can be less than 1″ both at a low velocity of 15 (″)/s and at a high velocity of 170 (″)/s. In addition, the tracking error remains less than 1 arcsec when the de-rotator operated at a variable and high velocity of 7.86°sin(2π×0.5t).
Keywords:telescope;dual motor anti-backlash;large caliber and field of view;de-rotator control;Alt-azimuth telescope
Abstract:In this study, a method to improve robot positioning accuracy was proposed, given that the accuracy performance of industrial robots still does not satisfy the requirements of high-end manufacturing. First, a kinematics error model based on pose differential transformation and kinematics error model based on coordinate error transformation were summarized. Second, a kinematic parameters identification method based on BAS-PSO algorithm was proposed. Finally, accuracy characteristics of different error models were compared and analyzed via experiments. The experimental results indicate that the average comprehensive position/attitude error of the TX60 robot, after identification by the proposed algorithm, decreases from (0.312 mm, 0.221°) to (0.093 8 mm, 0.044 2°). The average position error and average attitude error of the robot after direct identification based on the forward kinematics model correspond to 0.097 5 mm and 0.098 6°, respectively. The BAS-PSO algorithm proposed in the study displays good performance in terms of identification accuracy and convergence speed. Furthermore, robot kinematic parameters directly identified by the forward kinematics model exhibit better identification stability and accuracy.
Abstract:As the micro-vibrations generated by the mechanical shutter (with a large aperture) of a survey camera decreases the imaging resolution of an optical telescope also gets suppressed. Static and dynamic balancing methods cannot be used due to limitations in mass and volume of the shutter. The passive vibration isolation method cannot simultaneously meet the requirements of isolation efficiency and structural stability due to the low operating frequency of the shutter, which is only 0.77 Hz. Therefore, an arrangement in which two partially open shutter blades are facing each other was proposed. The excitation force of the shutters was eliminated in one direction (y-direction) but is doubled in the other (z-direction). Based on the characteristics of partially open shutters, a method was proposed in which the excitation force along the z-direction was suppressed by optimizing the driving curve. The excitation force of the shutter was modeled mathematically and its absolute value along the z-direction was minimized by optimizing the driving curve of the shutter using simplex method. Simulation test results show that the medium- and high-frequency components of the micro-vibrations are suppressed. The amplitude of the excitation force along the y-direction is reduced by 97.5% (from 0.73 N to 0.018 N). The excitation force along the z-direction is reduced by 41% (from 2.92 N to 1.72 N). The results of finite element analysis verify the correctness of the suppression method, which can be applied to the design and optimization of other similar rotating mechanisms.
Abstract:When the Newton-Raphson algorithm is used to solve the inverse kinematics of a 6-degree-of-freedom parallel Hexapod platform with axis offset Hooke joints, the high-precision calculation requirements can decrease the response speed of the platform. To solve this issue, a new method based on Brent’s method is proposed in this study. First, the kinematics model based on the space circle auxiliary model is established by considering the geometric characteristics of the platform and motion characteristics of the axis offset Hooke joint structure. Subsequently, an inverse kinematics algorithm based on the space circle distance solution method is proposed, along with a compensation scheme for the derivative motion of the screw–nut pair. Subsequently, Brent’s method to determine roots is introduced and applied to the inverse kinematics algorithm. Finally, the experiment is used to verify the necessity of the derivative motion error compensation scheme, compare it with the inverse kinematics based on the Newton–Raphson algorithm, and test the response speed and accuracy of the Hexapod platform controlled by the proposed algorithm. In the experimental results, a comparison of the inverse kinematics based on Brent’s method with those based on the Newton-Raphson algorithm indicates that the comprehensive response rate increases by approximately 0.5 s under the operating requirements of repeatability of less than 10 μm and 5". Therefore, the inverse kinematics method based on Brent’s method improves the response speed of the Hexapod platform.
Abstract:The installation error of a wave generator is a common source of error in deformation measurement of a flexspline. This installation error leads to a large deviation in the functional relationship between the deformation of the flexspline and rotation angle. In this study, we attempt to solve the problem of the installation error between the center of the wave generator and center of rotation by analyzing the measurement error of the flexspline and proposing a compensation method for the radial deformation of the flexspline. First, based on the principle of coordinate change, a mathematical model of eccentricity error characterization is established to obtain the installation eccentricity and actual structural parameters of the wave generator. Second, based on the eccentric radial deformation function of the wave generator, a correction model for the radial deformation error of the flexspline is developed to correct and compensate for the actual radial deformation function of the flexspline under eccentric installation conditions. The experimental results indicate that the peak-to-valley deviation of the flexspline deformation function corresponds to 0.134 mm under the effect of installation error, which significantly differs from the theoretical value. Following correction and compensation using the method proposed in the study, the error obtained is approximately 0.012 mm, and the radial deformation of the flexspline differs from the theoretical value. The trend is essentially identical. The method can effectively improve measurement accuracy of the radial deformation of the flexspline and establish an experimental and theoretical foundation to optimize tooth profile parameters of the harmonic reducer.
