Deep Learning using Spatial Distances, Normalized Coordinates, Scaling and Visual Words for Large Datasets

Abstract

Image Retrieval has become popular and crucial task and the number of digital images on (web servers) amplified, it became increasingly very difficult to classify and track images. Many methods have been used to make image exploration effective and reliable, such as search based on the file name, image tagging, etc. but none have proved a good idea to work in real scenario. Our proposed methodology applies deep learning using spatial distances, normalized coordinates, scaling and visual words for large data sets for retrieval of images with highest accuracy. The proposed methodology has three basic steps: the first is Content Analysis. The image is passed through coarser intervolving phase, second is CNN, third is RGB color evaluation, fourth is Retrieved feature vectors and fifth is results derivation. Proposed methodology was applied on the three famous datasets namely, Cifar-100, FTVL and Fashion. Experiments conducted on these datasets have shown outstanding results.