An Improved Convolutional Neural Network Algorithm and Its Application in Multilabel Image Labeling.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Automatic image annotation is crucial for efficient text-based image retrieval in the age of abundant image resources.
• Imbalanced training data, particularly with low-frequency semantic labels, leads to poor annotation accuracy.
• Existing methods struggle with accuracy and training efficiency when dealing with imbalanced datasets.

Purpose of the Study:

• To develop an improved automatic image labeling model that addresses the challenge of imbalanced semantic annotations.
• To enhance the accuracy and efficiency of automatic image annotation, particularly for low-frequency labels.
• To introduce a novel Dual-Channel Convolutional Neural Network (DCCNN) for superior image labeling performance.

Main Methods:

• Designed a Dual-Channel Convolutional Neural Network (DCCNN) integrating two CNN channels with distinct structures.
• One channel is optimized for low-frequency samples, increasing their proportion during training.
• The second channel trains on the entire dataset; outputs are fused for the final labeling decision.

Main Results:

• The DCCNN achieved 93.4% labeling accuracy on the Pascal VOC 2012 dataset, surpassing CNN by 8.9% and traditional methods by 15%.
• DCCNN reached high accuracy in significantly fewer training iterations (100) compared to CNN (2,500).
• The model demonstrated stable performance (>93% accuracy) on a large dataset (50,000 images) and improved low-frequency label accuracy by ~10% over CNN.

Conclusions:

• The proposed DCCNN effectively enhances automatic image labeling accuracy, especially for imbalanced datasets.
• DCCNN offers a more efficient and reliable solution for image annotation compared to existing CNN and traditional methods.
• The model's ability to improve low-frequency label accuracy validates its robustness and practical applicability.