Point Cloud Analysis Combining Gated Self-Calibration Mechanism and Graphical Convolutional Network
Xu Jiali, Fang Zhijun, and Wu Shiqian
Point clouds, unlike images represented by dense grids, are characterized by irregularity and disorder, making it difficult to precisely reason out the shape features in point cloud data. The internal-external shape son volution for point sets (IE-Conv) is proposed to address the limitations of current research. The local shape inside the point set is treated separately from the global shape outside the point set using an efficient bilateral design. Rich inter-point relationships are selectively studied in a gate-based manner within the point set, while point-by-point and local features are optimized by self-calibration functions; outside the point set, global shapes are constructed using graph convolution and focus on long-range dependencies between point sets. Finally, the organic fusion of the bilateral outputs is performed. This paper performs classification and segmentation experiments on the standard ModelNet40 and ShapeNet datasets by hierarchically embedding IE-Conv into the shape-reasoning convolutional network (SR-Net). The experimental results show that the classification task achieves an accuracy of 93.9% and the segmentation task achieves the mean intersection over the union of 86.4%, which verifies the good performance of SR-Net in point cloud analysis.
  • May. 23, 2022
  • Laser & Optoelectronics Progress
  • Vol.59 Issue, 12 1210017 (2022)
  • DOI:10.3788/LOP202259.1210017
Fine-Grained Image Recognition of Wild Mushroom Based on Multiscale Feature Guide
Zhang Zhigang, Yu Pengfei, Li Haiyan, and Li Hongsong
Deep learning technology is proposed to solve the social problem of the frequent occurrences of wild mushroom poisoning in China. However, due to the small difference between classes and complex image backgrounds, fine-grained recognition accuracy is low. To solve this problem, this paper proposes an improved ResNeXt50 network. First, a multiscale feature guide (MSFG) module is designed, which guides the network to learn and use low and high-level features fully through short connections. Then, the improved attention mechanism module is used to reduce the network’s learning for complex backgrounds. Finally, the different hierarchical features in the model are fused, and the obtained joint features are used for recognition. Experimental results show that the accuracy of the proposed network on the test set can reach 96.47%, which is 2.64 percentage points higher than the unimproved ResNeXt50 network. Comparison results show that the accuracy of the improved network model is 8.10 percentage points, 5.13 percentage points, 3.24 percentage points, 3.30 percentage points, and 4.25 percentage points better than VGG19, DenseNet121, Inception_v3, ResNet50, and ShuffleNet_v2, respectively.
  • May. 23, 2022
  • Laser & Optoelectronics Progress
  • Vol.59 Issue, 12 1210016 (2022)
  • DOI:10.3788/LOP202259.1210016
Remote Sensing Image Segmentation Using Super-Pixel and Dot Product Representation of Graphs
Zhang Daming, Zhang Xueyong, Liu Huayong, and Li Lu
The image segmentation method using division-combination mitigates the limitations of the traditional pixel-based remote sensing image segmentation algorithm, such as noise interference, low segmentation efficiency, and poor segmentation effect. Thus, this paper proposes a new split-merge-based remote sensing image segmentation method using the super-pixel and dot product representation of graphs. First, the image is divided into super-pixels using the simple linear iterative clustering (SLIC) algorithm. Second, the texture feature of each super-pixel area is measured and distance between any two areas is calculated with respect to spatial proximity. Third, each super-pixel area is mapped as a vertex of the graph. Therefore, the dot product representation of graphs is modified and used to construct a similarity matrix; thereafter, all vertices (i.e., super-pixel areas) are mapped as new vectors clustered by angular-based k-means algorithm to get the final segmentation results. The experimental results show that the proposed method has stable segmentation results, improves the accuracy of the segmentation, and achieves a better visual segmentation effect.
