Loading...

Design and Implementation of Infrared Imaging Array and Image Processing using Machine Learning

Amiri, Mohammad Javad | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57987 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Fakharzadeh, Mohammad
  7. Abstract:
  8. In this project, the physics of the thermal infrared spectrum was first studied, focusing on thermal modeling of hidden objects, and field experiments were conducted. Then, with the main goal of the project, which was to design and build an economical thermography system to be used as a complement to other scanners such as millimeter wave or X-ray, the main parameters of thermal infrared image detectors (FPA) available on the market were reviewed and the appropriate option was selected and purchased. In the next step, a comprehensive study was carried out on the specifications and performance of the selected sensor, and the image reconstruction mechanism was described and implemented in detail. Then, to ensure the sensor's performance, the image reconstruction process was implemented on an ARM series test board, and the initial image was reconstructed. After ensuring the correct operation of the sensor and considering the project requirements, the conceptual design of the required boards was carried out, and then their schematics and layouts were designed. The designed boards consisted of three main units: sensor board, power board, and main board or microcontroller. These boards were ready for use after design, printing, assembly, and functional testing. The selection of an economical sensor brought challenges such as low resolution and noise in the signal. In order to overcome these hardware limitations, dedicated preprocessing and super-resolution methods were designed and implemented. These methods included an innovative chain including automatic multi-frame weighting, multi-sensor data fusion in a quad array, and estimation with the Lucy-Richardson method, which were presented in the form of a signal processing chain. Also, a chain of classical image processing techniques in the spatial and frequency domains was implemented to improve the quality of the images. For example, in the multi-frame method, the structural similarity index (SSIM) value was improved from 0.24 to 0.38. Also, in the multi-sensor method, the SSIM value was improved from 0.6 to 0.82. After the improved image reconstruction, object recognition algorithms were implemented using deep AI networks such as YOLO and Faster RCNN. The implementation of this step required a database containing comprehensive practical data from different conditions. For this purpose, a data collection platform including 20 prohibited objects with location labels was designed and implemented. Finally, more than 5000 samples of different people with various temperature conditions, gender, and clothing were collected. After training the algorithms, the recognition accuracy of YOLO reached 81.16% and the accuracy of FRCNN reached 94.89%, indicating the high efficiency of these methods in the defined conditions
  9. Keywords:
  10. Thermal Infrared Images ; Thermography ; Multisensory Systems ; Data Fusion ; Deep Learning ; Focal Plane Area (FPA) ; Image Processing

 Digital Object List

 Bookmark

No TOC