Bag of Visual Words Model with Deep Spatial Features for Geographical Scene Classification.

Area of Science:

• Computer Science
• Geographic Information Science
• Artificial Intelligence

Background:

• Geotagging images are increasingly common, driving research in geographical scene classification.
• Effective spatial feature selection is crucial for improving classification performance.
• Traditional Bag of Visual Words (BoVW) methods rely on well-matched feature extractors for geographical scene classification.

Purpose of the Study:

• To optimize feature extraction for geographical scene classification using geotagged images.
• To improve the selection of visual vocabularies for enhanced classification accuracy.
• To compare the performance of a CNN-optimized approach against the standard BoVW method.

Main Methods:

• Utilizing convolutional neural networks (CNNs) to develop an optimized feature extractor.
• Training the CNN to learn suitable visual vocabularies directly from geotagged image data.
• Evaluating the proposed method on three diverse datasets containing various scene categories.

Main Results:

• The CNN-optimized feature extractor demonstrated superior performance compared to the standard BoVW.
• The approach successfully learned more relevant visual vocabularies for geographical scenes.
• Consistent performance improvements were observed across all tested datasets.

Conclusions:

• Convolutional neural networks offer a powerful method for optimizing feature extractors in geographical scene classification.
• The proposed CNN-based approach provides a more effective solution than traditional BoVW for analyzing geotagged images.
• This research contributes to advancing the accuracy and efficiency of geographical scene classification systems.