2024, Volume 46, Issue 2
2024, 46(2): 119-128.
Polarization image fusion improves overall image quality by combining spectral and polarization information. It is used in different fields, such as image enhancement, spatial remote sensing, target identification and military defense. In this study, based on a review of traditional fusion methods using multi-scale transform, sparse representation, pseudo-coloration, etc. we focus on the current research status of polarization image fusion methods based on deep learning. First, the research progress of polarization image fusion based on convolutional neural networks and generative adversarial networks is presented. Next, related applications in target detection, semantic segmentation, image defogging, and three-dimensional reconstruction are described. Some publicly available high-quality polarization image datasets are collated. Finally, an outlook on future research is presented.
2024, 46(2): 129-137.
Ultraviolet image intensifiers are imaging devices that are sensitive to ultraviolet radiation. Defects in the field of view are the main factors restricting the imaging effect of ultraviolet image intensifiers. Currently, the field-of-view defect detection technology is mainly divided into artificial and machine vision. This paper explains the definitions and detection standards for field defects. Subsequently, the difficulties in field defect detection are analyzed from the perspectives of defect-overlapping proximity, size, and quantity. Next, the research status of the field-of-view defect detection technology of ultraviolet image intensifiers is introduced. Combined with the current detection requirements and deficiencies, the defect detection effect of deep-learning technology in other fields was investigated. Finally, a theoretical feasibility analysis is presented, and the concept of field defect detection based on deep learning is proposed. The purpose is to provide a new solution for field defect detection of ultraviolet image intensifiers and promote their development in a practical and intelligent direction.
2024, 46(2): 138-143.
In spacecraft thermal control technology, infrared heating cages are used to simulate the external heat flow reaching each surface; however, the rationality and accuracy of this method must be re-evaluated when the feature size of the simulated surface decreases progressively. In this study, simulation analysis and design optimization of infrared heating cages were conducted based on a micro-space camera with the most sensitive heat dissipation surface for external heat flow. The finite element method was used to establish a simulation model of the infrared heating cage black-sheet heat flow meter system, and the influence of the traditional infrared heating cage control method on the total arrival energy and heat flow density uniformity of the simulated surface was analyzed. The results show that the heat flow density uniformity of the simulated surface was improved by appropriately enlarging the heating cage and adjusting the position of the heat flow meter paste to ensure that the total arrival energy satisfied the conservative design principle. The statistical variances in the heat flow density of the radiator surface before and after the optimized design were 102.0 and 27.0, respectively, and the homogenization effect was significant. This study can be used as a reference for the accurate simulation of heat flows on other tiny space surfaces.
Design of Adaptive Inversion Proportional-Integral-Derivative Control System for Fast-Steering Mirror
2024, 46(2): 144-149.
The influence of unmeasurable disturbances in a fast-steering mirror system must be considered to improve the beam-tracking performance of a compound-axis system. For measurable disturbances, an adaptive feedforward control algorithm is designed. Inspired by this, an adaptive inversion proportional-integral-derivative(PID) control system for suppressing unmeasurable disturbances was designed. An adaptive algorithm was used to improve the steady-state accuracy of the system and the adaptability to different disturbances. In addition, a PID controller was used to further correct the error signals and improve the dynamic performance of the system. The simulation results show that compared with that of the PID control algorithm, the mean square difference of the error of the adaptive inversion PID control system decreases by 34.76%. Compared with that of the adaptive control algorithm, the mean square difference of the error of the adaptive inversion PID control system decreases by 13.3%. The accuracy of the compound control system significantly improved compared with that of the classical PID and adaptive control systems. When using the compound algorithm, the rise time decreases by 48.9% compared with the adaptive algorithm, and the overshoot decreases by 80.5% compared with the classical PID algorithm. Overall, the dynamic performance of the system improved significantly.
2024, 46(2): 150-154.
Fast-steering mirrors require a rapid dynamic response and an anti-interference ability. In this study, an improved sliding-mode controller based on an extended state observer is proposed to solve the uncertain interference caused by self-movement, external interference, and other factors in the working environment of a fast-steering-mirror system based on the analysis and mathematical modeling of the fast-steering-mirror system. The proposed system uses an extended state observer to observe the unknown disturbance and directly compensate it to the controller, which effectively reduces chattering and facilitates engineering implementation. The simulation results show that compared with the traditional sliding-mode controller, the improved sliding-mode controller based on the extended state observer decreases the rise and adjustment times by 50.4% and 39.1%, respectively, and increases the tracking accuracy by 30.5%. These values satisfy the working requirements of the fast mirror and improves the dynamic performance.
2024, 46(2): 155-161.
