FFENet: frequency-spatial feature enhancement network for clothing classification.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Clothing classification is crucial in computer vision but challenged by complex real-world scenes.
• Existing methods struggle with feature extraction due to interference from complex environments, impacting contour and texture analysis.
• Poor classification results arise from relying solely on spatial information in cluttered clothing datasets.

Purpose of the Study:

• To propose a clothing classification network that effectively integrates frequency and spatial domain information.
• To enhance the extraction of clothing features by leveraging both frequency and spatial data without channel compression.
• To improve the accuracy of clothing classification, particularly in complex, real-world scenarios.

Main Methods:

• Developed a novel network architecture based on frequency-spatial domain conversion for clothing classification.
• Integrated frequency domain features with spatial domain features, maintaining uncompressed feature map channels.
• Introduced a frequency domain feature enhancement module for preliminary clothing feature extraction.
• Created and utilized a new dataset, Clothing-8, specifically for clothing analysis in complex scenes.

Main Results:

• Achieved a top-1 model accuracy of 93.4% on the challenging Clothing-8 dataset.
• Attained a high top-1 accuracy of 94.62% on the Fashion-MNIST dataset.
• Demonstrated superior performance on the DeepFashion dataset, achieving top-3 and top-5 metric benchmarks.

Conclusions:

• The proposed frequency-spatial domain conversion network effectively enhances clothing feature extraction.
• Integrating frequency and spatial information significantly improves clothing classification accuracy in complex scenes.
• The network shows strong generalization capabilities across diverse clothing datasets, including real-world scenarios.