Multi-scale Structure Aggregation with Instance-aware Enhancement for Object Detection

Object detection performance of low-quality images is crucial for ensuring safety in surveillance, medical diagnosis, and autonomous driving. In response to the challenge of object feature blurring caused by a large amount of noise in low-quality images, a multi-scale structure aggregation with instance-aware enhancement network (MSA-IE) is proposed for low-quality image object detection, which contains the multi-level topological graph fusion and the instance-aware spatial correlation enhancement modules. The former devises a graph feature pyramid network with a parallel multi-scale structure aggregation paradigm for capturing semantic and spatial information of the objects. The latter designs a gate selection to adaptively learn features of prominent foreground instances that are more beneficial for detection tasks, which effectively suppresses noises of low-quality images. Finally, extensive experiments conducted on two common datasets with eight baselines demonstrate exceptional performance of MSA-IE in the object detection of low-quality images.