A Comparative Study about Content-Based Image Retrieval using Features from Deep Convolutional Neural Networks

"Content-Based Image Retrieval (CBIR)" is a popular approach that is frequently used to handle the image retrieval problem. Despite the fact that the CBIR approach has seen tremendous improvements, the fundamental challenge of finding and retrieving pertinent photos with high levels of accuracy remains. A lot of demand exists for the creation of efficient CBIR systems that utilise modern techniques. Convolutional neural networks (CNN) were used, which led to a substantial paradigm shift in machine learning and outstanding performance. This study's major objective is to ascertain whether deep CNN models can be employed to boost picture retrieval effectiveness. In a range of image-processing tasks, such as object identification, classification, and clustering, among others, CNNs excel very well. This work proposed the CBIR methodology, which implements the extraction of features by using a range of pre-trained CNN models. For this experimental research project, radiology and meteorological datasets were used. The findings demonstrated that pre-trained deep CNN models' learnt features may be quickly applied to image retrieval tasks and outperform hand-crafted features utilized in conventional CBIR techniques in terms of performance. A CBIR framework using features from a "pre-trained deep CNN model (VGG16)" was provided in accordance with the study's findings. Put the recently created CBIR algorithm through its paces using a "multi-class dataset" made up of images connected to the weather. The innovative CBIR technique resulted in an average retrieval precision of 86.73%. The standard CBIR design, which relies on manually created features (such color, texture, and form attribute), has produced an average retrieval precision of 73.25% in comparison to this. The unique methodology outperformed the traditional method, which showed a 13.48% larger gain in average precision.