CM-supplement network model for reducing the memory consumption during multilabel image annotation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• The exponential growth of digital image resources necessitates efficient retrieval and organization methods.
• Image auto-annotation is a fundamental yet challenging task in image retrieval.
• Existing multi-label annotation methods suffer from low accuracy and high memory consumption.

Purpose of the Study:

• To address the limitations of current image auto-annotation techniques.
• To propose a novel CM-supplement network model for improved multi-label image annotation.
• To enhance the accuracy and reduce the memory footprint of image annotation systems.

Main Methods:

• Developed a CM-supplement network model integrating cavity convolutions, Inception modules, and a supplement network.
• Utilized cavity convolutions to expand the receptive field without increasing parameters.
• Employed Inception modules for multi-scale feature extraction with reduced memory usage.
• Incorporated a supplement network to capture negative image features.

Main Results:

• The CM-supplement network achieved an overall annotation accuracy of 94.5% after 100 training iterations on the PASCAL VOC 2012 dataset.
• This represents a 10.0 and 1.1 percentage point increase over traditional Convolutional Neural Network (CNN) and Double-Channel CNN (DCCNN) methods, respectively.
• The model reached a stabilized accuracy of 96.4%.
• The CM-supplement network has over 1.5 times fewer parameters than DCCNN, indicating significantly lower memory resource consumption.

Conclusions:

• The proposed CM-supplement network offers a superior solution for multi-label image annotation.
• It achieves comparable or better annotation performance than existing methods while consuming less memory.
• This advancement is crucial for efficient image retrieval and management in the era of big data.