Innovative Approach to Duplicate Image Detection: Leveraging Graph Neural Networks and Spectral Clustering for Optimal Performance

Visual information in smart applications consumes extensive storage due to its high density and frequent replication, leading to increased retrieval complexity and latency. Existing duplicate image detection methods still rely on manual inspection and face copyright concerns, creating the need for a more efficient automated solution. To address these issues, this study proposes an Optimised Locally Inspired High-Level Graph Convolutional Neural Network (OLI-HGCNN) model for accurate image duplication detection. Four benchmark datasets-D-Rep, CoMoFoD, CASIA CMFD, and COVERAGE are employed to collect input images, containing both original and duplicate samples. A density-based feature extraction module is first applied to generate mask images that assist in identifying authentic samples. Subsequently, a Locally Scaled Spectral Clustering model is used to detect identical images based on the generated masks. During the detection stage, a High-Level Graph Convolutional Neural Network extracts global dependencies among images, where high-level node representations capture more informative and discriminative features. The Parametric Rectified Linear Unit (PReLU) activation function enhances computational flexibility. The proposed framework finally classifies images as non-duplicate or duplicate. Experimental results demonstrate superior performance with a deduplication rate of 0.38 bytes and a complexity rate of 0.13 bits.