MBAB-YOLO: A Modified Lightweight Architecture for Real-Time Small Target Detection

Abstract

Current target detection methods have achieved high accuracy for detecting large and medium-sized targets. However, due to factors such as the small number of pixels and features available for targets in images, the detection performance for small targets is generally unsatisfactory. In addition, the real-time performance of target detection is also critical. In conclusion, a modified lightweight architecture for real-time small target detection, i.e., MBAB-YOLO, is proposed based on You Only Look Once (YOLO) model by combining channel-wise attention block, space-attention block and multi-branch-ConvNet (Convolutional Neural network) structure. Specifically, our method is more suitable for the rich scale information of small targets through proposed adaptive multi-receptive-field focusing, and then combines proposed blended attention block (BAB) to re-calibrate small target information to make it more prominent and improve the discriminability of small target features. Finally, extensive experiments have been conducted on the open source data set for the proposed real-time small target detection method, i.e., MBAB-YOLO. The results of ablation experiment and contrast experiment show that our method has excellent performance, not only with high detection accuracy, but also with fast detection speed. Compared with the various benchmark methods, it achieves a good trade-off between the two aspects mentioned above. In addition, this paper gives a comprehensive and detailed review of the current work about small target detection from different several perspectives, which can be used as a reference for future researchers.

Keywords

• Deep learning
• target detection
• channel-wise attention
• space-attention
• YOLO