Open AccessA Comparative Approach on Classification of Images with Convolutional Neural Networks
Author(s) -
Ravikant Kholwal,
Shishir Maurya
Publication year - 2021
Publication title -
international journal of engineering and advanced technology
Language(s) - English
Resource type - Journals
ISSN - 2249-8958
DOI - 10.35940/ijeat.d2483.0410421
Subject(s) - computer science , convolutional neural network , architecture , artificial intelligence , convolution (computer science) , simple (philosophy) , pattern recognition (psychology) , image (mathematics) , noise (video) , distortion (music) , contextual image classification , network architecture , artificial neural network , machine learning , art , amplifier , computer network , philosophy , epistemology , computer security , bandwidth (computing) , visual arts
Image degradation, such as blurring, or varioussources of noise are common reasons for distortion happeningduring image procurement. In this paper, we will study in asystematical manner the efficiency of various ConvolutionalNeural Networks (CNN) approaches, in respects to the type ofarchitecture and optimization strategies, with two mainobjectives in mind. Firstly, we examine the CNN performance inclassifying clean images, with a dataset containing 8 classes andmore than 18,000 images, observing comparatively the obtainedresults from training on a standard architecture with thoseobtained from training on a hyper parameters fine-tuned networkand lastly, from training on a wider pre fine-tuned network.Secondly, training our model after a degradation function isapplied, and after analyzing the results, we propose an approachwhich will gently balance the efforts stemming from difficultarchitecture de-sign or adopting the best optimization decisionswith obtaining a satisfactory efficiency in a simple manner. Wehave offered a standard convolution architecture as a solution forclassifying images which are distorted, and our results suggestthat, departing from a simple design, with possible alterations ofhyper parameters and other optimizing routes, the efficiencycould massively increase.