Journals > > Topics > Optics in Computing
Optics in Computing|30 Article(s)
Influence of Hyperparameters on Performance of Optical Neural Network Training Algorithms
Wen Cao, Meiyu Liu, Minghao Lu, Xiaofeng Shao, Qifa Liu, and Jin Wang
An optical neural network (ONN) based on fast Fourier transform (FFT) is constructed for digital image recognition in optical devices. Herein, ONN uses Mach-Zehnder interferometer (MZI) as its linear optical processing unit. These MZIs are connected in a grid-like layout and modulate the passing optical signals to achieve multiplication and addition. Subsequently, MZIs achieve classification and recognition for images. In this study, the influence of main hyperparameters (e.g., momentum coefficient and learning rate of the training algorithm) on the performance of ONN in recognizing handwritten digital images is investigated. First, the performance of ONN with four training algorithms in recognizing handwritten digital images under different learning rates is compared. These algorithms connect with different nonlinear functions and different number of hidden layers, namely, stochastic gradient descent (SGD), root mean square prop (RMSprop), adaptive moment estimation (Adam), and adaptive gradient (Adagrad). Additionally, the accuracy, running memory, and training time of ONN with the SGD algorithm connected with different nonlinear functions and different number of hidden layers are analyzed under different momentum coefficients. The recognition performance of ONN with SGD and RMSprop training algorithms is also compared after the introduction of momentum, where the learning rate is 0.05 and 0.005. The experimental results show that when the learning rate changes from 0.5 to 5 × 10-5, the FFT-typed ONN with the RMSprop training algorithm, two hidden layers, and the nonlinear function of Softplus has the highest recognition accuracy, reaching 97.4%. Furthermore, for the momentum coefficient of 0, the ONN with two hidden layers and the nonlinear function of Softplus trained by the SGD algorithm exhibits the highest recognition accuracy of 96%, when the momentum coefficient increases to 0.9, the accuracy of ONN is improved to 96.9%. However, the RMSprop algorithm with momentum leads to nonconvergence or slow convergence of network recognition accuracy. An optical neural network (ONN) based on fast Fourier transform (FFT) is constructed for digital image recognition in optical devices. Herein, ONN uses Mach-Zehnder interferometer (MZI) as its linear optical processing unit. These MZIs are connected in a grid-like layout and modulate the passing optical signals to achieve multiplication and addition. Subsequently, MZIs achieve classification and recognition for images. In this study, the influence of main hyperparameters (e.g., momentum coefficient and learning rate of the training algorithm) on the performance of ONN in recognizing handwritten digital images is investigated. First, the performance of ONN with four training algorithms in recognizing handwritten digital images under different learning rates is compared. These algorithms connect with different nonlinear functions and different number of hidden layers, namely, stochastic gradient descent (SGD), root mean square prop (RMSprop), adaptive moment estimation (Adam), and adaptive gradient (Adagrad). Additionally, the accuracy, running memory, and training time of ONN with the SGD algorithm connected with different nonlinear functions and different number of hidden layers are analyzed under different momentum coefficients. The recognition performance of ONN with SGD and RMSprop training algorithms is also compared after the introduction of momentum, where the learning rate is 0.05 and 0.005. The experimental results show that when the learning rate changes from 0.5 to 5 × 10-5, the FFT-typed ONN with the RMSprop training algorithm, two hidden layers, and the nonlinear function of Softplus has the highest recognition accuracy, reaching 97.4%. Furthermore, for the momentum coefficient of 0, the ONN with two hidden layers and the nonlinear function of Softplus trained by the SGD algorithm exhibits the highest recognition accuracy of 96%, when the momentum coefficient increases to 0.9, the accuracy of ONN is improved to 96.9%. However, the RMSprop algorithm with momentum leads to nonconvergence or slow convergence of network recognition accuracy.
