MRTD (Minimum Resolvable Temperature Difference) is an important parameter for comprehensive evaluation of temperature resolution and spatial resolution of infrared imaging systems. It has become one of the necessary detection parameters for manufacturers of thermal imaging cameras. The traditional subjective MRTD parameter test method is gradually replaced by objective test methods due to its long test time and high labor cost. At present, the objective test method has developed the video MTF method and the photometric camera method, but both methods have their corresponding limitations. This paper proposes a new objective MRTD parameter test method based on CNN neural network. Firstly, the four-bar target image used to test the MRTD parameters is analyzed. It is concluded that the process of testing the MRTD parameters is essentially an image classification, which lays a foundation for the learning of CNN neural networks. Then the network model of CNN neural network interpretation of four-bar target image is expounded, and the accuracy of MRTD test results under different network models is analyzed. It was found that the network structure should not be complicated in the classification process of the four-bar target image. Based on the classic CNN neural network LeNet model, this paper proposes a CNN neural network suitable for four-bar target image classification problem by optimizing the convolution layer size, changing the activation function and adjusting the network structure. The experimental results show that the optimized CNN neural network improves the accuracy and repeatability of the MRTD parameter test.

Date Published: 12 March 2020
PDF: 7 pages
Proc. SPIE 11439, 2019 International Conference on Optical Instruments and Technology: Optoelectronic Measurement Technology and Systems, 1143910 (12 March 2020); doi: 10.1117/12.2544098

Published in SPIE Proceedings Vol. 11439:
2019 International Conference on Optical Instruments and Technology: Optoelectronic Measurement Technology and Systems
Jigui Zhu; Kexin Xu; Hai Xiao; Sen Han, Editor(s)