Abstract: To tackle the problems of low accuracy, high error rate, and high missed rate for small targets in target
detection tasks, an improved YOLOv4 small target detection algorithm named scSE-IYOLOv4 was proposed. The
experiment employed the VEDAI small target dataset and the K-means++ algorithm to optimize the anchor frame of
the target sample, thereby improving the accuracy of the algorithm. Based on the YOLOv4 algorithm, studies and
analyses were respectively conducted concerning the effect of the scSE attention module embedded in the model’s
different positions and that of the addition of the SPP module to the model’s neck on the algorithm’s detection
performance. The experiments proved that embedding the scSE attention module after the “Add” and “concat” layers
of the backbone network of the YOLOv4 algorithm model, and adding the SPP module to the model’s neck can
enhance the algorithm’s detection accuracy for small targets. The results showed that the mAP@0.5 both increased by
2.4% on the VEDAI test set. Finally, the scSE-IYOLOv4 algorithm was proposed based on the above experimental results of the improved backbone network and neck of the YOLOv4 model. It was proved by the experiment that
scSE-IYOLOv4 can significantly heighten the detection accuracy for small targets. The value of mAP@0.5 on the
VEDAI test set increased by 4.1%, compared with YOLOv4, and rose by 2.2% on the PASCAL VOC dataset.

Key words: small target detection, YOLOv4, scSE attention, spatial syramid pooling, K-means++