Laser & Optoelectronics Progress
- Publication Date: Nov. 25, 2023
- Vol. 60, Issue 22, 2220001 (2023)
Marine Creature Detection Based on Sample Iterative Fusion
Lidong Wu, Zongju Peng, Xin Li, Tao Su, Fen Chen, and Xiaodong Wang
Occlusion caused by gathering of marine creatures together is an important reason for false and missed detections. Therefore, this study proposes a marine creature detection method based on iterative fusion of sample-assisted network training. First, an improved deep hole residual structure is selected as the feature extraction network, which improves the feature extraction ability of the network. Second, because of the occlusion and dense characteristics of marine creature images, the loss function is improved to avoid false and missed detections. Finally, to solve the problems of target occlusion and data imbalance, a sample iterative fusion method is proposed to generate an extended training set of simulated images. This improves the effectiveness of network training and the ability to detect marine creatures with a small sample size. The experimental results show that the proposed method can achieve a detection accuracy of 91.36% on the URPC2018 dataset and 90.27% on the Taiwan fish dataset. The detection accuracy and speed of the proposed method are higher than those of existing target detection algorithms. Occlusion caused by gathering of marine creatures together is an important reason for false and missed detections. Therefore, this study proposes a marine creature detection method based on iterative fusion of sample-assisted network training. First, an improved deep hole residual structure is selected as the feature extraction network, which improves the feature extraction ability of the network. Second, because of the occlusion and dense characteristics of marine creature images, the loss function is improved to avoid false and missed detections. Finally, to solve the problems of target occlusion and data imbalance, a sample iterative fusion method is proposed to generate an extended training set of simulated images. This improves the effectiveness of network training and the ability to detect marine creatures with a small sample size. The experimental results show that the proposed method can achieve a detection accuracy of 91.36% on the URPC2018 dataset and 90.27% on the Taiwan fish dataset. The detection accuracy and speed of the proposed method are higher than those of existing target detection algorithms.
Laser & Optoelectronics Progress
- Publication Date: Jan. 25, 2023
- Vol. 60, Issue 2, 0220001 (2023)
Application of Improved Sparrow Algorithm in Sky-Wave Radar Location
Shen Chen, Yian Liu, and Hailing Song
As complex battlefield environment requires rapid and accurate positioning of sky-wave radar, a positioning model based on multi-strategy improved sparrow search algorithm is proposed. First, cubic chaotic mapping, dynamic adjustment of step factor, reverse learning, and mixed mutation operator are used to invent an improved sparrow search algorithm. Then, the improved sparrow search algorithm is used to find the best-fit kernel function parameters and weight coefficient of mixed kernel of hybrid kernel extreme learning machine (HKELM). Finally, the optimized HKELM is used to locate the target detected by the skywave radar. The results show that the accuracy and stability of improved sparrow search algorithm are not only superior to the HKELM location model which is optimized by the basic sparrow search algorithm, but also stronger than the extreme learning machine (ELM) location model. In other words, the effectiveness of the method is proved. As complex battlefield environment requires rapid and accurate positioning of sky-wave radar, a positioning model based on multi-strategy improved sparrow search algorithm is proposed. First, cubic chaotic mapping, dynamic adjustment of step factor, reverse learning, and mixed mutation operator are used to invent an improved sparrow search algorithm. Then, the improved sparrow search algorithm is used to find the best-fit kernel function parameters and weight coefficient of mixed kernel of hybrid kernel extreme learning machine (HKELM). Finally, the optimized HKELM is used to locate the target detected by the skywave radar. The results show that the accuracy and stability of improved sparrow search algorithm are not only superior to the HKELM location model which is optimized by the basic sparrow search algorithm, but also stronger than the extreme learning machine (ELM) location model. In other words, the effectiveness of the method is proved.
Laser & Optoelectronics Progress
- Publication Date: May. 25, 2023
- Vol. 60, Issue 10, 1020001 (2023)
Three-Dimensional Circular Hole Recognition Algorithm Based on Point Cloud Normal and Projection Fusion
Haoyu Li, Yunjie Yang, Hao Yang, and Yu Fang
This paper proposes a three-dimensional (3D) circular hole recognition algorithm based on the fusion of point cloud normal and projection to solve the problems of poor extraction accuracy and the incomplete edge point extraction of the existing 3D circular hole recognition algorithms. Firstly, k-dimensional tree(KD-tree) assists in establishing the spatial topological relationship of the point cloud. Secondly, K-NearestNeighbor(KNN) is used to search the k neighborhood points closest to the point. The point greater than the threshold is determined as the boundary point by defining the distance threshold. Finally, the fusion of point cloud normal and projection are combined to realize the distinction between feature points and noise points at the edge of the point cloud, and the 3D circular hole features of the point cloud data are extracted. The experimental results show that the algorithm can effectively realize point cloud edge extraction and 3D circular hole recognition. This paper proposes a three-dimensional (3D) circular hole recognition algorithm based on the fusion of point cloud normal and projection to solve the problems of poor extraction accuracy and the incomplete edge point extraction of the existing 3D circular hole recognition algorithms. Firstly, k-dimensional tree(KD-tree) assists in establishing the spatial topological relationship of the point cloud. Secondly, K-NearestNeighbor(KNN) is used to search the k neighborhood points closest to the point. The point greater than the threshold is determined as the boundary point by defining the distance threshold. Finally, the fusion of point cloud normal and projection are combined to realize the distinction between feature points and noise points at the edge of the point cloud, and the 3D circular hole features of the point cloud data are extracted. The experimental results show that the algorithm can effectively realize point cloud edge extraction and 3D circular hole recognition.
