Recognition From Web Data: A Progressive Filtering Approach.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Training deep learning models like convolutional neural networks (CNNs) often requires large datasets.
• Web data offers abundant resources but frequently contains inaccurate labels (noise).
• Noisy labels can significantly degrade the performance of trained CNN models.

Purpose of the Study:

• To develop a robust method for training CNNs using noisy web image data.
• To improve image classification accuracy by effectively filtering and utilizing web-sourced images.
• To address the challenge of inaccurate or incomplete labeling in large-scale web image datasets.

Main Methods:

• An iterative approach is proposed, alternating between filtering noisy web labels and fine-tuning the CNN model.
• The method leverages the model's improving capability to refine web image label accuracy.
• A multi-label assignment strategy is employed to mitigate the impact of single, potentially incorrect, labels for complex web images.

Main Results:

• The proposed iterative method significantly enhances CNN discriminative ability and web image selection accuracy.
• Utilizing 0.5 million crawled web images demonstrated notable improvements over baseline methods lacking web data.
• The approach achieved competitive recognition accuracy compared to state-of-the-art methods.

Conclusions:

• The iterative filtering and multi-labeling strategy effectively cleans noisy web data for CNN training.
• This method enables the use of large, unlabeled web image datasets to build more accurate CNN models.
• The proposed approach offers a viable solution for data-scarce scenarios in image classification.