A novel de-noising algorithm and multi-edge enhancement for middle wave infrared thermography

  • Gloria (Mingyu) Liao

    Student thesis: Doctoral Thesis

    Abstract

    An infrared (IR) camera is one of the common tools to take images for thermography analysis in various fields, including medical diagnosis. The field of IR imaging normally requires vast amounts of data to be stored and transmitted. In addition, IR images are susceptible to various types of noise, as the noise sources are distributed almost everywhere. Therefore, image enhancement and de-noising techniques are of utmost importance to improve the quality of IR images and meet the requirements from applications. This research presents a novel de-noising algorithm and multi-edge enhancement for middle wave infrared thermography. Object recognition using IR images captured with a Middle Wave Infrared (MWIR) camera is studied. An experimental study is conducted to identify the capability of MWIR thermography in distinguishing an object from its background with varying emissivity and temperature levels. A histogram equalization approach is devised to enhance blurry MWIR images. Useful methods to recognize the edges and the inner characteristics of MWIR images as well as to remove noise are also proposed. A program developed based on the MATLAB Image Processing Toolbox (IPT) is used to enhance the MWIR images of the target object. In addition,, a novel image de-noising algorithm based on Cohen-Daubechies-Feauveau wavelets with low-pass filters of the length of 9 and 7 (CDF 9/7) wavelets combined with Set Partition in Hierarchical Trees (SPIHT) coding algorithm and entropy coding techniques is proposed in image processing. This is to demonstrate that choice of decomposition level plays a very important role in achieving superior wavelet compression performances. Image quality is assessed objectively by parameters of compression ratio, peak signal-to-noise ratio (PSNR) and mean structural similarity index (MSSIM).In this research, it was first observed that object recognition becomes difficult when the emissivity difference is within 0.01 and when the temperature difference between the object and the background for the MWIR system is within 1oC. The experimental results demonstrate that the histogram equalization method is able to expand the dynamic range expansion of image gray levels so that the output image becomes clear and visible for object recognition, as an MWIR camera has been employed to capture images with close temperature and emissivity between the object and the background. A number of image processing methods is developed in this research using the MATLAB Image Processing Toolbox (IPT) for the function of “dilate” and “erode” processing, thinning processing as well as “imrotate” and canny edge detection to recognize the object contour, internal characteristics and to remove noise. In addition, from the experimental results from several tests of IR images by using the de-noising algorithm, under the noisy condition with σ=0.2 and density=20% cases, the mean square error (MSE) of the proposed method decreases by 83%; while PSNR and MSSIM increase by 14% and 67% with the same condition, respectively. The proposed threshold is applied to our compression methods. The choice of decomposition level plays a very important role in achieving superior wavelet compression performance. The proposed method provides a better compression ratio approximately 88%, highest PSNR values and the performance of MSSIM measurements. It is achieving as overall correct recognition rate as 99.30 % value for the IR images, and this is more suitable for this category of images. We have also shown that CDF 9/7 wavelet transform and compression method provides effective results which are useful for the de-noising of IR image. The possibility of better results can be achieved in image segmentation and image pre-processing for noise removal and can be further extended by the use of wavelet transform applied for image decomposition. Besides, image quality is assessed objectively by parameters of compression ratio, peak signal-to-noise ratio (PSNR), and mean structural similarity index (MSSIM).
    Date of Award2014
    Original languageEnglish
    SupervisorXu Huang (Supervisor) & Dharmendra Sharma AM PhD (Supervisor)

    Cite this

    '