Development Of Modified Deep Learning Techniques For Effective Retrieval Of Images

Abstract

Technological advances have resulted in generation of large newlinevolume of images.It is advantageous that there is quantitative newlinerepresentative for every scenario. However managing such voluminous newlineset of images and processing these images to arrive at conclusion is newlinechallenging . Consistently researches are contributing in this area of newlineimage interpretation / classification / identification / retrieval . Even a newlinedomain specific technique has not been standardized. Hence research is newlinecontinued in this area. newlineIn this research work, developing classification techniques to newlineact as as the foundation for content based image retrieval is focused. Three newlinesets of image database are considered. Two datasets are obtained from newlineweb and a thermal potato dataset is created for the research work. newlineInitially the work began with conventional model of newlineclassification where two different models one for feature extraction and newlinethe other for classification is used. newlineImpact of feature extraction techniques and classifiers on image newlineclassification his studied in terms of accuracy, sensitivity and specificity. newlineTo overcome the weakness in appropriate feature selection, the newlineparadigm then shifted to conventional neural networks. Feasibility of newlinevarious conventional neural networks namely Alexnet, Darknet, Dense newlinenote, Resnet and VGG net for image classification is studied. Also the newlinevii newlineimpact of each layer, number of strides, filters, fully connected network newlineon the performance of image classification is understood. newlineHaving understood the above, modifications in the existing newlinearchitectures of alexnet ,darknetand densenet are made. The modified newlinenetwork are then subjected to training from the standard database. Then newlinetransfer learning is used for classification of the images from three newlinedifferent data sets used in the research work. Performance is measured in newlineterms of sensitivity, specificity and accuracy newline