RS Invariant Image Classification and Retrieval with Pretrained Deep Learning Models

CBIR deals with seeking of related images from large dataset, like Internet is a demanding task. Since last two decades scientists are working in this area in various angles. Deep learning provided state-of-the art result for image categorization and recovery. But pre-trained deep learning models are not strong enough to rotation and scale variations. A technique is proposed in this work to improve the precision and recall of image retrieval. This method concentrates on the extraction of high-level features with rotation and scaling invariant from ResNet18 CNN (Convolutional Neural Network) model. These features used for segregation of images using VGG19 deep learning model. Finally, after classification if the class of given query image is correct, we will get the 100% results for both precision and recall as the ideal requirement of image retrieval technique. Our experimental results shows that not only our proposed technique outstrip current techniques for rotated and scaled query images but also it has preferable results for retrieval time requirements. The performance investigation exhibit that the presented method upgrades the average precision value from 76.50% for combined features DCD (Dominant Color Descriptor), wavelet and curvelet to 99.1% and average recall value from 14.21% to 19.82% for rotated and scaled images utilizing Corel dataset. Also, the average retrieval time required is 1.39 sec, which is lower than existing modern techniques.