SSD-EMB: An Improved SSD Using Enhanced Feature Map Block for Object Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• Deep learning has advanced object detection, but detecting small objects remains a significant challenge.
• Existing methods often struggle with accuracy or speed when identifying small objects in complex scenes.

Purpose of the Study:

• To develop an improved single-shot multibox detector (SSD) specifically designed for enhanced small object detection.
• To introduce the enhanced feature map block (EMB) to improve the accuracy and efficiency of small object detection.

Main Methods:

• Proposed an enhanced feature map block (EMB) integrating an attention stream and a feature map concatenation stream.
• The attention stream utilizes channel averaging and normalization to focus on object regions.
• The feature map concatenation stream adds semantic information without compromising detection speed.

Main Results:

• The SSD-EMB model demonstrated high accuracy in small object detection across multiple datasets.
• Achieved a mean average precision (mAP) of 80.4% on PASCAL VOC 2007 and 79.9% on VOC 2012.
• Maintained a detection speed of 30 frames per second on an RTX 2080Ti GPU, with an mAP of 26.6% on MS COCO.

Conclusions:

• The proposed SSD-EMB effectively addresses the challenges of small object detection in computer vision.
• The enhanced feature map block significantly improves detection accuracy while maintaining a competitive detection speed.
• The model shows strong performance on benchmark datasets, indicating its practical applicability.