Search

Search Results

• USING FAST PROTOTYPING FACILITIES FOR IMPLEMENTATION OF A CONVOLUTION NEURAL NETWORK ON A FPGA

  V. V. Bakhchevnikov , V. A. Derkachev , A. N. Bakumenko

  2020-10-11

  Abstract ▼

  Research in the field of artificial intelligence is carried out with increasing interest every
  year. The fields of application of artificial intelligence are quite extensive: automation, analysis of
  a large amount of data, smart home technology, machine vision, etc. Artificial intelligence technologies
  are based on the use of artificial neural networks, which are based on the principles of
  the animal nervous system. In this case, the actual issue is the implementation of artificial neural
  networks on various software and hardware platforms: programmable logic integrated circuits of
  the FPGA type (Field Programmable Gate Array), on special purpose integrated circuits (Application-
  Specific Integrated Circuit, ASIC), GPU, CPU etc. FPGA performs best in low-power mobile
  systems. ASIC demonstrates the highest performance at a fairly high development cost.
  The problem of rapid prototyping of projects based on the use of artificial neural networks for
  FPGAs using conventional methods (using HDL languages, HDL encoders, graphic programming)
  is that either such a project is complex and time-consuming to debug (HDL languages), or
  the resulting code is not optimal (HDL encoders), or the duration of the project development and
  the complexity of reconfiguring the neural network (graphical programming) are high. Therefore,
  in the framework of this work, an effective method for designing fully connected and convolutional
  neural networks for their implementation on FPGAs using the Xilinx System Generator for DSP
  and Matlab / Simulink package is considered. Artificial neural networks generated in this way are
  easily reconfigurable and allow solving the following problems: image recognition, optimal filtering
  (for example, for problems of subsurface radar).

• MODEL OF SCATTERING OF RADAR SIGNALS FROM UAV

  V. A. Derkachev

  2021-07-18

  Abstract ▼

  In this article, a model of scattering of radar signals from unmanned aerial vehicles (UAVs)
  of a multi-rotor type is considered for the formation of training data for a neural network classifier.
  Recently, there has been an increased interest in studying the issue of detecting and classifying
  small unmanned aerial vehicles (UAVs), which is associated with the development of the UAV
  range in sales and production. In addition to the development of UAVs, an increase in the performance
  of computers made it possible to create classifiers using new neural network algorithms.
  This model generates radar images obtained as a result of the reflection of a chirp radar signal
  from an unmanned aerial vehicle, taking into account the configuration, characteristics, current
  location and flight parameters of the observed object. When calculating the reflected signal, the
  angles of rotation of the UAV (pitch, roll and yaw), flight speed, size and location of propellers in
  the current UAV configuration are taken into account. The resulting model can be useful for the
  formation of a training set of a classifier of unmanned aerial vehicles of a multi-rotor type, builtusing convolutional neural networks. The need to use a model that generates data for a neural
  network is due to the requirement for a large number of training and verification samples, as well
  as a wide variety of configurations of unmanned aerial vehicles, which greatly increases the complexity
  and cost of creating a training dataset using experimental measurements. In addition to
  training the neural network itself, this model can be used to assess the detection and classification
  of various types of multi-rotor UAVs, in the development of a specialized radar station for detecting
  this type of objects.

• CLASSIFICATION OF RADAR IMAGES OF MULTI-ROTOR UNMANNED AERIAL VEHICLES USING THE YOLO11 ALGORITHM

  V.А. Derkachev

  171-180

  2025-07-24

  Abstract ▼

  This article discusses a classifier of radar images of unmanned aerial vehicles based on a neural network built on the YOLO algorithm version 11. Solving the problem of detecting and classifying unmanned aerial vehicles has become one of the priority tasks at present. The increase in the number of modifications of unmanned aerial vehicles greatly complicates the use of statistical classification methods, which requires the use of new approaches to solving the classification problem. The development of neural network methods, simultaneously with an increase in the performance of computers for training, on the one hand, and embedded solutions, on the other, allows for the classification of aircraft using radar images in real time. The use of the YOLO11 algorithm allows, in addition to determining the class of the target, to estimate the range to the observed object. The use of radar images is justified due to the fact that visual observation is not always possible due to difficult weather conditions and darkness. To train the neural network, it is proposed to use a set of radar images obtained using the author's model of data generation with an arbitrary configuration of unmanned aerial vehicles. The neural network of the Detection YOLO11s class (9.4 million parameters) was trained on a sample of radar images of two classes, a total of 8192. As a result of training, an accuracy of 0.99 was obtained for classification in 2 classes of objects (on test model data). Tests were conducted using natural data taken using the TI IWR1642 millimeter-range radar system, as a result of which error-free classification of objects on a small sample was achieved

