图像处理技术与应用
本专题全面探讨了图像处理领域的创新技术与应用，涵盖了从图像增强、去噪到精确分割的多个方面。专题中的研究文章不仅针对传统算法的不足提出了改进方案，如结合生成对抗网络和检测网络的学习模型，而且还探索了如何通过层次聚类和多线程技术优化图像分割过程。这些研究成果不仅提升了图像处理的效率和精度，还为合成孔径雷达图像、彩色图像等多种图像类型提供了有效的处理方法。本专题旨在汇聚图像处理领域的最新研究成果，促进技术交流，推动图像处理技术在科学研究和工业应用中的进一步发展。
关键词：暂无
2024, 46(2):224-237.
DOI: 10.11887/j.cn.202402023
Abstract:
Since existing low-illumination image enhancement networks have insufficient ability to perceive and express feature information of different scales, a low-illumination image enhancement network model based on pyramid asymptotic fusion was proposed. The network performed multiple down-sampling operations on the image to form a feature pyramid. It fused the feature maps at different scales by adding skip connections to three different branches of the feature pyramid. Fine recovery module further extracted the refined information, and restored the feature map to a normal light image. Results indicate that, the network model not only effectively enhances the brightness of the overall low-illumination image, but also maintains the detailed information and clear edge contours of the objects in the image. Moreover, it can effectively suppress the dark noise, and make the overall enhanced image realistic and natural.
Since existing low-illumination image enhancement networks have insufficient ability to perceive and express feature information of different scales, a low-illumination image enhancement network model based on pyramid asymptotic fusion was proposed. The network performed multiple down-sampling operations on the image to form a feature pyramid. It fused the feature maps at different scales by adding skip connections to three different branches of the feature pyramid. Fine recovery module further extracted the refined information, and restored the feature map to a normal light image. Results indicate that, the network model not only effectively enhances the brightness of the overall low-illumination image, but also maintains the detailed information and clear edge contours of the objects in the image. Moreover, it can effectively suppress the dark noise, and make the overall enhanced image realistic and natural.
2024, 46(2):238-246.
DOI: 10.11887/j.cn.202402024
Abstract:
In view of the structural residual noise in the weighted nuclear norm minimization algorithm and the inability to maintain the edge structure of the image, a denoising method that minimizes the weighted kernel norm and improves the wavelet threshold was adopted. The total variation model to perform preliminary denoising of the noise image, and the noisy image to subtract the preliminary denoised image were used. An improved wavelet threshold function was used to denoise the noise difference image obtained after subtraction. The denoised residual image was superimposed with the preliminary denoised image, and the superimposed image was finally denoised using an iterative weighted kernel norm minimization algorithm based on the residual noise level. Compared with the more popular denoising algorithms, the PSNR and SSIM processed by this algorithm are improved, the texture structure of the image can be maintained, and the effect is better in a high-noise environment.
In view of the structural residual noise in the weighted nuclear norm minimization algorithm and the inability to maintain the edge structure of the image, a denoising method that minimizes the weighted kernel norm and improves the wavelet threshold was adopted. The total variation model to perform preliminary denoising of the noise image, and the noisy image to subtract the preliminary denoised image were used. An improved wavelet threshold function was used to denoise the noise difference image obtained after subtraction. The denoised residual image was superimposed with the preliminary denoised image, and the superimposed image was finally denoised using an iterative weighted kernel norm minimization algorithm based on the residual noise level. Compared with the more popular denoising algorithms, the PSNR and SSIM processed by this algorithm are improved, the texture structure of the image can be maintained, and the effect is better in a high-noise environment.
2023, 45(5):164-172.
DOI: 10.11887/j.cn.202305019
Abstract:
The distortion in the imaging of space-based optical cameras for actual in-orbit earth observations needs to be suppressed by geometric correction. At present, the small size and high frame rate sequential images obtained by mainstream area array cameras for ground observation are difficult to meet the requirements of traditional geometric correction methods for calculating the number of control points and spatial distribution of frame-by-frame solution, and the computational complexity is huge. To address this issue, a geometric correction method by using equivalent bias angle sparse measurement for sequential observation image of area-array sensor was proposed. The problem of parameters solution for each frame converted to the problem of equivalent bias angle recovery under the time-domain sparse measurement. The requirement for the number and spatial distribution of control points in a single frame image were reduced by using the time-frequency information of equivalent bias angle signal. Meanwhile, the real image data from the area-array camera of Gaofen-4 satellite was used to verify the validity and low computation of the proposed method.
