Annual: 2019

EM052 »
The system of automated recognition of the boundaries of territories with a cloud filtering algorithm for satellite imagery.
📁Machine Learning
👤Sergey Postolnik
 ( Robotics center "Robot.ON")
📅Oct 08, 2019
Regional Final



👀 2432   💬 2

EM052 » The system of automated recognition of the boundaries of territories with a cloud filtering algorithm for satellite imagery.

Description

The system of automated recognition of the boundaries of territories with a cloud filtering algorithm for satellite imagery. By taking satellite images of the same terrain in different spectral channels and overlapping them with each other, determining the boundaries of certain territories and objects on it with the ability to predict parts that are covered by clouds and other interference that are between the satellites and the territory under study by using computer vision and neural networks based on FPGA. The main task is to get a vectorized map of the area with the exact contours of all the objects on it.

Demo Video

  • URL: https://youtu.be/TZNHeOffH4E

  • Project Proposal

    1. High-level Project Description

    In the process of developing the project, we talked to farmers, cartographers, designers, landscape designers, auditors, and field damage assessors after natural disasters and tried to solve some problems when working with images of territories.

    In the process of creating our project and further development, we will try to solve the following problems - to make a real vector map of agricultural fields with an indication of the boundaries and coordinates. Then it can be applied to the official cadastral map and compare the result. We will also try to make a relief grid of the area that designers can use. The alignment of the territories with the help of artificial intelligence will help cartographers, designers, designers to save time and will allow you not to do manual monotonous and repetitive work.

    Also in this project we will try using FPGA to solve the problem of clouds in the image by teaching the neural network an algorithm for constructing the boundaries of the territory in the absence of a part of the image.

    The system of automated recognition of the boundaries of territories with a cloud filtering algorithm for satellite imagery. By taking satellite images of the same terrain in different spectral channels and overlapping them with each other, determining the boundaries of certain territories and objects on it with the ability to predict parts that are covered by clouds and other interference that are between the satellites and the territory under study by using computer vision and neural networks based on FPGA. The main task is to get a vectorized map of the area with the exact contours of all the objects on it.

     

    2. Block Diagram

     

    3. Intel FPGA Virtues in Your Project

    Using parallel computing, presented by FPGA Terasic
    The DE10-Nano Kit (due to its high computational power) makes the boundary detection process faster and allows you to perform image processing at satellite orbital speed, which eliminates the possibility of the tracking device exiting the orbit before processing the previous image.
    FPGA also has such advantages as low power consumption and high speed of information processing, which can give a chance to try to launch a high-precision neural network in field conditions, i.e. implement the project in a portable device. Perhaps such a device will find its application in the military sphere.
    Recent studies in the field of neural networks have shown that they do a good job with many tasks related to the classification and processing of images, audio and video data. The dimension and computational complexity during the classification is so great that even powerful general-purpose CPUs do a poor job with computations. For full-fledged work with modern neural networks, powerful and as a result expensive GPUs (video cards) are used. The hardware implementation of the neural network (NN) largely depends on the effective implementation of a single neuron. FPGA-based reconfigurable computing architectures are suitable for the hardware implementation of neural networks.

    Advantages: 

     

    1) The use of FPGA (due to the high processing power) allows you to implement a device that can process images in high resolution without loss of quality

    2) Low power consumption allows you to achieve maximum autonomy and mobility of the platform, which makes it possible to process images using one device, which are obtained in different locations of the antennas

    3) Ability to switch from floating point calculations to fixed point calculations

    4) Provides the ability to send processed images over a LAN

    
     
    
     
    
     

    4. Design Introduction

      This development is aimed at eliminating the human factor in the process of building maps of the area with the exact boundaries of the objects that are located on it. The use of FPGA makes it possible to increase the accuracy of boundary construction and increase the speed of recognition.
      The project implemented a recognition system for the boundary lines of the Earth’s surface using Hough transforms and the high-level Tensorflow code. The entire software part of the project was implemented using Verilog. The developed device can be used not only to detect the boundaries of crops and forest stands, but also to recognize urban infrastructure. Using the well-known architecture of the neural network, as well as the availability of code for training it on an arbitrary data set, allows the device to be used by enthusiasts and is easily reprogrammed to fit their tasks.
    
     

    5. Function Description

    To implement the project, it was decided to use machine vision algorithms such as Hough algorithms, to search for lines in the image. In the process of searching for lines, the following steps can be distinguished: Translation of the image into gray gradients, changing the contrast and brightness to highlight the main lines, searching for all lines in the image, filtering lines. After that, the neural network algorithms search for dashed lines caused by obstacles in the form of clouds and guess the possible ends of the missed lines along the lines around and draws them on the map. Then the linear card is overlaid on top of the original image.
    

     

    6. Performance Parameters

    For the implementation of the project, all the resources of the Terasic DE10-Nano board were used to organize capacities aimed at highly accurate recognition of the boundaries of ground-based objects by implementing a neural network with the capabilities of using machine vision technologies to ensure highly productive work.

    Сompilation results:

    LE: 7031/113560

    Pins: 0/378

    Memory bits:115430/12492800

    DSP Blocks: 83/342

    Maximum frequency: 21 MHz

    Number of clocks to get the result for single frame: 382475

     

    
     
    
     
    
     

    7. Design Architecture



    2 Comments

    Bilal Zafar
    If the goal is to teach "the neural network an algorithm for constructing the boundaries of the territory in the absence of a part of the image", it's it better to do this on a CPU? What is the advantage of using FPGAs for this application? With a software-only approach you will have both ease of implementation and performance advantage.
    🕒 Jul 04, 2019 04:38 AM
    EM052🗸
    Recent studies in the field of neural networks have shown that they do a good job with many tasks related to the classification and processing of images, audio and video data. The dimension and computational complexity during the classification is so great that even powerful general-purpose CPUs do a poor job with computations. For full-fledged work with modern neural networks, powerful and as a result expensive GPUs (video cards) are used. The hardware implementation of the neural network (NN) largely depends on the effective implementation of a single neuron. FPGA-based reconfigurable computing architectures are suitable for the hardware implementation of neural networks.
    🕒 Jul 04, 2019 01:13 PM

    Please login to post a comment.