Keywords:harmonic drive;radial deformation of the flexspline;installation eccentricity;error compensation;wave generator
Abstract:This study aimed to solve the problems of ambient light interference and speckle loss during small-format measurement of the weld seam of the steel plate under arc welding. For this, a non-contact deformation measurement method is proposed based on the principles of yellow light narrow-pass filter, an improved digital image correlation method (DIC), and Gaussian low-pass filter technology. The method uses the DIC optical measurement method to address the limitations of the contact measurement equipment, that it cannot work under high temperatures. Thus, matte gray high temperature glue and black high temperature were used paint to ensure effective high temperature speckle resistance. To reduce the interference of ambient light, a narrow-pass filter was installed, which filtered out yellow light. Further, two optical compensation coefficients were added to the DIC matching formula, and Gaussian low-pass filtering technology was used to denoise high frequencies in strong light areas. The incremental interpolation test was used to obtain the best filter cutoff frequency. A constant displacement comparison experiment of DIC was performed on the filtered image, and it was verified that Gaussian low-pass filtering can effectively improve the matching accuracy of DIC. The results of the proposed measurement method were compared by using two sets of experiments. First, they were compared with those of ordinary processing; the matching process was stable and had a better cloud image quality. Second, they were compared with those of large-format experiments with verified accuracy; the measurement accuracy of the selected two points was higher than 1%, which is a high matching accuracy. The maximum displacement of the center line of the weld was 1.23 mm. It can be concluded that the proposed method can effectively measure and match the high-temperature weld image with small format.
Abstract:A lightweight convolutional neural network is proposed to realize driver pose estimation. This architecture is validated using a new data set containing video clips of 26 drivers gathered for this study. Firstly, by mathematical modeling, the driver pose estimation task can be reformulated as seeking a mapping function between a confidence map of ground-truth joint labels and a confidence map of predicted value when the loss function is minimized. To develop a fully convolutional neural network, we used hourglass as the backbone model for each stage, the residual block as the basic unit, and employed batch normalization and activation functions. We performed feature aggregation using the features from the preview stage as the input for the current stage, and features from different stages were aggregated to obtain both local detailed information and global context information. At each stage, the computed beliefs provided an increasingly refined estimate for the location of each part. The use of multiple loss functions allowed the network model to learn more detailed and accurate representations. A comparative experiment was performed to verify the feasibility of the model, which shows that the cascaded network structure and multi-task learning strategy improve the prediction accuracy of the model by 3.84%. These extensive experiments demonstrate that the proposed architecture can be executed quickly at a low computational of 0.7 M model parameter number and has an average prediction accuracy of 95.74%. The number of calculation and parameters of the model are much lower than those of other human pose estimation models, and the real-time detection of driving posture can be accomplished on the vehicle side.
Abstract:To improve the low efficiency of video data utilization and the weak target recognition ability of optical measuring equipment, we propose a method to establish a database with massive amounts of video and then construct an infrared target recognition system. Firstly, a fast infrared target detector to extract the target region from video frames and establish a database by classifying these subimages is designed. Secondly, according to a specific task, a cluster of convolutional neural networks with different structures is designed, and a selection strategy based on a mean value statistical analysis of test accuracy and parameter scale is designated. Consequently, we obtain a simple network with good generalization ability and a reasonable number of training epochs. Finally, the optimized model and its parameters are loaded as a classifier, which is combined with the detector to perform real-time detection and classification of infrared target events. Simulation results show that the average target classification accuracy can exceed 95% and the rate is approximately 50 FPS. The design scheme and selection strategy from the convolutional neural network structure is effective, and the constructed system exhibits real-time infrared target recognition while meeting accuracy requirements.
