A multi-scale semantically enriched feature pyramid network with enhanced focal loss for small-object detection

Recent deep-learning methods have achieved higher accuracy in the detection of large objects. However, the performances of these methods are relatively poor for small objects. Studies have revealed several reasons for this weakness. First, small objects have limited spatial dimensions, causing them to have less semantic information. Second, the imbalance between foreground and background features complicates the process of distinguishing small objects from the background. This study presents a multi-scale semantic information enhancement module that captures sensitive details from the higher prediction layers of a feature pyramid network. This module infuses the captured details into its lowest prediction layer, which is responsible for detecting small objects. We approached the class imbalance issue using a novel enhanced focal loss method that embeds a special weighting factor to enhance the contribution of easy examples in the object detection training process. The empirical results show that, compared with baseline methods, our method achieves better performance in detecting small objects.