Improved Small Object Detection Algorithm CRL-YOLOv5.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Small object detection in images is challenging due to limited pixel data and feature extraction difficulties.
• Existing methods often result in missed or false detections of small objects.

Purpose of the Study:

• To enhance the accuracy and precision of small object detection algorithms.
• To introduce an improved algorithm, CRL-YOLOv5, addressing the limitations of current small object detection methods.

Main Methods:

• Integrated the Convolutional Block Attention Module (CBAM) into the C3 module of the backbone network for improved localization.
• Incorporated the Receptive Field Block (RFB) module to broaden the receptive field and leverage contextual information.
• Restructured the network architecture with an additional detection layer for small objects, enabling deeper feature extraction from shallow layers.

Main Results:

• CRL-YOLOv5 achieved a mean Average Precision (mAP50) of 39.2% on the VisDrone2019 small object dataset.
• Demonstrated a 5.4% improvement in mAP50 compared to the original YOLOv5 algorithm.
• Effectively enhanced the detection precision for small objects in complex image scenarios.

Conclusions:

• The proposed CRL-YOLOv5 algorithm offers a significant improvement for small object detection.
• The integration of CBAM and RFB modules, along with architectural modifications, effectively addresses the challenges of detecting small objects.
• The enhanced algorithm shows strong potential for real-world applications requiring precise small object identification.