SSD-TSEFFM: New SSD Using Trident Feature and Squeeze and Extraction Feature Fusion.

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• The Single Shot Multi-Box Detector (SSD) struggles with small object detection due to insufficient scale contextual information and semantic deficiency in shallow layers.
• Existing methods like Faster R-CNN and DF-SSD have limitations in accurately detecting small objects.

Purpose of the Study:

• To enhance the accuracy of the Single Shot Multi-Box Detector (SSD) for small object detection.
• To introduce a novel detector, SSD-TSEFFM, that addresses the limitations of the original SSD.

Main Methods:

• Proposes the SSD-TSEFFM, integrating a trident feature module (TFM) for scale contextual information using dilated convolution and a squeeze and excitation feature fusion module (SEFFM) for enhanced semantic information.
• The trident feature module (TFM) leverages dilated convolutions to capture multi-scale features, improving robustness to scale variations.
• The squeeze and excitation feature fusion module (SEFFM) enriches feature representations by adaptively recalibrating channel-wise feature responses.

Main Results:

• The proposed SSD-TSEFFM achieved high accuracy in small object detection, outperforming Faster R-CNN (2015), SSD (2016), and DF-SSD (2020) on PASCAL VOC 2007 and 2012 datasets.
• Achieved mean Average Precision (mAP) of 80.4% on PASCAL VOC 2007 and 80.2% on PASCAL VOC 2012.
• Demonstrated an average improvement of approximately 2% in mAP over existing models, indicating superior performance.

Conclusions:

• The SSD-TSEFFM effectively improves small object detection accuracy by integrating scale contextual and semantic information.
• The model shows robustness to scale changes and provides better overall accuracy compared to previous state-of-the-art methods.
• The proposed approach offers a significant advancement in object detection, particularly for challenging small objects.