Epipolar rectification is a projection transformation method for the original image pair of a binocular camera such that the corresponding polar lines of the corrected image are on the same horizontal line, no vertical parallax occurs, and stereo-matching is optimized as a one-dimensional search problem. A polar correction method based on a binocular camera translation matrix is proposed to address the shortcomings of current polar correction methods. First, the new corrected rotation matrix is derived using the translation matrix of singular value decomposition. Second, a new camera internal reference matrix is established based on the image relationship before and after correction to complete the polar correction. The proposed method was used to verify multiple groups of binocular images in different scenes in the SYNTIM database. The experimental results show that the average correction error is within 0.6 pixels. The image produces minimal distortion, and the average deviation is approximately 2.4°. The average operation time is 0.2302 s. With its application value, this method fully satisfies polar correction requirements, solves the error, and improves the tedious calculation process caused by the mechanical deviation of the camera during the stereo matching of binocular cameras.
2024, 46(2): 162-167.
Object detection is a popular research topic and fundamental task in computer vision. Anchor-based object detection has been widely used in many fields. Current anchor selection methods face two main problems: a fixed size of a priori values based on a specific dataset and a weak generalization ability in different scenarios. The unsupervised K-means algorithm for calculating anchor frames, which is significantly influenced by initial values, generates less variation in anchor points for clustering datasets with a single object size and cannot reflect the multiscale output of the network. In this study, a multiscale anchor (MSA) method that introduces multiscale optimization was developed to address these issues. This method scales and stretches the anchor points generated by clustering according to the dataset characteristics. The optimized anchor points retain the characteristics of the original dataset and reflect the advantages of the multiple scales of the model. In addition, this method was applied to the preprocessing phase of training without increasing the model inference time. Finally, the single-stage mainstream algorithm, You Only Look Once (YOLO), was selected to perform extensive experiments on different scenes of the infrared and industrial scene datasets. The results show that the MSA method can significantly improve the detection accuracy of small-sample scenes.
2024, 46(2): 168-175.
Special materials can fake finger veins to deceive recognition systems, and the large amount of computation is required for finger vein recognition using convolutional neural networks. In this study, a finger vein recognition system with an in vivo detection function and a lightweight convolutional neural network structure was designed. Changes in finger pulse waves were detected using Photo Plethysmo Graphy to determine whether the object was living. The ResNet-18 convolutional neural network was improved using a pruning and channel recovery method, and L1 regularization was combined to improve the feature selection ability of the convolutional neural network. On the basis of improving the accuracy of the algorithm, the network can effectively reduce the consumption of computing resources. The experimental results show that by using the improved pruning and channel recovery optimization structure, the number of parameters decreases by 75.6%, the amount of calculations decreases by 25.6 %, and the iso-error rates obtained in the Shandong University and Hong Kong Polytechnic University finger vein databases are 0.025% and 0.085%, respectively, which are significantly lower than those of RESNET-18 (0.117% and 0.213%).
2024, 46(2): 176-182.
This study proposes a super-resolution reconstruction algorithm for infrared gray images based on deep residual neural networks. Initially, the residual convolution module is employed to deepen the network, enhancing its learning capacity. This enables the convolutional layer to utilize more neighborhood information during learning, resulting in improved ability to process complex scenes. Subsequently, we used the skip connection method to increase the high-frequency information transmission to enhance the image details. Experimental results show that the proposed network can effectively enrich the details of the reconstructed image, and that the target contour in the reconstructed image is significantly improved.
2024, 46(2): 183-189.
A method for generating infrared images in battlefield environments based on JRM was developed to satisfy the performance requirements of a new generation of infrared-imaging target simulation systems and generate realistic infrared images of battlefield environments for infrared-imaging-guided weapons to conduct hardware-in-the-loop simulation tests. First, a three-dimensional model of the target was established using 3DSMAX. The satellite image data and elevation data of the target area were combined to establish a three-dimensional model of the background area to satisfy the requirements of the seeker's field of view. Next, JRM's Genesismc, Sigsim, and Sensim tools were used to model the physical material characteristics, target heat source, environmental characteristics, and sensor characteristics. Finally, the OSV tool was used to render the infrared image in real time. The experimental results show that the method satisfies the requirements of real-time generation of infrared images in an infrared-imaging target simulation system and has the advantages of strong flexibility and good effectiveness.
2024, 46(2): 190-198.
Traditional fusion methods cannot select an effective fusion strategy based on the different characteristics of dual-mode infrared images. A mimic fusion method for the difference between the infrared intensity and polarization images was developed in this study. First, the degree of difference between image features was calculated to roughly screen the difference features, and the selection rules of the main difference feature types were formulated to determine the main difference features of the image groups. Next, the degree of feature fusion was constructed to establish the mapping between the difference features and variables in each layer of the mimic variable set and to determine the hierarchical structure of the variables. Finally, in the hierarchical structure of the variables, the variables of each layer of the main difference feature type were selected. The degrees of feature fusion of the difference features between combined variables of different mimic structures were compared to determine the mimic structure with the highest proportion of its maximum value and form a variant. The experimental results show that the visual effect of the proposed method was better than that of the comparison method after a subjective analysis. After objective evaluation, the results obtained using the proposed method indicate effective fusion. Therefore, this method realizes adaptive selection of the fusion strategy and improves image fusion quality.
