A novel feature representation: Aggregating convolution kernels for image retrieval.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Feature maps from Convolutional Neural Networks (CNNs) are effective but computationally intensive for large datasets.
• High-dimensional float-format feature maps consume significant memory and processing power.
• Existing methods face challenges with ultra-large datasets due to computational complexity.

Purpose of the Study:

• To propose a novel image representation method that reduces computational complexity and memory footprint.
• To develop an efficient image retrieval system for ultra-large datasets.
• To enhance the discriminative power of image representations.

Main Methods:

• A new image representation technique called Aggregating Convolution Kernels (ACK) is introduced.
• ACK activates specific convolution kernels and extracts their top-n index numbers as discrete integer representations.
• A novel distance measurement based on ordered sets is defined for calculating position-sensitive similarities.

Main Results:

• The proposed ACK method achieves competitive performance in image retrieval tasks.
• ACK significantly lowers computational cost compared to traditional feature map representations.
• Experiments on benchmark datasets (Oxford Buildings, Paris, Holidays) demonstrate ACK's effectiveness.

Conclusions:

• ACK offers a more efficient and memory-saving approach to image representation for large-scale applications.
• The method outperforms conventional feature map-based representations in terms of efficiency and performance.
• ACK provides a promising direction for optimizing deep learning models in computer vision.