PRIVACY-PRESERVING CBIR SYSTEM USING SIAMESE TWIN NETWORK WITH SEGNET ARCHITECTURE-BASED HIGH-LEVEL REGION DETECTION

Abstract

The work introduces a Content-based image retrieval (CBIR) approach that can preserve the privacy content of the picture using two deep learning architectures namely Siamese twin network and SegNet architectures. In this approach, the pictures that are uploaded to the cloud are initially separated into two level components namely low and high. The low-level components are encrypted using a block-based permutation approach to preserve the picture privacy content. The resultant image is uploaded to the cloud, where the cloud environment uses a SegNet architecture to segment the high-level components. The high-level component and the low-level encrypted regions are merged to extract features. The SegNet architecture results in a segmentation accuracy, recall, and precision of 98.14%, 96.74%, and 97.63% respectively when evaluated using the Corel-10k dataset. The descriptors are then collected from the merged image and clustered utilizing a recursive tuneable clustering approach. During the retrieval process, the Siamese network is utilized to match the selected leader and followers estimated by the clustering algorithm. The recursive tunable clustering approach reduces the complexity during the retrieval process. The suggested CBIR system was evaluated utilizing the scale such as time complexity and mean average precision (mAP) with the datasets namely Corel-10k and Inria Holiday databases. The proposed CBIR system results in a mAP of 69.27% and 64.53% when evaluated using the Corel-10k and Inria Holiday dataset respectively which is higher than similar recent CBIR systems.