ISSN : 2319-7323

INTERNATIONAL JOURNAL OF COMPUTER SCIENCE ENGINEERING

ABSTRACT

Title	:	Feature Enhancement of Retinal Images in Deep Convolution Pipeline, Its Reconstruction, Extraction and Analytical Evaluation
Authors	:	Mahua Nandy Pal, Minakshi Banerjee
Keywords	:	Deep convolution; Transposed convolution; Feature map; PSNR; AMBE;
Issue Date	:	Nov-Dec 2020
Abstract	:	Purpose: Retinal image analysis and segmentation gives us information regarding different ocular and cardio-vascular diseases. Retinal image enhancement is the important pre-requisite of retinal image analysis and segmentation. In recent era, deep network has been extensively used in different research fields. In this paper, we attempt to exploit the application of deep convolution in retinal image enhancement and evaluate it against traditional enhancement techniques which are most prevalently used for retinal image enhancement. Method: We have utilized successive convolution and transposed convolution to enhance features of a retinal image. Feature maps are reconstructed from deep convolution layers and enhanced image is extracted successfully. We have evaluated the quality of the extracted enhanced image, with respect to three traditional enhancement techniques as well as different combinations of them. These traditional techniques are applications of contrast limited adaptive histogram equalization (CLAHE), adaptive gamma correction (AGC) and Tophat transformation. We evaluated all the methods on the basis of image quality assessment (IQA) metrics. Both statistical error based IQA metrics and visual information based IQA metrics are evaluated for this purpose. The metrics are peak signal to noise ratio (PSNR) and absolute mean brightness error (AMBE). Results: Deep convolution enhanced retinal images are reconstructed and extracted successfully and compared with other enhancement schemes. In most of the experiments deep convolution based enhancement performs the best among all schemes in terms of both types of IQA metrics. Conclusion: Deep convolution based enhancement can be used prior to retinal image segmentation and analysis instead of single or different arbitrary combinations of more than one single enhancement schemes for better precision in the relevant fields.
Page(s)	:	375-385
ISSN	:	2319-7323
Source	:	Vol. 9, No. 6