Laser & Optoelectronics Progress
- Publication Date: Apr. 25, 2022
- Vol. 59, Issue 8, 0820002 (2022)
Power Prediction of Photovoltaic Generation Based on Improved Temporal Convolutional Network
Guilan Li, Jie Yang, and Manguo Zhou
To improve the efficiency of photovoltaic (PV) power forecasting, the method of feature fusion combined with improved temporal convolutional network (TCN) is proposed. The correlation coefficient approach is utilized to examine the time series features, and the effective input for feature fusion is calculated. To increase the accuracy of generating power forecasting, the TCN expansion parameters and connection modes are adjusted. The proposed method is evaluated on two different power plant data sets in South China, and it is compared to the classical algorithms LSTM, GRU, 1D-CNN, and TCN, as well as diverse weather samples. The results reveal that the approach described in this paper achieves a decisive coefficient of 0.982 and outperforms other algorithms in terms of fitting ability. The training time of the model is only 30 s, and the prediction efficiency is greatly improved. To improve the efficiency of photovoltaic (PV) power forecasting, the method of feature fusion combined with improved temporal convolutional network (TCN) is proposed. The correlation coefficient approach is utilized to examine the time series features, and the effective input for feature fusion is calculated. To increase the accuracy of generating power forecasting, the TCN expansion parameters and connection modes are adjusted. The proposed method is evaluated on two different power plant data sets in South China, and it is compared to the classical algorithms LSTM, GRU, 1D-CNN, and TCN, as well as diverse weather samples. The results reveal that the approach described in this paper achieves a decisive coefficient of 0.982 and outperforms other algorithms in terms of fitting ability. The training time of the model is only 30 s, and the prediction efficiency is greatly improved.
Laser & Optoelectronics Progress
- Publication Date: Apr. 25, 2022
- Vol. 59, Issue 8, 0820001 (2022)
Super-Resolution Reconstruction Algorithm of Images Based on Improved Enhanced Super-Resolution Generative Adversarial Network
Yuanxue Xin, Fengting Zhu, Pengfei Shi, Xin Yang, and Runkang Zhou
To solve the problem of insufficient detail processing in the existing image super-resolution reconstruction algorithm, a super-resolution reconstruction algorithm of images based on improved enhanced super-resolution generative adversarial network (ESRGAN) is proposed. Firstly, the deep information extraction module of the improved ESRGAN generation network is improved using multiscale dense block (MDB) instead of dense block (DB), and by adding channel attention mechanism after MDB to adjust the characteristic response values of different channels. Secondly, the shallow feature extraction module of the improved ESRGAN model is used to extract the original features of the low resolution images, and the deep information extraction module is used to extract the depth residual features of the low resolution images. The original features and the depth residual features are fused by adding the corresponding elements. Finally, the reconstruction module is used to complete the image super-resolution reconstruction. The proposed algorithm's two and four times super-resolution reconstructions are tested on Set5, Set14, and BSD100 datasets and compared to Bicubic, FSRCNN, and ESRGAN methods. The results show that the proposed algorithm's reconstructed image has a clearer edge, and it can provide more details, which greatly improves the image's visual effect. Compared to ESRGAN, the proposed algorithm improves the average peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of 2-fold super-resolution reconstruction images by 0. 467 dB and 0. 005, respectively; At the same time, the proposed algorithm improves the average PSNR and SSIM of 4-fold super-resolution reconstruction images by 0.438 dB and 0.015, respectively. To solve the problem of insufficient detail processing in the existing image super-resolution reconstruction algorithm, a super-resolution reconstruction algorithm of images based on improved enhanced super-resolution generative adversarial network (ESRGAN) is proposed. Firstly, the deep information extraction module of the improved ESRGAN generation network is improved using multiscale dense block (MDB) instead of dense block (DB), and by adding channel attention mechanism after MDB to adjust the characteristic response values of different channels. Secondly, the shallow feature extraction module of the improved ESRGAN model is used to extract the original features of the low resolution images, and the deep information extraction module is used to extract the depth residual features of the low resolution images. The original features and the depth residual features are fused by adding the corresponding elements. Finally, the reconstruction module is used to complete the image super-resolution reconstruction. The proposed algorithm's two and four times super-resolution reconstructions are tested on Set5, Set14, and BSD100 datasets and compared to Bicubic, FSRCNN, and ESRGAN methods. The results show that the proposed algorithm's reconstructed image has a clearer edge, and it can provide more details, which greatly improves the image's visual effect. Compared to ESRGAN, the proposed algorithm improves the average peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) of 2-fold super-resolution reconstruction images by 0. 467 dB and 0. 005, respectively; At the same time, the proposed algorithm improves the average PSNR and SSIM of 4-fold super-resolution reconstruction images by 0.438 dB and 0.015, respectively.
