Supervised Deep Feature Embedding With Handcrafted Feature.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Deep convolutional neural networks (CNNs) excel in computer vision tasks like image retrieval.
• Supervised deep metric learning and handcrafted features can improve discriminative feature embeddings.
• Existing methods may not fully leverage the synergy between deep and handcrafted features.

Purpose of the Study:

• To propose a novel supervised deep feature embedding method integrating handcrafted features into CNNs.
• To develop a general fusion unit (Fusion-Net) for combining diverse feature types.
• To enhance image retrieval and re-identification performance through improved feature representation.

Main Methods:

• A novel Fusion-Net architecture designed to fuse handcrafted feature information into CNNs.
• Implementation of a supervised deep metric learning approach using a custom network loss function.
• Extensive experimentation on benchmark datasets for image retrieval, person re-identification, and vehicle re-identification.

Main Results:

• The proposed Fusion-Net method significantly outperforms state-of-the-art techniques in image retrieval on the Stanford online products and in-shop clothes datasets.
• Demonstrated superior performance in person re-identification and vehicle re-identification tasks.
• Validated the effectiveness and efficiency of the integrated deep and handcrafted feature embedding approach.

Conclusions:

• The proposed Fusion-Net model effectively integrates handcrafted features with deep learning for enhanced image representation.
• The method offers significant improvements in image retrieval and re-identification tasks.
• This approach provides a robust and efficient solution for advanced computer vision applications.