Abstract:A high-resolution non-destructive magneto-optical imaging restoration method based on slip-induced magnetic anisotropy and magneto-optical effect has been proposed in this study. The study investigated the change law of magnetic field distribution due to slip-induced magnetic anisotropy which is caused by plastic deformation of ferromagnetic materials. A model of damaged characters recovery system based on magneto-optical effect was developed, and the factors influencing magneto-optical imaging resolution were analyzed. The experimental system was built in accordance with the damaged character recovery system. The study also accomplished the recovery recognition of damaged characters and the gray analysis of the restored image. The results of the experiment showed that as the brightness of the image decreases, the angle φ between the polarizer and the analyzer increases, while the resolution first increases and then decreases. The imaging results were best when φ was approximately 75°. The embossed characters provide the clearest image recovery among the various methods of character formation. The results also showed that the maximum gray mutation and gray gradient decrease as the depth of the character damage increases. Due to this reason, the resolution decreases, and the characters become more difficult to recover and identify. The maximum damage depth for effective recovery for 0.4 mm embossed characters is 0.6 mm. This study demonstrates the validity of the slip-induced magnetic anisotropy theory and provides a theoretical foundation for developing an efficient portable and high-resolution non-destructive character recovery recognition device.
Keywords:damaged characters;magneto-optical imaging;image restoration;slip-induced magnetic anisotropy
Abstract:Infrared small target detection has been widely used for early warning, guidance, and in other fields of national defense. However, infrared small target occupies less pixels and lacks shape and texture features, which makes detection of infrared dim and small targets a challenging task. In this study, a simple and efficient real-time infrared small target detection network has been proposed. The detection network uses the adaptive receptive field fusion module to increase the contextual information around the targets. In addition, we used spatial attention mechanism to determine the relationship among different regions, which can strengthen the correlation and compactness among different regions. To improve the ability of network for locating the target and estimating the positive and negative samples, GIOU Loss and Focal Loss were used to design the loss function. The experiments were conducted on three infrared small target sequences and a single frame image set. The proposed network achieved 91.62%, 71.54%, 81.77% and 90.67% AP, respectively, and maintained high detection speed at approximately 165 FPS. The experimental results showed that the proposed infrared dim and small target detection network has good detection performance on the infrared small targets with complex background and low SNR.
Keywords:infrared small target detection;attention mechanism;convolutional neural network;deep learning
Abstract:To fully exploit the dependency of information captured from transmission map estimation and image dehazing, this paper proposes a dual vision attention network for jointly estimating transmission map and clear image from hazy image. The network consists of an image dehazing layer and a transmission estimation layer, with each layer containing a recurrent attentive network and an encoder-decoder network. The haze attention map is generated in the image dehazing layer under the supervision of the transmission map in the recurrent attentive network and guides the subsequent encoder-decoder network to estimate the dehazing result. Similarly, with clear image supervision, the recurrent attentive network in the transmission estimation layer generates the scene attention map and guides the subsequent encoder-decoder network to predict the transmission map. Based on this architecture, we adopted the recurrent units in the two recurrent attentive networks to exchange hidden information. The haze concentration information can then be used to estimate the scene attention map, and the scene information can be used to predict the transmission map. The experimental results demonstrate that our method not only achieves a good dehazing effect on both synthetic and real images, but also outperforms the existing methods in terms of quality and quantity. The average processing time for a single hazy image is 0.043 s. Therefore, our method can be used in the engineering practice of image dehazing.
Abstract:Computational imaging is a new interdisciplinary subject that has gained widespread research attention in recent years. However, its efficiency and recovery effect restrict its development in engineering applications. In this paper, an efficient image deblurring method based on a region selection network is proposed to tackle the restoration task in the fields of wavefront coding imaging and single lens computational imaging. In contrast to traditional image restoration methods, which usually involve construction of an objective function and addition of reasonable image priors to restore blurry images, the proposed method is based on a combination of a deep learning method and a traditional restoration algorithm. The traditional method is used for the main image restoration process, while the deep learning method is used to intervene in the kernel estimation region selection. The deep learning method involves constructing and training a deep binary classification network, which can automatically eliminate the flat overexposure, short texture, and other areas in the global image, and select the most suitable block area for kernel estimation. On this basis, traditional restoration methods perform kernel estimation, non-blind image restoration, and image enhancement processing. The experimental results show that the proposed method can achieve a good and stable restoration effect, that the proposed region selection method can reduce the computational complexity, and that the point spread function can be estimated well. When the error rate is limited to 1.5, the restoration success rate is improved by at least 2.1%, and the average peak signal-to-noise ratio (PSNR) is increased by at least 0.5 dB.
