Novel Deep Learning Technique to Improve Resolution of Low-Quality Finger Print Image for Bigdata Applications

High-resolution mages are highly in demand when they are utilized for different analysis purposes and obviously due to their quality aesthetic visual impact. The objective of image super-resolution (SR) is to reconstruct a high-resolution (HR) image from a low-resolution (LR)image. Storing, transferring and processing of high-resolution (HR) images have got many practical issues in big data domain. In the case of finger print images, the data is huge because of the huge number of populations. So instead of transferring or storing these finger print images in its original form (HR images), it cost very low if we choose its low-resolution form. By using sampling technique, we can easily generate LR images, but the main problem is to regenerate HR image from these LR images. So, this paper addresses this problem, a novel method for enhancing resolution of low-resolution fingerprint images of size 50 x 50 to a high-resolution image of size 400 x 400 using convolutional neural network (CNN) architecture followed by sub pixel convolution operation for up sampling with no loss of promising features available in low-resolution image has been proposed. The proposed model contains five convolutional layers, each of which has an appropriate number of filter channels, activation functions, and optimization functions. The proposed model was trained using three publicly accessible fingerprint datasets FVC 2004 DB1, DB2, and DB3 after being validation and testing were done using 10 percent of these fingerprint data sets. In terms of performance measures like Peak Signal to Noise Ratio (PSNR), Mean Squared Error (MSE), Structural Similarity Index (SSIM) and loss functions, the quantitative and qualitative results show that the proposed model greatly outperformed existing state-of-the-art techniques like Enhanced deep residual network (EDSR), wide activation for image and video SR (WDSR), Generative adversarial network(GAN) based models and Auto-encoder-based models.