  • May. 23, 2022
  • Laser & Optoelectronics Progress
  • Vol.59 Issue, 12 1210015 (2022)
  • DOI:10.3788/LOP202259.1210015
Reconstruction of Magnetic Resonance Images Based on Dual-Domain Crossed Codec Network
Zhang Dengqiang, Liu Xiaohan, and Pang Yanwei
Magnetic resonance imaging (MRI) has outstanding soft-tissue contrast and provides unparalleled benefits in various diagnoses. It is an important way of observation in current clinical practice. The scanning period of an MRI, however, is long, which greatly limits the diagnostic efficiency. Obtaining undersampled K-space data through partial scanning at a specific acceleration magnification is a critical approach to save scanning time. Existing approaches only rebuild the K-domain or the image domain alone or alternately process the two domains through serially coupled image domain and K-domain convolution, resulting in poor reconstruction performance. A dual-domain parallel codec structure that processes image domain and K-domain data simultaneously is presented to provide high-quality reconstruction of undersampled K-space data at high acceleration rates. The proposed technique reconstructs the undersampled image domain and K-domain data using two parallel codec networks, respectively, then combines the features of the K-domain branch into the image domain using the inverse Fourier transform, considerably enhancing reconstruction quality. For presampling data with varying acceleration magnifications, experimental results reveal that the proposed method outperforms other U-Net-based image reconstruction methods. This proposed method is projected to develop into a high-performance, high-acceleration-magnification MRI undersampling data reconstruction method that can be used in clinical MRI reconstruction.
  • May. 23, 2022
  • Laser & Optoelectronics Progress
  • Vol.59 Issue, 12 1210014 (2022)
  • DOI:10.3788/LOP202259.1210014
Visual Tracking Combining Attention and Feature Fusion Network Modulation
Xu Keying, Shu Ping, and Bao Hua
The existing tracking algorithms for network modulation ignore high order feature information, so they are prone to drift when dealing with large scale changes and object deformations. An object tracking algorithm that combines the attention mechanism and feature fusion network modulation is proposed. First, an efficient selective kernel attention module is embedded in the feature extraction backbone network, so that the network pays more attention to the extraction of target feature information; second, a multiscale interactive network is used for the extracted features to fully mine the multiscale information in the layer, and high order feature information is fused to improve the ability of target representation, to adapt to the complex and changeable environment in the tracking process; finally, the pyramid modulation network is used to guide the test branch to learn the optimal intersection over union prediction to achieve an accurate estimation of the targets. Experimental results show that the proposed algorithm achieves more competitive results than other algorithms in tracking accuracy and success rate on VOT2018, OTB100, GOT10k, TrackingNet, and LaSOT visual tracking benchmarks.
  • May. 23, 2022
  • Laser & Optoelectronics Progress
  • Vol.59 Issue, 12 1210013 (2022)
  • DOI:10.3788/LOP202259.1210013
Image Feature Point Matching Algorithm Based on Oriented Fast and Rotated Brief Algorithm and Hue, Saturation and Value
Shan Yusi, Chen Bo, and Cheng Pengfei
An image feature point matching algorithm based on the oriented fast and rotated brief (ORB) algorithm and hue, saturation and value (HSV) is proposed and the experimental research is carried out. Firstly, the image is preprocessed by the combination of bilateral filtering and mean filtering. Secondly, the ORB algorithm is used to extract feature points. Thirdly, the K-D Tree algorithm and Hamming distance are used for matching of feature points roughly, and then the HSV information of the image are used for the secondary screening of matched feature point pairs. The experimental results show that, in the image preprocessing stage, the weighted average of variance, vollath and information entropy is used as the evaluation index, and compared with the original image, histogram equalization and bilateral filtering results, the evaluation index value obtained by the combination of bilateral filtering and mean filtering is the best. In the stage of feature point matching and image mosaic, the average matching correct rate of feature points is improved by 12.60 percentage points after using HSV information, and the quality of image mosaic result is better, as its natural image quality evaluation (NIQE) index value is smaller.
  • May. 23, 2022
  • Laser & Optoelectronics Progress
  • Vol.59 Issue, 12 1210012 (2022)
  • DOI:10.3788/LOP202259.1210012
Remote Sensing Aircraft Detection Based on Smooth Label and Multipath Aggregation Network
Li Kewen, Zhang Baohua, Lv Xiaoqi, Gu Yu, Wang Yueming, Liu Xin, Ren Yan, Li Jianjun, and Zhang Ming
In order to solve the problem of complex background in remote sensing images and the large variation of aircraft target size, a new algorithm for remote sensing aircraft detection based on the smooth label and multipath aggregation network is proposed. Considering the difficulty of aircraft target identification in remote sensing images, an associative attention mechanism is used to capture the target area and narrow the search range. Then, the improved path aggregation network is used to extract the four feature layers in the backbone network, so as to effectively extract the shallow feature information. When the features of each layer are normalized, they are fused to predict the position of the target. In order to avoid the training model relying too much on the prediction labels, resulting in over fitting, technology for the smooth label is used in the network to reduce the inter-class distance, which effectively improves the generalization ability of the training model. The effectiveness of the proposed algorithm is verified by a large number of experiments on two public data sets RSOD and HRRSD. The experimental results show that the average accuracy in the RSOD data set and the HRRSD data set is 0.967 and 0.993 respectively. Compared with the related algorithms, the detection accuracy of the proposed algorithm has been significantly improved.
