Building a smart dynamic kernel with compact support based on deep neural network for efficient X-ray image denoising

ABSTRACT

Gaussian filtering is a successful computer operation vision to reduce noise and calculate the gradient intensity change of an image. However, it’s well known that in scale space context, the Gaussian kernel has some drawbacks, loss of information caused by the unavoidable Gaussian truncation and the prohibitive processing time due to the mask size. To give a solution to both problems, a new kernel family with compact support and its separable version were presented in the literature. The theoretical study of these kernels shows that the new family kernel is parameterised by a scale parameter and generated in such a way that fine scale structures are successively suppressed when the scale parameter is increased. Moreover, the scale parameter is increased, the image is blurred and details and border are removed. All these disadvantages are related to the static nature of these kernels. In this paper, we propose a smart kernel based on deep neural networks (dnn) to create a dynamic kernel with compact support called DSKCS. The parameter involved in the filtering process is calculated in real time and supervised by deep neural networks. The filter is continuously variable to detect, clean and avoid noisy areas of the image. Extensive experiments show that the proposed kernel can improve the classic kernel and presents a solution for its limitations related to its static nature. Furthermore, different metrics calculated illustrate our approach efficiency. As stated in the filtering performance, which reveals the highest PSNR and NCC with the metrics results (PSNR = 32.18, NCC = 0.95). Also, we recorded more than 0.89 for area under curves of the classification results using DBN-DNN technique.

KEYWORDS:

• Efficient noise reduction
• adaptive filtering
• compact kernel support
• deep neural network
• Covid19

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Notes on contributors

Zouhair Mbarki

Zouhair Mbarki was born in Tunisia, he received the master degree in Automatic from the Higher School of Sciences and Techniques of Tunis, in 2011 and the thesis degree in image processing in 2017. His research interests are focused on image processing, smoothing, restoration and segmentation.

Amine Ben Slama

Amine Ben Slamais a PhD Doctor in Biophysics and medical imaging from the University of Tunis El Manar, ISTMT. He is a member of research group in Laboratory of Biophysics and medical technologies at the same university. His research interests include signal and image processing, and deep learning methods.

Hassene Seddik

Hassen Seddik was born in 15 October 1970 in Tunisia, he has obtained the electromechanical engineer degree in 1995 and followed by the master degree in “signal processing: speaker recognition” and the thesis degree in image processing “watermarking using non-conventional transformations”. He has over 14 international journals papers and 65 conference papers. His domain of interest is: Audio-image and video processing applied in filtering, encryption and watermarking. He belongs to the CEREP research unit and supervises actually five thesis and 08 masters in the field.

Hedi Trabelsi

Hedi Trabelsiborn in 1975 in Tunis (Tunisia), he received the B.Sc. degree in Biophysics from the faculty of Sciences of Tunis, He is currently Professor at the higher institute of medical technologies of Tunisia University of Tunis El Manar, member of the Laboratory of biophysics and medical technologies. His research interests are focused on diffuse optical Tomography and physics simulation.