Bioinspired Scene Classification by Deep Active Learning With Remote Sensing Applications.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Cognitive Science

Background:

• Scene classification is crucial for intelligent systems but current deep models lack human visual perception encoding.
• Existing methods fail to explicitly represent gaze movements and cognitive processes in scene analysis.

Purpose of the Study:

• To propose a biologically inspired deep model for scene classification that incorporates human gaze behaviors.
• To enhance scene classification accuracy by mimicking human visual perception through a unified deep active learning (UDAL) framework.

Main Methods:

• Decomposed scenes into object patches using an objectness measure for varied object sizes.
• Developed a local-global feature fusion scheme to integrate multimodal features with automatic weight calculation.
• Implemented UDAL to hierarchically represent human gaze behavior, combining salient region detection and gaze shifting path (GSP) learning with partial semantic tags and sparsity penalty.

Main Results:

• The proposed UDAL framework successfully discovered and represented human gaze behaviors for scene classification.
• The model achieved competitive performance on six benchmark scene datasets, including remote sensing images.
• The approach effectively integrated multimodal features and avoided redundant low-level regional features.

Conclusions:

• The biologically inspired deep model offers a novel approach to scene classification by encoding human visual perception.
• The UDAL framework and learned deep GSP features provide a robust method for analyzing complex sceneries.
• This research advances computer vision by bridging the gap between deep learning models and human cognitive processes in scene understanding.