Architectural Formation with Deep Learning and Algorithmic Bindings for Cross-Domain Information Retrieval

Abstract

Efficient strategies for index search are crucial elements involved in categorizing and retrieving simple as well as complex image collections and libraries. In this paper, new algorithm is presented aimed at refining the selection of images to be clustered and more accurate identification of ROIs in many clustered objects.  The relations to other features are also expected to be provided, including the RGB image features and the other feature sets obtained with the use of Convolutional Neural Networks (CNNs) for achieving the scale invariance. Despite, GoogleNet and AlexNet and ResNet exist this algorithm has the deep feature and spatial data point of view for improving the image classification. Feature coefficient computation further enables the application of norms L1 and L2 on over the images of RGB. The ‘Scale invariance’ encompasses predicting the scaling of keypoints, computation of coefficients between two successive octaves along with expressions of virtual intra octave expressions. In the process of maxima selection, interpolation, non-maxima suppression, and cumulative thresholding the algorithm applies ROI detection. The presented multimodal approach significantly enhances the identification of objects particularly in a setting as depicted in this paper with high density of other similar objects. The color feature sets and CNN feature sets that are integrated in constructing the Bag-of-Words (BoW) model improve image indexation and image search. From the quantitative analysis, there is promising average precision (AP) and average recall (AR) when the presented algorithm is tested using data from Corel-10K, Tropical Fruits and Cifar-10 datasets.