PicSOM-self-organizing image retrieval with MPEG-7 content descriptors.

Area of Science:

• Computer Science
• Artificial Intelligence
• Information Retrieval

Background:

• Content-based image retrieval (CBIR) has historically lacked standardized methods for describing visual content.
• The MPEG-7 international standard offers a framework and specific descriptors for content description, addressing this limitation.

Purpose of the Study:

• To apply MPEG-7 visual content descriptors within the PicSOM system, a neural, self-organizing technique for CBIR.
• To compare the performance of the PicSOM system using MPEG-7 descriptors against a reference system based on vector quantization (VQ).

Main Methods:

• Development of the PicSOM system, utilizing tree-structured self-organizing maps (SOMs) based on pictorial examples and relevance feedback (RF).
• Integration and application of MPEG-7 visual content descriptors within the PicSOM architecture.
• Comparative analysis against a vector quantization (VQ) based image indexing technique.

Main Results:

• MPEG-7 content descriptors are compatible with the PicSOM system, despite Euclidean distance not being optimal for all descriptors.
• The PicSOM system demonstrates a slower initial retrieval phase compared to the VQ reference system.
• PicSOM's retrieval precision significantly surpasses the reference system once its relevance feedback (RF) mechanism is engaged.

Conclusions:

• The PicSOM system effectively leverages MPEG-7 descriptors for enhanced content-based image retrieval.
• The relevance feedback (RF) mechanism is crucial for PicSOM's superior retrieval precision.
• PicSOM presents a viable and powerful alternative for advanced image retrieval applications.