An End-to-End Trainable Multi-Column CNN for Scene Recognition in Extremely Changing Environment.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Scene recognition is crucial for vision-based robot navigation.
• Deep learning methods often rely on pre-extracted features for scene recognition.
• Existing approaches have limitations in capturing multi-level and multi-layer image perceptions.

Purpose of the Study:

• To present a novel end-to-end trainable Multi-column Convolutional Neural Network (MCNN) for scene recognition.
• To interpret scene recognition as a region-based image retrieval problem.
• To enhance intra-class compactness and inter-class separability for improved recognition.

Main Methods:

• Developed a novel Multi-column Convolutional Neural Network (MCNN) architecture.
• Employed filters with varying receptive fields for multi-level and multi-layer image perception.
• Utilized VGG16 (first seven layers) for feature extraction, Inception-A for deeper representation, and Large-Margin Softmax Loss (L-Softmax) for classification.

Main Results:

• The proposed MCNN architecture demonstrated robustness and accuracy in scene recognition tasks.
• Experimental results on three popular datasets showed superior performance compared to state-of-the-art methods.
• The approach achieved enhanced intra-class compactness and inter-class separability.

Conclusions:

• The novel MCNN approach offers a significant advancement in vision-based robot navigation scene recognition.
• The method's effectiveness is validated through extensive experiments on benchmark datasets.
• This represents a new application of this specific architecture for scene recognition in the literature.