A Deep Learning-Based Framework for Feature Compression and Similarity in Tattoo Recognition

Tattoo recognition is used in forensic and security applications, particularly in scenarios where conventional biometric modalities are unavailable or unreliable. Traditional approaches based on hand-crafted features and keypoint matching often show limited performance under variations in lighting, occlusion, and deformation. This work presents a deep learning-based framework that incorporates feature compression using convolutional autoencoders alongside hybrid similarity metrics for tattoo retrieval. The framework reduces the dimensionality of tattoo images while preserving essential structural and semantic information, combining cosine similarity with SIFT and ORB descriptors to support matching. The system was evaluated on a data set of 5000 tattoo images and showed consistent reconstruction quality and retrieval coherence. Although no direct comparison with existing methods was included, the results indicate that the approach is reliably effective in retrieving visually similar tattoos under varying conditions. The framework is intended as a modular baseline for future extensions, such as benchmarking and integration of alternative architectures.