  • May. 23, 2022
  • Laser & Optoelectronics Progress
  • Vol.59 Issue, 12 1210011 (2022)
  • DOI:10.3788/LOP202259.1210011
Registration Method for Power Equipment Infrared and Visible Images Based on Improved Curvature Scale Space Algorithm
Li Yunhong, Luo Xuemin, Su Xueping, Zhu Yaolin, Yao Lan, and Duan Jiaojiao
Aiming at the problems of difficult registration and long registration time of existing power equipment infrared and visible images, a registration method for power equipment infrared and visible images based on improved curvature scale space (CSS) algorithm is proposed. Firstly, the Freeman chain code difference is introduced to improve the feature point extraction accuracy of the CSS algorithm; Secondly, each feature point is assigned the main direction of the vertical distance from the point to the string, and the feature description operator is obtained using the speed up robust features (SURF) algorithm; Finally, the two-sided fast library for approximate nearest neighbors (FLANN) search matching and random sample consensus (RANSAC) method are used to obtain the correct matching point pair, and the affine transformation model parameters are obtained. The experimental results show that compared with the SURF, scale invariant feature transform (SIFT), and CSS registration methods, the improved CSS image registration method has significantly improved performance indicators and its average root mean square error (RMSE) is reduced by 77.73%, 80.32% and 7.63% and its average matching time is reduced by 30.82%, 40.12% and 10.57%, respectively. It improves the registration efficiency of the power equipment infrared and visible images.
  • May. 23, 2022
  • Laser & Optoelectronics Progress
  • Vol.59 Issue, 12 1210010 (2022)
  • DOI:10.3788/LOP202259.1210010
Hyperspectral Fast Spectral Clustering Algorithm Based on Multi-Layer Bipartite Graph
Li Siyuan, Zheng Zhiyuan, Du Xiaoyan, Liu Tong, and Yang Xiaojun
Large-scale hyperspectral image clustering algorithms are widely used in the field of remote sensing, including K-means clustering and spectral clustering algorithms. However, the spectral clustering algorithm still has its limitations. Because of its high computational complexity, it is not suitable for large-scale problems. The spectral clustering algorithm based on the anchor graph can reduce the computational cost to a certain extent. However, in the large-scale hyperspectral image data processsing, the anchor points need to be dense enough, otherwise reasonable accuracy cannot be obtained. This makes the computing cost of the clustering algorithm increase sharply. In order to overcome these problems, a new fast spectral clustering algorithm based on multi-layer bipartite graph is proposed. Firstly, the anchor points are selected by the binary tree,and the multi-layer anchor points are selected to construct the multi-layer anchor point graph. Then a multi-layer bipartite graph is constructed, and finally the spectrum of the graph is analyzed. The high efficiency of the proposed algorithm is proved by experiments.
  • May. 23, 2022
  • Laser & Optoelectronics Progress
  • Vol.59 Issue, 12 1210009 (2022)
  • DOI:10.3788/LOP202259.1210009
PCB Image-Denoising Algorithm Based on Image Difference and Residual Learning
Ran Guangzai, Xu Lei, Li Dashuang, and Guo Zhanling
Current printed circuit board (PCB) image-denoising algorithms can easily produce excessive edge smoothing and detail loss in the denoising process. To improve the effect of PCB image denoising, this paper proposes a PCB image-denoising algorithm based on residual learning and image difference. First, an image downsampling method is used to expand the receptive field of the image based on the idea of residual learning. Thereafter, a residual block is designed to extract the noise characteristics of the PCB image. Meanwhile, batch normalization and ReLU activation function are added to the residual convolutional neural network element to improve the denoising efficiency. Finally, the noise is removed through the image difference process. The experimental denoising performance of various algorithms is compared under different noise levels and the results show that the algorithm proposed in this paper has better performance than other algorithms in terms of peak signal-to-noise ratio (PSNR) and structural similarity (SSIM).
  • May. 23, 2022
  • Laser & Optoelectronics Progress
  • Vol.59 Issue, 12 1210008 (2022)
  • DOI:10.3788/LOP202259.1210008