• MODEL OF ALGORITHM FOR STREAMING LABELING OF WIDE FORMAT IMAGES

  А.N. Bakumenko, V. А. Derkachev, V.V. Bakhchevnikov, V.T. Lobach

  2024-05-28

  Abstract ▼

  This article proposes a wide-format image processing algorithm for use in systems operating in real
  time with a high-speed video data stream. The issue of image preprocessing, its clustering, segmentation
  and labeling is of particular importance for systems for processing high-resolution video streams in real
  time. In addition, when implementing such algorithms, there is an urgent issue of minimizing the cost of
  computational resources of programmable logic integrated circuits (FPGAs), on which the direct deployment
  of streaming image processing algorithms takes place. Minimal resource consumption is ensured by
  single-pass marking algorithms, which eliminate the need for image buffering, which is especially important
  when processing high-resolution wide-format images. However, when implementing a single pass
  of an image through the processing system, many additional markers may be created that are subject to
  further combining, especially when analyzing images with high resolution. The additional markers created
  require an increase in the requirements for the number of usable memory cells on the FPGA. The algorithm
  for streaming high-resolution wide-format images described in the article makes it possible to label
  high-resolution streaming video images, reducing the likelihood of creating additional tags that need to be
  further combined. The essence of improving the algorithm relative to the standard one-pass one is to add
  additional elements to the scanning mask, which avoid the appearance of different labels corresponding to
  the same object, which allows, with a minimal increase in the amount of memory used on the FPGA, to
  avoid duplication of labels and overuse of device memory. The algorithm was simulated for implementation
  on an FPGA using the Xilinx System Generator for DSP tool in conjunction with the Matlab Simulink
  environment for model-based design (MBD). The results of the algorithm are presented on images obtained
  from a high-speed linear camera TELEDYNE DALSA LA-CC-04K05B-00-R using the Integre
  Technologies LLC FMC-200-A mezzanine, as well as the Xilinx ZYNQ Ultrascale+ MPSoC ZCU106 development
  board.

• IMPLEMENTATION OF A MATCHED FILTER IN THE FREQUENCY DOMAIN ON FPGA

  V.V. Bakhchevnikov, V.А. Derkachev, А. N. Bakumenko

  2023-06-07

  Abstract ▼

  The use of filters matched to radio signals is quite common in radar, which helps to improve
  range resolution, as well as in communication systems and many other radio engineering systems,
  allowing you to increase the output signal-to-noise ratio (SNR). Designing digital devices on field
  programmable gate array (FPGA) allows us to configure them quite flexibly and create prototypes
  of radio engineering systems for further implementation of DSP algorithms, on applicationspecific
  integrated circuits (ASIC ), GPU, CPU, etc. FPGA digital devices are most used in low
  power mobile systems, while ASICs show the highest performance with high development costs.
  In this work, special attention is paid to the design and implementation of a filter matched to a
  complex chirp signal in the frequency domain on an FPGA using the Xilinx System Generator for
  DSP library of Matlab/Simulink. The results of the hardware-software model operation are presented
  in paper both for a single point object and for three point objects with different sampled
  delays. The dependence of the output on the input SNR for a linear and quadratic envelope detector
  is shown. The analytical curve SNROUT(SNRIN) is compared with the curve obtained using the
  developed hardware-software model implemented on the FPGA. The paper shows the benefits of
  using Xilinx System Generator for rapid prototyping of DSPs on FPGAs, and it provides an analysis
  of the used FPGA resources for the developed matched filter.

• MULTI-ROTOR UAV CLASSIFIER

  V.А. Derkachev, V.V. Bakhchevnikov, А.N. Bakumenko

  2023-06-07

  Abstract ▼

  This article discusses a classifier of radar signals reflected from unmanned aerial vehicles
  (UAVs), based on neural networks. In the proposed classifier, for the formation of training data, a model
  of scattering of radar signals from UAVs is used. Recently, the demand for UAV classification has
  been quite high due to a significant increase in the number of models and sales of these devices. Increasing
  the computing power of processors and the development of the theory of neural networks allows you
  to create new types of classifiers. When using models, it is possible to create a set of training data that is
  acceptable for training a classifier neural network. The convolutional neural network of the classifier is
  trained using radar images obtained using the proposed model of scattering of radar signals from
  UAVs. The resulting radar images are modeled taking into account the UAV orientation angles relative
  to the UAV normal coordinate system, flight speed, and various propeller parameters of the simulated
  UAV. To form training data, in addition to the signal structure, white noise of a certain configuration is
  added, which helps to increase the diversity of training samples to improve the learning ability of the
  convolutional neural network. The use of data obtained using the model for training a neural network is
  due to the need to use a large number of training samples with various UAV movement parameters,
  such as height, speed, direction, orientation in space, as well as a wide variety of possible configurations
  of unmanned aerial vehicles: tricopter (three propellers), quadcopter (four propellers), hexacopter
  (six propellers), or octocopter (eight propellers). which complicates the use of experimental data to
  create classifiers of this type.

• CLASSIFIER OF IMAGES OF AGRICULTURAL CROPS SEEDS USING A CONVOLUTION NEURAL NETWORK

  V. A. Derkachev, V. V. Bakhchevnikov, A. N. Bakumenko

  2020-11-22

  Abstract ▼

  This article discusses the creation of a convolutional neural network architecture that classifies
  images of crops (in particular wheat) for subsequent use in an optical seed separator (photo
  separator). Interest in the design of neural networks for classifying images has recently increased
  significantly, which is associated both with the development of the theory of deep neural networks
  and the increased computing power of desktop computers, as well as the transfer of computing to
  graphic processors. The aim of the article is to develop the architecture of a neural network that
  allows the separation of the input flow of wheat seeds into two classes: “good” seeds and “bad”
  (with defects in shape and color) seeds. The architecture of the resulting neural network is convolutional,
  because, unlike a fully connected one, this class of neural networks is within certain limits
  immune to changes in the scale and angle of rotation of objects in the input data. In the work,
  for the formation of training, validation and test samples, seed images obtained using a household
  camera were used, which negatively affected the results of training and testing the neural network
  regarding the possible result of application in a real photo separator. The architecture of the developed
  neural network is preliminarily optimized for use on FPGAs, however, in the considered
  case, the transition from the values of weighting factors from the data type from a floating point to
  an integer type has not been made, which can lead to a decrease in the accuracy of the neural
  network, while significantly reducing the amount of resources FPGA. Application of the proposed
  architecture allows one to obtain a fairly accurate estimate of classified wheat seeds from verification
  and test data sets.