The distortion in the imaging of space-based optical cameras for actual in-orbit earth observations needs to be suppressed by geometric correction. At present, the small size and high frame rate sequential images obtained by mainstream area array cameras for ground observation are difficult to meet the requirements of traditional geometric correction methods for calculating the number of control points and spatial distribution of frame-by-frame solution, and the computational complexity is huge. To address this issue, a geometric correction method by using equivalent bias angle sparse measurement for sequential observation image of area-array sensor was proposed. The problem of parameters solution for each frame converted to the problem of equivalent bias angle recovery under the time-domain sparse measurement. The requirement for the number and spatial distribution of control points in a single frame image were reduced by using the time-frequency information of equivalent bias angle signal. Meanwhile, the real image data from the area-array camera of Gaofen-4 satellite was used to verify the validity and low computation of the proposed method.
2023, 45(3):127-135.
DOI: 10.11887/j.cn.202303015
Abstract:
In order to solve the defects of edge artifacts and semantic discontinuity when inpainting large and irregular distortions in an image with fine texture and complex background, a residual connection image restoration algorithm based on parallel generation convolution was proposed. The damaged image was inputted into a two-column parallel convolutional structure inpainting network to obtain two image components with different sizes of reception fields. The two image components were combined by the shared decoding and the L2 loss of the output was calculated to optimize the network. The output of the coarse network was sent into the fine inpainting network which contained the residual connection and the attention mechanism to fuse context information and improve the ability of repairing fine details. The global and local discriminators and visual geometry group net were used to calculate the loss and to optimize the overall discrimination network and enhance the global and local consistency of the inpainting result. The performance of the proposed algorithm was validated on internationally recognized databases, and experimental results show that the proposed algorithm can effectively repair large and irregular missing areas under complex background and fine texture, improve the authenticity and integrity of image details, semantics and structure. Its peak signal-to-noise ratio and structural similarity are superior to the state-of-the-art.
In order to solve the defects of edge artifacts and semantic discontinuity when inpainting large and irregular distortions in an image with fine texture and complex background, a residual connection image restoration algorithm based on parallel generation convolution was proposed. The damaged image was inputted into a two-column parallel convolutional structure inpainting network to obtain two image components with different sizes of reception fields. The two image components were combined by the shared decoding and the L2 loss of the output was calculated to optimize the network. The output of the coarse network was sent into the fine inpainting network which contained the residual connection and the attention mechanism to fuse context information and improve the ability of repairing fine details. The global and local discriminators and visual geometry group net were used to calculate the loss and to optimize the overall discrimination network and enhance the global and local consistency of the inpainting result. The performance of the proposed algorithm was validated on internationally recognized databases, and experimental results show that the proposed algorithm can effectively repair large and irregular missing areas under complex background and fine texture, improve the authenticity and integrity of image details, semantics and structure. Its peak signal-to-noise ratio and structural similarity are superior to the state-of-the-art.
2022, 44(5):53-62.
DOI: 10.11887/j.cn.202205006
Abstract:
For the common image lossless compression methods cannot work well. Thus, a lossless compression method for synchrotron radiation source images based on image difference and neural network was proposed. The image difference method was used to reduce the linear correlations among images. The neural network was trained to learn the nonlinear correlations in the images sequence, and the pixel value was compressed with arithmetic coding using the predicted distribution. To reduce the predicting time and coding time, the pixel value was splitted into two parts for parallel compression. The tests based on the images of Shanghai Synchrotron Radiation Facility show that the proposed method can improve more than 20% in compression ratio compared to PNG(portable network graphics), JPEG2000, FLIF(free lossless image format), and the pixel value split can reduce 30% of the time in predicting and coding.
For the common image lossless compression methods cannot work well. Thus, a lossless compression method for synchrotron radiation source images based on image difference and neural network was proposed. The image difference method was used to reduce the linear correlations among images. The neural network was trained to learn the nonlinear correlations in the images sequence, and the pixel value was compressed with arithmetic coding using the predicted distribution. To reduce the predicting time and coding time, the pixel value was splitted into two parts for parallel compression. The tests based on the images of Shanghai Synchrotron Radiation Facility show that the proposed method can improve more than 20% in compression ratio compared to PNG(portable network graphics), JPEG2000, FLIF(free lossless image format), and the pixel value split can reduce 30% of the time in predicting and coding.
6 Smooth principal component analysis network image recognition algorithm with fusion graph embedding
2022, 44(3):16-22.
