Linear distance coding for image classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Traditional feature coding-pooling frameworks excel in image classification by generating discriminative and robust representations.
• However, information loss from feature quantization and pooling's spatial dependence limit classification performance.

Purpose of the Study:

• To propose a linear distance coding (LDC) method that captures lost discriminative information and reduces pooling's spatial dependence.
• To enhance image classification accuracy by addressing limitations in existing coding methods.

Main Methods:

• Developed a linear distance coding (LDC) method transforming local image features into discriminative distance vectors.
• Employed robust image-to-class distance and encoded distance vectors into sparse codes to capture salient image features.

Main Results:

• LDC is theoretically and experimentally shown to be complementary to traditional coding methods.
• Combining LDC with traditional methods achieves higher classification accuracy.
• Demonstrated effectiveness on diverse datasets (Flower 102, PFID 61, Scene 15, Indoor 67, Caltech 101, Caltech 256).

Conclusions:

• The proposed LDC method generally outperforms traditional coding methods in image classification.
• LDC achieves state-of-the-art or comparable performance across various object and scene recognition tasks.