Scalable Discrete Matrix Factorization and Semantic Autoencoder for Cross-Media Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Hashing methods are crucial for multimedia tasks but suffer from quantization error and scalability issues.
• Existing supervised cross-modal methods require large similarity matrices, leading to high computational costs.

Purpose of the Study:

• To propose a novel, scalable algorithm for generating accurate and semantically meaningful binary hash codes.
• To overcome the limitations of large matrix computations and quantization errors in existing hashing techniques.

Main Methods:

• Introduced a two-stage method: scalable discrete matrix factorization (SDMSA) and a semantic autoencoder.
• Stage 1: Utilized matrix factorization with label matrices to learn latent semantic information and generate binary codes, avoiding large similarity matrices.
• Stage 2: Employed an autoencoder with an encoder-decoder paradigm to preserve semantic and feature information in the learned projections.

Main Results:

• Developed two algorithms, SDMSA-lin and SDMSA-ker, under the SDMSA framework.
• Demonstrated that SDMSA generates more semantically meaningful binary hash codes.
• Achieved promising performance across multiple experimental databases.

Conclusions:

• SDMSA offers an effective and scalable solution for supervised cross-modal hashing.
• The proposed method successfully addresses quantization error and computational complexity.
• SDMSA provides a robust framework for learning high-quality binary hash codes for multimedia applications.