AutoRet: A Self-Supervised Spatial Recurrent Network for Content-Based Image Retrieval.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Multimedia data is rapidly expanding, increasing the need for efficient image retrieval.
• Current image retrieval systems often rely on labeled data, which can be costly and time-consuming to obtain.
• Self-supervised and unsupervised learning offer alternatives but can be sensitive to class numbers and struggle with mixed labeled data.

Purpose of the Study:

• To introduce AutoRet, a novel deep convolutional neural network (DCNN) based self-supervised image retrieval system.
• To develop a system capable of effective image retrieval using pairwise constraints, enabling both self-supervised and partially supervised training.
• To address the limitations of existing methods regarding data labeling costs and the inability to integrate available labeled data.

Main Methods:

• Utilized a deep convolutional neural network (DCNN) to extract image embeddings from multiple patches.
• Implemented a strategy to fuse these embeddings, enhancing the quality of information for retrieval.
• Trained the system using pairwise constraints, allowing flexibility for self-supervised and partially supervised learning.

Main Results:

• The proposed AutoRet method demonstrated superior performance in a self-supervised setting across three benchmark datasets.
• The system showed highly convincing performance when a small portion of labeled data was incorporated.
• Benchmarking confirmed the effectiveness of the DCNN-based approach and embedding fusion technique.

Conclusions:

• AutoRet provides an effective self-supervised image retrieval solution, overcoming the reliance on extensive labeled data.
• The system's ability to leverage partially labeled datasets makes it a versatile tool for real-world applications.
• The proposed method represents a significant advancement in self-supervised learning for image retrieval systems.