Deep Features for Training Support Vector Machines.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Traditional computer vision relied on handcrafted features.
• Convolutional Neural Networks (CNNs) now learn features automatically.
• Extracting and utilizing features from inner CNN layers is challenging due to high dimensionality.

Purpose of the Study:

• To develop a generic computer vision system leveraging features from trained CNNs.
• To combine multiple learned features for diverse image classification tasks.
• To improve the performance of standard CNNs through novel feature extraction and combination methods.

Main Methods:

• Extracted features from inner layers of CNNs.
• Utilized Support Vector Machines (SVMs) with combined features.
• Applied dimensionality reduction techniques, including Discrete Cosine Transform (DCT) on each channel.
• Tested ensemble methods with different CNN topologies and global mean thresholding pooling.

Main Results:

• The proposed system, using DCT on inner CNN features, significantly boosted performance on diverse image datasets.
• Dimensionality reduction techniques were effective for SVM integration.
• Ensemble models achieved state-of-the-art results on a benchmark image virus dataset.

Conclusions:

• Learned features from inner CNN layers can be effectively combined for generic computer vision tasks.
• DCT is a powerful technique for enhancing feature representation and performance.
• The developed system offers a robust and high-performing solution for image classification.