DOI: 10.11887/j.cn.202203003
Abstract:
PCANet (principal component analysis network) is a simple deep learning algorithm with excellent performance in the field of image recognition. Integrating the idea of graph embedding into PCANet, a new image recognition algorithm Smooth-PCANet was proposed. In order to verify the effectiveness of the Smooth-PCANet algorithm, adequate experiments were performed on different data sets such as face, handwritten characters, and images. Compared with several image recognition algorithms based on deep learning, the experiments demonstrated that the Smooth-PCANet achieves higher recognition performance than the PCANet and avoids overfitting more effectively, with a significant advantage in small samples training.
PCANet (principal component analysis network) is a simple deep learning algorithm with excellent performance in the field of image recognition. Integrating the idea of graph embedding into PCANet, a new image recognition algorithm Smooth-PCANet was proposed. In order to verify the effectiveness of the Smooth-PCANet algorithm, adequate experiments were performed on different data sets such as face, handwritten characters, and images. Compared with several image recognition algorithms based on deep learning, the experiments demonstrated that the Smooth-PCANet achieves higher recognition performance than the PCANet and avoids overfitting more effectively, with a significant advantage in small samples training.
7 Target detection in SAR images based on joint generative adversarial network and detection network
2022, 44(3):164-175.
DOI: 10.11887/j.cn.202203020
Abstract:
Aiming at the problem of difficult and limited sample acquisition in SAR(synthetic aperture radar) image target detection, a learning model of joint GAN (generative adversarial network) and detection network was proposed. The original training set was used to pretrain the specially designed faster regional convolutional neural network. The deep convolutional GAN based on attention mechanism was employed to generate extensive synthetic samples, which were input into the detection network for prediction. The corresponding annotation information of the new samples was determined by the prediction information and probability equivalent class label allocation strategy, and the annotated new samples were used to expand the original dataset with a certain proportion. The detection network was retrained with the expanded dataset. Simulation results show that the proposed framework can improve the detection efficiency and performance of the network effectively.
Aiming at the problem of difficult and limited sample acquisition in SAR(synthetic aperture radar) image target detection, a learning model of joint GAN (generative adversarial network) and detection network was proposed. The original training set was used to pretrain the specially designed faster regional convolutional neural network. The deep convolutional GAN based on attention mechanism was employed to generate extensive synthetic samples, which were input into the detection network for prediction. The corresponding annotation information of the new samples was determined by the prediction information and probability equivalent class label allocation strategy, and the annotated new samples were used to expand the original dataset with a certain proportion. The detection network was retrained with the expanded dataset. Simulation results show that the proposed framework can improve the detection efficiency and performance of the network effectively.
2022, 44(3):194-200.
DOI: 10.11887/j.cn.202203023
Abstract:
Based on the GBS(graph-based segmentation) algorithm and the hierarchical clustering algorithm, a method to solve the under-segmentation of GBS algorithm was constructed. Meanwhile, the way of multi-threaded parallel processing of data was used to effectively improve the processing speed of the traditional hierarchical clustering algorithm. In the RGB color space, the GBS algorithm was used to obtain the initial segmentation result of each pixel in the image. The pixel value in each type of region was extracted and the hierarchical clustering was carried out to obtain the category label of pixel value in each type of region. According to the category label obtained by hierarchical clustering and the preset category range, the initial segmentation result of each pixel was modified. A new segmentation graph was generated according to the region merging criterion. Experimental results show that compare with the K-means-SLIC algorithm and the GBS algorithm, this method solves the phenomenon of under-segmentation, and produces a semantic segmentation graph with high segmentation accuracy.
Based on the GBS(graph-based segmentation) algorithm and the hierarchical clustering algorithm, a method to solve the under-segmentation of GBS algorithm was constructed. Meanwhile, the way of multi-threaded parallel processing of data was used to effectively improve the processing speed of the traditional hierarchical clustering algorithm. In the RGB color space, the GBS algorithm was used to obtain the initial segmentation result of each pixel in the image. The pixel value in each type of region was extracted and the hierarchical clustering was carried out to obtain the category label of pixel value in each type of region. According to the category label obtained by hierarchical clustering and the preset category range, the initial segmentation result of each pixel was modified. A new segmentation graph was generated according to the region merging criterion. Experimental results show that compare with the K-means-SLIC algorithm and the GBS algorithm, this method solves the phenomenon of under-segmentation, and produces a semantic segmentation graph with high segmentation accuracy.
