Deep Reinforcement Learning for Image Hashing
| dc.contributor.author | Athira Musaliyath Dinesh 1NH20CE005; Dhanush Y J 1NH20CE010; Anusha Sai 1NH20CE019 | |
| dc.date.accessioned | 2025-05-31T10:13:10Z | |
| dc.date.available | 2025-05-31T10:13:10Z | |
| dc.date.issued | 2024 | |
| dc.description.abstract | The rapid increase in digital images necessitates efficient retrieval and storage methods, where traditional hashing techniques struggle with scalability and accuracy. This project introduces Deep Reinforcement Learning (DRL) for Image Hashing (DRLIH), leveraging DRL's sequential decision-making to enhance hashing function learning and retrieval performance. Traditional hashing methods like Locality-Sensitive Hashing (LSH) and spectral hashing treat hashing functions independently, ignoring correlations between them and leading to suboptimal accuracy. Deep learning methods using convolutional neural networks (CNNs) for hash code generation also lack mechanisms to refine hashing functions sequentially based on previous errors. DRLIH models the hashing process as a Markov Decision Process (MDP), incorporating state representation from recurrent neural networks (RNNs) and CNNs, an action space of possible hashing functions, and a reward function providing feedback on hash code accuracy. The DRL agent, which features an RNN-based architecture with Long Short-Term Memory (LSTM) units, is trained using the policy gradient method to maximize cumulative rewards through iterative interactions. By iteratively correcting errors and leveraging historical context, DRLIH offers a robust, scalable solution for large-scale image retrieval, setting a new standard in the field. Future work will explore advanced RL techniques to further enhance performance. | |
| dc.identifier.uri | http://192.168.75.5:4000/handle/123456789/19193 | |
| dc.language.iso | en | |
| dc.publisher | NHCE | |
| dc.title | Deep Reinforcement Learning for Image Hashing | |
| dc.type | Learning Object |