2024, 46(2): 199-207.
Infrared small-target detection refers to the detection of small targets in infrared images with low signal-to-noise ratios and complex backgrounds. Infrared small-target detection is essential in applications, such as maritime rescue and traffic management. However, because of factors such as low image resolution, small target size, and inconspicuous features, infrared targets are prone to submergence in a background that contains noise and clutter. The accurate detection of the shape information of small infrared targets remains a challenge. An infrared small-target detection algorithm based on a hierarchical regression transformer (HRformer) network was constructed to address these problems. Specifically, the PixelUnShuffle operation was leveraged to downsample the original image and obtain the input of different network levels to obtain multiscale information while minimizing the loss of the original image information. The PixelShuffle operation upsamples the output feature map of each level, improving the flexibility of the network. Next, a cross-attention fusion module that includes the spatial and channel attention calculation branches realizes efficient feature fusion and information complementarity to realize the information interaction between different levels of features in the network. Finally, combined with the ordinary Transformer structure, which has a large receptive field, and the window-based Transformer, which has the advantage of minimal computational complexity, a local–global transformer structure is proposed to further improve the detection performance of the network and reduce computational costs. The proposed structure can model global dependencies while extracting local context information. The experimental results show that the proposed algorithm has a higher detection accuracy and fewer parameters than some advanced infrared small-target detection algorithms. Therefore, the proposed algorithm is suitable for solving practical problems.
2024, 46(2): 208-215.
The brazing process of TC4/Ni for a Dewar cold finger with a miniature Joule–Thomson-cooled infrared focal plane detector was selected, and the brazing process of TC4/Ni was investigated based on the microstructure of the brazing method and solder type and the reliability of the joint structure. The results show that research on the brazing process of the TC4 /Ni cold finger end-face structure has practical significance in engineering by combining the simulation results of stress, deformation, and cooling time with the analysis results of rust prevention and corrosion. Through an orthogonal test, an improved process scheme of the high-temperature vacuum brazing + AgCu28 brazing filler metal combination was determined, which could control element segregation and reduce the formation of the welding brittle phase. Based on the penetration rate, stamping, high-voltage holding, and shear strength tests, the conical weld was the best welding structure.
2024, 46(2): 216-224.
As a key mode of transportation, bridges bear the high pressure of traffic flow. Many bridges have defects before reaching their designed service life. Bridge-defect recognition based on visible light uses grayscale defect images and regional edge gradient information, which have limitations in complex environments. The radiation and absorption of spectral band signals by objects are detected by hyperspectral imaging, and the signals are transformed into images and graphics. The physical properties of the measured object are analyzed based on the position and intensity of the absorption peak. In this study, a method based on hyperspectral vision is proposed to identify exposed reinforcement bar defects in bridge concrete. Based on the spectral lines and spatial features of hyperspectral images of exposed reinforcement defects in bridge concrete combined with processing——Smooth filtering multivariate scattering calibration (SG-MSC), feature space transformation——First derivative method (FD), and feature variable selection algorithm——Competitive adapative reweighted sampling (CARS), the original spectral curve data were transformed into feature space to extract the corresponding feature values and display the band. The dataset was constructed based on spectral curve feature vectors, and a support vector machine algorithm was used to establish a prediction model for identifying exposed reinforcement defects. Considering a cross-river bridge as an example, a hyperspectral visual testing system was used to identify actual exposed reinforcement bar defects of the bridge. By performing smooth feature space transformation and feature extraction on the original spectral data, the differences were amplified, reducing the dimensionality of the 254 band data to 23 band data and achieving a model prediction accuracy of 94.6%. Hyperspectral vision has higher dimensional information than visible-light vision. Hence, the proposed model can effectively characterize material properties, is feasible, and has broad application prospects.
2024, 46(2): 225-232.
There are numerous problems with infrared imaging using power equipment, such as dark brightness and low contrast. To solve these problems, an enhancement algorithm using a color-model space was proposed. In this method, the contrast and brightness enhancement of the image are processed in the HSV and RGB spaces, respectively. First, the high gray levels of the image are preprocessed, and the mixed filtering method is adopted to suppress the noise in the image. An enhancement function is used to improve the brightness of the image. Finally, the enhanced image is converted into the HSV space, the H, S and V component images are extracted, the gamma transform and CLAHE algorithms are used to improve the brightness of V component, and a nonlinear saturation correction function is used to process component S to improve the image contrast. Finally, the enhanced image in the HSV space is obtained by the corresponding fusion of each processing and extraction component, and is transferred back to the RGB space to obtain the final output image. Experimental results show that the proposed algorithm can significantly improve the contrast and brightness of infrared images. The average gray mean and standard deviation of the enhanced 6 groups of images were 115.94 and 78.65, respectively, which are improvements of 81.59 and 36.17 compared with the original image.