Efficient rotated and scaled digital image retrieval model using deep learning-based hybrid features extraction

• Published in: Multimedia tools and applications 2024-04, Vol.83 (12), p.34733-34758
• Main Authors: Hire, D. N., Patil, A. V., Charles, Priya
• Format: Article
• Language: English
• Subjects: Color; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Deep learning; Digital imaging; Feature extraction; Image enhancement; Image retrieval; Multimedia Information Systems; Queries; Query processing; Retrieval; Rotation; Scaling; Special Purpose and Application-Based Systems
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-023-17016-y

The previous couple decades have seen extensive study in digital image recovery. Content-based image retrieval finds intellectually and aesthetically similar images to the query image. Recent deep learning (DL) approaches enhanced CBIR automation accuracy. After improving accuracy, such systems have failed to overcome image threats such scaling and rotation. Present methods severely decrease scaled and rotated query image performance. The Rotation and Scaling aware Digital Image Retrieval (RS-DIR) system is our unique CBIR framework. The main objective of RS-DIR is to enhance query image retrieval for rotated and scaled images. We build a hybrid feature extraction technique for scaled and rotated images. Image pre-processing, hybrid feature extraction, and retrieval are all processes in the RS-DIR system. In a pre-processing step, the input query image is transformed into the standard size followed by either rotation or scaling threat. The hybrid feature extraction is performed by extracting the DL features and handcrafted (HC) features. We designed a modified 18-layer fast SqueezeNet DL model to extract the high-level image features automatically. However, retrieval efficiency for rotated and scaled input query images cannot be reached using solely high-level characteristics. Therefore, we extracted HC features and combined them with DL features in the RS-DIR model. The features such as color histogram, color auto-correlogram, color moments, and wavelet moments are extracted to build HC features. The experimental results show the efficiency of the proposed model compared to underlying solutions using different datasets.