Laser & Optoelectronics Progress
- Publication Date: Feb. 20, 2022
- Vol. 59, Issue 4, 0420002 (2022)
Soft Pseudo-Label and Multi-Scale Feature Fusion for Person Re-Identification
Hao Chen, Baohua Zhang, Xiaoqi Lü, Yu Gu, Yueming Wang, Xin Liu, Yan Ren, Jianjun Li, and Ming Zhang
The traditional unsupervised domain adaptive person re-identification algorithm suppressed the noise of pseudo-label poorly and lack inter-domain generalization ability. For the above problems, an unsupervised domain adaptive person re-identification algorithm was proposed which based on soft pseudo-label and multi-scale feature reconstruction. In order to suppress pseudo-label noise, the predicted value of the parallel network is used as the soft tag, and pseudo-label noise is corrected by cross-proofreading methods, which provides a more robust soft false tag for unsupervised domain adaptive tasks. In order to enhance the generalization ability between domains, multi-scale feature reconstruction and Hadamard product feature fusion methods are used to process the deep and shallow feature layer information, realize the style conversion from source domain data to target domain, and solve the problem of poor adaptability of residual network domain with instance normalization and batch normalization network, so as to enhance the generalization ability of the network to source domain and target domain. Experimental results show that the proposed algorithm has achieved good performance in both Market to Duke and Duke to Market unsupervised domain adaptive tasks, which is significantly better than the related algorithms. The traditional unsupervised domain adaptive person re-identification algorithm suppressed the noise of pseudo-label poorly and lack inter-domain generalization ability. For the above problems, an unsupervised domain adaptive person re-identification algorithm was proposed which based on soft pseudo-label and multi-scale feature reconstruction. In order to suppress pseudo-label noise, the predicted value of the parallel network is used as the soft tag, and pseudo-label noise is corrected by cross-proofreading methods, which provides a more robust soft false tag for unsupervised domain adaptive tasks. In order to enhance the generalization ability between domains, multi-scale feature reconstruction and Hadamard product feature fusion methods are used to process the deep and shallow feature layer information, realize the style conversion from source domain data to target domain, and solve the problem of poor adaptability of residual network domain with instance normalization and batch normalization network, so as to enhance the generalization ability of the network to source domain and target domain. Experimental results show that the proposed algorithm has achieved good performance in both Market to Duke and Duke to Market unsupervised domain adaptive tasks, which is significantly better than the related algorithms.
Laser & Optoelectronics Progress
- Publication Date: Dec. 25, 2022
- Vol. 59, Issue 24, 2420001 (2022)
Reconstruction Technology for Three-Dimensional Emission Computerized Tomography Based on Linear Interpolation Theory
Mingzhe Li, Jia Wang, Dangjuan Li, Mi Zhou, Junxia Cheng, and Shenjiang Wu
Three-dimensional (3D) emission computerized tomography (ECT) is superior to the "slicing" process of the traditional two-dimensional ECT technology and reconstructs the test zone as a whole. As the projections are not limited to the same horizontal plane, 3D ECT can solve the problems of limited detection positions and assembly errors encountered during actual combustion tests, which is necessary for 3D combustion imaging and testing. In this study, a 3D weight matrix calculation algorithm was developed using a mathematical model of camera imaging in a 3D space and the linear interpolation theory to reduce the data amount and improve the calculation accuracy and efficiency of the weight matrix for 3D ECT. The accuracy of the algorithm was verified via numerical simulations using an algebraic reconstruction algorithm for tomography reconstruction. An ECT system with multiple cameras was established, and the proposed algorithm was used to reconstruct the combustion flame. The results have a significant reference value for improving the accuracy and efficiency of tomographic reconstruction. Three-dimensional (3D) emission computerized tomography (ECT) is superior to the "slicing" process of the traditional two-dimensional ECT technology and reconstructs the test zone as a whole. As the projections are not limited to the same horizontal plane, 3D ECT can solve the problems of limited detection positions and assembly errors encountered during actual combustion tests, which is necessary for 3D combustion imaging and testing. In this study, a 3D weight matrix calculation algorithm was developed using a mathematical model of camera imaging in a 3D space and the linear interpolation theory to reduce the data amount and improve the calculation accuracy and efficiency of the weight matrix for 3D ECT. The accuracy of the algorithm was verified via numerical simulations using an algebraic reconstruction algorithm for tomography reconstruction. An ECT system with multiple cameras was established, and the proposed algorithm was used to reconstruct the combustion flame. The results have a significant reference value for improving the accuracy and efficiency of tomographic reconstruction.
