Optical Memory and Neural Networks最新文献

This study investigates the spatiotemporal dynamics of broad-area vertical-cavity surface-emitting lasers (VCSELs) under external optical injection. Using a semiconductor-adapted Maxwell-Bloch model, we demonstrate that weak optical injection effectively suppresses modulation instability—a major constraint to achieving coherent VCSEL emission. Through linear stability analysis and numerical simulations, we found and showed the dependence of spatial patterns on pump current and injection amplitude. Our results reveal that controlled optical injection transforms chaotic emission into ordered structures, including stripes, hexagons, labyrinths, and their hybrid forms, with the pattern scale governed by the wavenumber of maximum growth increment. Notably, reducing the laser aperture size promotes the formation of defect-free patterns. These findings offer key insights for stabilizing VCSEL emission and leveraging self-organized patterns for advanced photonic applications, such as optical computing and on-chip communication systems.

In this paper, we consider the beam diffraction on quantized non-paraxial optical elements. We show that the formation of diffraction orders depends on both the numerical aperture and the focal length. In non-paraxial propagation, the main energy falls on the first diffraction orders, and it is also possible to eliminate a larger number of orders compared to the paraxial case. To simultaneously increase the number of longitudinal and transverse orders in the non-paraxial region, we propose using multifocal lenses as focusing elements. We expect that the results obtained can be used in (de-)multiplexing and signal detection problems.

A method of feature selection based on hyperspectral data classification is proposed. The features are selected using iterative training of Decision Tree classifiers for further construction of a thick cloud classifier based on spectral features. Classifiers are trained with different hyperparameters for different set of features by recursive elimination. Feature selection is determined by analyzing the correlation between the decrease in Gini impurity and classification accuracy, combined with mean feature importance. Classification model training is performed for three types of surfaces: ocean, vegetation, and urbanized areas. Feature selection improves the accuracy of Random Forest classifier by choosing a limited set of features from the NIR and the lower part of the SWIR spectrum ranges according to their importance.

Effective control of the spatial distribution of laser beam energy plays a key role in material processing technologies. The integration of diffractive optical elements (DOE) with digital tools opens new prospects for the control of laser processes, enabling precise energy distribution, adaptive adjustment of processing parameters, and improved predictability of outcomes. This paper explores digital approaches to the design and application of diffractive optics in laser processing. It describes methods for tuning laser beam parameters, the use of digital twins, and machine learning algorithms to enhance processing accuracy. The paper also thoroughly examines the potential applications of digital technologies for optimizing micro- and nano-processing, laser welding, and improving the quality and stability of materials during laser modification. The future development of these technologies, including integration with adaptive systems and optimization algorithms, is presented, opening new horizons for precision manufacturing processes.

The influence of vortex phase singularity and wave aberrations on the distortion of the pattern of formation of light curves is investigated. The Gillis transform was used to form the parametric type of the curve, and the diffraction optical elements were calculated using the Whittaker integral. It was found that the influence of vortex phase singularity mainly affects the change in the peripheral part of the amplitude of the formed curve and leading to the formation of zones with zero intensity in the central part. It is also shown that aberrations in optical systems distort the wavefront and degrading the image quality of formed light curves. Different types of aberrations introduce characteristic changes in the intensity distributions of light curves, which can be further used for detection and recognition of aberrations.

This paper explores the application of soil type classification algorithms (red clay, loamy soil, chernozem) in agricultural fields based on hyperspectral imaging data (400–1000 nm) acquired from an unmanned aerial vehicle (UAV). As part of the dataset preparation, the data were processed using a set of algorithms, from classical approaches to deep models. The ensemble NM3D-CNN achieved the best performance—0.874 weighted F1. Results of other approaches—PCA segmentation: 0.864, 1D-CNN: 0.842, SSFTT: 0.829, SVM: 0.764. These results indicate that Vis-NIR UAV-borne HIS enables reliable soil mapping, while spectral-spatial deep ensembles provide the strongest gains.

This paper presents a prototype application for generating synthetic data used to train neural network models for recognizing actions and interaction objects of an agricultural drone operator. This application has been implemented with the ability to customize various generation parameters. The obtained generation results are saved as images and text files in YOLO format.

We explore the diffraction of a plane wave by an opaque rectangular screen and show that at any on-axis distance from the obstacle plane, a multitude of diffraction fringes occur in a screen-parallel plane, characterized by a (negative) canonical backflow. Remarkably, the localized fringe recurrence interval is found to decrease with larger distance from the screen edge, with the first fringe of the canonical energy backflow departing from the edge with increasing distance from the optical axis. The energy backflow is shown to occur in the diffraction pattern areas characterized by subwavelength values of phase and amplitude modulation, i.e. where the local wave-vector expressed through the phase gradient is larger than the incident wave wave-vector.

Video surveillance utilization has developed significantly in sectors such as traffic monitoring, private institution protection, and landmark protection. Identifying an object in a captured surveillance image is difficult because of the poor excellence of the images. The quality of Low-Resolution (LR) images can be enhanced using an image Super-Resolution (SR) reconstruction method. Sophisticated Deep learning methods have been utilized in order to attain state-of-the-art performance in SR. Nevertheless, these techniques are typically prone to losing essential information and perform poorly on complex computations. To overcome these challenges, this research develop a deep Attention-Multistage Generative Adversarial Network (DA-MGAN) for image super-resolution and integrate SDR (Sparse Dense fusion R-CNN) with Non-Maximum Suppression (NMS) to enhance object detection accuracy in surveillance images. DA-MGAN is used to generate high-resolution surveillance images by utilizing attention mechanisms for improved feature extraction and integrating a multistage GAN that progressively enhances the image quality at each stage. After image resolution, Sparse Dense Fusion R-CNN (SDR) is used for object detection in super-resolved images to improve feature extraction through Sparse Dense Fusion. The R-CNN leverages these enhanced features to accurately detect and segment objects at the pixel level. Subsequently, Non-Maximum Suppression (NMS) was applied to improve localization by eliminating overlapping bounding boxes and minimizing false positives. This integrated method boosts overall detection precision and reliability in real time surveillance scenarios. The proposed model achieves a Super Resolution Error Rate (SRER) of 0.19%, a Bit Error Rate (BER) of 0.125%, a Packet Error Rate (PER) of 0.0990%, and a Deep End-to-End Image Metric (DEEIM) of 0.04963%, showcasing its superior performance. In contrast with existing methodologies, these results highlight the effectiveness of the suggested approach in reducing error rates and enhancing image quality metrics. As a result, these methods are ideally suited for real-time applications, particularly in high-resolution scenarios and object detection within surveillance systems.

Background estimation is an important part of many computer vision applications. However, it is a challenging task due to illumination changes, camouflage, occlusion, dynamic background, rain or snow fall, and shadows. In this paper, we propose a method to predict the background of videos recorded by fixed cameras. The proposed algorithm is unsupervised and online. It takes inspiration from the processing mechanisms of neural integrator circuits in recurrently connected networks. The neural activities of three distinct integrators, each responsible for processing a color channel in L * a * b color space, are updated according to the recent changes of the scene covering both spatial and temporal aspects. The maxima of the evolving activity distributions in color space are used to predict the background color value of each pixel. Evaluation results demonstrate that the proposed method outperforms several recent competitors on the Scene Background Initialization (SBI) and LASIESTA datasets, based on mean squared error (MSE) metrics.