Deep Learning Based Feature Selection Algorithm for Small Targets Based on mRMR.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Distinguishing small targets in complex backgrounds is challenging.
• Deep learning advances feature extraction but can introduce redundant relationships, reducing accuracy.
• Existing methods struggle with effective small target identification.

Purpose of the Study:

• To improve the accuracy and efficiency of small target feature recognition.
• To address the issue of redundant features in deep learning models.
• To enable reliable identification of small targets even in cluttered environments.

Main Methods:

• Utilized the YOLOv5 neural network for preliminary multi-dimensional feature extraction.
• Applied the Minimum Redundancy Maximum Relevance (MRMR) algorithm to de-redundancy candidate features.
• Implemented image pre-processing techniques to enhance feature recognition.

Main Results:

• Successfully reduced the dimensionality of the feature set by removing highly correlated features.
• Demonstrated effective identification of small target features through feature de-redundancy.
• Significantly improved recognition accuracy via pre-processing and optimized feature selection.

Conclusions:

• The MRMR algorithm effectively reduces feature dimensions and enhances small target recognition.
• Combining YOLOv5 with MRMR offers a robust solution for complex background scenarios.
• Image pre-processing further boosts the performance of small target detection systems.