Laser & Optoelectronics Progress
- Publication Date: Dec. 10, 2022
- Vol. 59, Issue 23, 2320002 (2022)
Performance Analysis of SCB-Spinal Code in Free-Space Optical Communication
Jing Zhang, Wenqing Li, Yang Cao, Xiaofeng Peng, and Fangding Du
Aiming at the problems of high decoding complexity and unequal error protection of Spinal codes in free space optical communication, a SCB-Spinal code scheme is proposed, in which cyclic redundancy code (CRC), Spinal codes, and BCH codes are concatenated. Premature termination of the segmented CRC check reduces the amounts of decoding calculations and the complexity of decoding, and also cascades the BCH code at the tail to perform error correction protection for error-prone tail information. The simulation results show that the SCB-Spinal code effectively reduces decoding complexity and obtains better bit error rate performance under different turbulence intensities. Compared to the traditional Spinal code, the SCB-Spinal code is approximately 62% lower with respect to signal-to-noise ratio and turbulence. In addition, it exhibited a performance rate of 0.04-0.17 bit/symbol. With respect to moderate and strong turbulence intensities, the complexity under SCB-Spinal code is reduced by 50%-60%, facilitating efficient application of Spinal code in free space optical communication. Aiming at the problems of high decoding complexity and unequal error protection of Spinal codes in free space optical communication, a SCB-Spinal code scheme is proposed, in which cyclic redundancy code (CRC), Spinal codes, and BCH codes are concatenated. Premature termination of the segmented CRC check reduces the amounts of decoding calculations and the complexity of decoding, and also cascades the BCH code at the tail to perform error correction protection for error-prone tail information. The simulation results show that the SCB-Spinal code effectively reduces decoding complexity and obtains better bit error rate performance under different turbulence intensities. Compared to the traditional Spinal code, the SCB-Spinal code is approximately 62% lower with respect to signal-to-noise ratio and turbulence. In addition, it exhibited a performance rate of 0.04-0.17 bit/symbol. With respect to moderate and strong turbulence intensities, the complexity under SCB-Spinal code is reduced by 50%-60%, facilitating efficient application of Spinal code in free space optical communication.
Laser & Optoelectronics Progress
- Publication Date: Dec. 10, 2022
- Vol. 59, Issue 23, 2320003 (2022)
Fitting Model of Laser-Induced Damage Threshold for Optical Elements with Periodic Surface
Yuan Li, Junhong Su, Junqi Xu, Lihong Yang, and Guoliang Yang
The laser-induced damage threshold of optical elements is a key indicator for measuring laser damage resistance. Optical elements with periodic surfaces have advantageous optical characteristics and potential applications in high-power laser systems. It is important to determine the laser-induced damage threshold accurately. In this paper, the main sources of uncertainty are analyzed, a calculation formula for the uncertainty of laser-induced damage threshold is established, and the processing methods for reducing the uncertainty of laser-induced damage threshold are provided. The results show that when the spot radius is 400 μm, the error is 10 μm, the laser energy error is 5%, and the uncertainty introduced by energy density is zero. Then, the main factors contributing to the uncertainty of the laser damage threshold are the uncertainty of the damage probability and that of the linear fitting. The precision of the laser-induced damage threshold can be further improved by increasing the number of measurements for each energy level. The laser-induced damage threshold of optical elements is a key indicator for measuring laser damage resistance. Optical elements with periodic surfaces have advantageous optical characteristics and potential applications in high-power laser systems. It is important to determine the laser-induced damage threshold accurately. In this paper, the main sources of uncertainty are analyzed, a calculation formula for the uncertainty of laser-induced damage threshold is established, and the processing methods for reducing the uncertainty of laser-induced damage threshold are provided. The results show that when the spot radius is 400 μm, the error is 10 μm, the laser energy error is 5%, and the uncertainty introduced by energy density is zero. Then, the main factors contributing to the uncertainty of the laser damage threshold are the uncertainty of the damage probability and that of the linear fitting. The precision of the laser-induced damage threshold can be further improved by increasing the number of measurements for each energy level.
Laser & Optoelectronics Progress
- Publication Date: Dec. 10, 2022
- Vol. 59, Issue 23, 2320001 (2022)
Topics
© Copyright 2018-2021 | Chinese Laser Press. All Rights Reserved 沪ICP备15018463号-20