Abstract:Given the low accuracy of delineation of the infiltration area of breast cancer lesions in DCE-MRI, variable structure and shape, large intensity heterogeneity changes, and low boundary contrast, the automatic segmentation of breast cancer lesions has the problems of low accuracy and mis-segmentation. For this reason, a two-stage breast cancer image segmentation framework is constructed, and a breast cancer lesion segmentation model UTB-net is proposed to integrate multi-scale and non-local at the encoding path and the end, respectively, which constructs attention-residuals in the decoding module. First, the benchmark U-net network model is used to achieve a rough delineation of the breast area, eliminating the influence of unrelated tissues, such as chest muscle, fat, and the heart, on breast tumor segmentation in the image. Then, based on the extracted ROI results, a multi-scale information fused and non-local module is constructed in the coding path of the model. Finally, an attention-residual hybrid decoding module is constructed in the decoding path, and a deep supervision mechanism is introduced to improve the segmentation accuracy of breast tumor lesions. Experimental results show that breast tumor segmentation indexes DICE, IOU, SEN, and PPV increase by 4%, 4.78%, 5.92%, and 3.94% respectively, in comparison with the U-Net benchmark model. The proposed model not only improves the segmentation results of breast cancer but also reduces the small-area mis-segmentation and calcification segmentation.
Keywords:Dynamic Contrast-Enhanced Magnetic Resonance Imaging(DCE-MRI);segmentation;breast tumor;U-net;attention;residual network
Abstract:The lossless compression algorithms for hyperspectral images based on the recursive least square filter have accurate prediction and fast convergence. However, the traditional compression algorithm is incapable of quickly identifying the optimal prediction length of each band, and the performance of the algorithm needs to be improved. In this paper, we propose a hyperspectral image compression algorithm based on the recursive least square lattice filter group. Considering the spatial correlation of the hyperspectral image, the proposed scheme first uses a single-sided Gaussian predictor for inner-band prediction. The lattice filter group is then used to calculate the optimal filter for each spectral band to obtain the prediction error. Further, the screening process of the optimal filters are simplified according to the characteristics of the chain sequence update of the lattice filter group; this significantly reduces the computational complexity. Finally, the arithmetic code is used to encode the residual data. Experiments on the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) 2006 hyperspectral image datasets reveal that the average compression results of the algorithm for 16-bit calibrated and uncalibrated images are 3.34 bits/pixel and 5.61 bits/pixel, respectively. The proposed algorithm therefore shows a competitive performance with regard to compression and requires less computation time compared to other outstanding algorithms.
Keywords:hyperspectral imagery;lossless compression;recursive least squares;Gaussian predictor;lattice filter group
Abstract:A novel neural network model based on "fusion gate" is proposed herein to improve the prediction performance of image captioning. The "fusion gate" network, which is based on the Encoder–Decoder structure, is a combination of the convolutional neural network and the recurrent neural network. The VGGNet-16 network was used during training to convolute the input image to generate 4096 output vectors. First, the output vectors were combined with the labeled statement to calculate the results, and these results were used as the input of the "fusion gate" network. Finally, the prediction result was obtained through the "fusion gate" network calculation, the preceding process was iterated successively according to time steps, and the network training was completed. Overall, the experimental results demonstrate that the CIDEr value is 10.56% higher than that of “Neural Talk” and other related evaluation indexes are significantly improved. The "fusion gate" neural network model not only has a high prediction index, but also involves 21.1% less parameters relative to the “Attention model” network. Moreover, the algorithm requires fewer computer resources and has a lower operating cost. This structure also makes it possible to deploy the neural network in edge computation processors, which plays a key role in promoting the wide range of image captioning applications.
Keywords:image captioning;Long Short Term Memory (LSTM);convolutional neural network;attention model
Abstract:Vessel detection has always been a popular research topic in fields such as remote sensing image processing, pattern recognition, and computer vision. As vessels are maritime transport carriers and important military objects, vessel detection plays an important role in a spectrum of related military and civil fields. It has broad application prospects and high practical significance. This paper provides an overview of the existing literature on vessel detection using optical satellite imagery. The development of optical satellites for vessel detection on the sea surface is reviewed. The physical characteristics of vessels in optical remote sensing imaging are analyzed. The global research status of vessel detection technology using optical remote sensing imaging is summarized. The related theories and key technologies pertaining to the target detection model and architecture are analyzed and compared in detail. The problems and challenges of vessel detection methods using optical remote sensing images are discussed. Aiming at the urgent demand for good performance and robustness of the algorithm in practical applications, some crucial problems that need to be solved are proposed. The development trend of future research on vessel detection using optical remote sensing images is discussed.
Keywords:remote sensing image;image processing;vessel detection;candidate region extraction;image dataset