Real-time detector design for small targets based on bi-channel feature fusion mechanism

YOLOv4-tiny is a simplified version of YOLOv4 detector, which is extremely fast and with few parameters. However, the detection performance of YOLOv4-tiny is poor while the recognition of small targets and occluded objects is weak. It is mainly attributed to the lack of feature extraction ability and learning ability of the backbone. Furthermore, the feature pyramid network (FPN) cannot adequately fuse adjacent feature maps in the process of multi-scale feature fusion. In this paper, a real-time detector with a bi-channel feature fusion mechanism is proposed based on YOLOv4-tiny, called BFF-YOLO, which effectively improves the detection of small targets and occluded objects. BFF-YOLO is composed of two main components: feature extraction and feature fusion. In the process of feature extraction, inspired by the idea of cross-stage partial connections (CSP), an enhanced CSP block (ECSPBlock) is proposed for enhancing the feature extraction of the backbone and the learning capability of the network. Moreover, the Maxpool layer in YOLOv4-tiny, which is used for downsampling and tends to lose fine-grained information, is replaced with a convolutional layer. In the process of feature fusion, a bi-channel feature fusion pyramid network (BFPN) is proposed to adequately fuse adjacent feature maps of different scales so that each detection head has both shallow and deep features. Finally, with a small increase in parameters, BFF-YOLO has achieved 36.5% AP and 85.1% mAP on the COCO and VOC datasets, respectively.