Go Deep or Broad? Exploit Hybrid Network Architecture for Weakly Supervised Object Classification and Localization.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Weakly supervised learning for object classification and localization typically uses image-level labels, lacking bounding box annotations.
• Conventional deep convolutional neural network (CNN)-based methods often focus on the most discriminative object parts, potentially harming classification and neglecting shallow features.

Purpose of the Study:

• To propose a novel hybrid network, DB-HybridNet, that enhances both classification and localization performance in a single frame.
• To integrate discriminative and complementary features from different network layers, including high-level semantic and low-level edge features.

Main Methods:

• A novel hybrid network, DB-HybridNet, combining deep CNNs with a broad learning network.
• Integration of multilevel features (semantic and edge features) via a global feature augmentation module.
• An iterative training algorithm based on gradient descent for an end-to-end framework.

Main Results:

• Achieved state-of-the-art classification and localization performance on CUB-200 and ILSVRC 2016 datasets.
• Demonstrated the effectiveness of combining deep and broad learning with multilevel feature integration.
• Validated the end-to-end training framework for improved weakly supervised object recognition.

Conclusions:

• DB-HybridNet offers a significant advancement in weakly supervised object classification and localization.
• The hybrid approach effectively leverages features from various network depths for superior performance.
• This method addresses limitations of conventional CNNs by incorporating broader learning and multilevel feature fusion.