YOLO-SASE: An Improved YOLO Algorithm for the Small Targets Detection in Complex Backgrounds.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Signal Processing

Background:

• Detecting small infrared targets in complex backgrounds presents significant challenges.
• Existing algorithms often struggle with low contrast and cluttered environments.

Purpose of the Study:

• To enhance the detection capabilities for infrared small targets.
• To improve the accuracy, recall, and stability of target detection algorithms.

Main Methods:

• An improved detection algorithm, YOLO-SASE, was developed based on the YOLO detection framework and SRGAN network.
• The algorithm incorporates super-resolution reconstructed images, SASE module, SPP module, and a multi-level receptive field structure.
• Feature weight exploration was used to adjust detection output layers and improve feature utilization efficiency.

Main Results:

• The YOLO-SASE algorithm demonstrated a 2% improvement in accuracy and a 3% improvement in recall rate compared to the original model.
• Significant improvements in the stability of results during the training process were observed.
• The algorithm effectively handles complex backgrounds for infrared small target detection.

Conclusions:

• The proposed YOLO-SASE algorithm offers a robust solution for infrared small target detection.
• The integration of super-resolution and specialized modules enhances detection performance and stability.
• This work contributes to advancements in computer vision for challenging detection tasks.