Due to the inherent diversity and complexity in detection environments,the use of algorithms in semantic analysis and sample mining for small object detection creates numerous challenges.For higher detection efficiency,most object detectors with a strong detection effect are anchor-based,that is,anchor boxes with a fixed scale and aspect ratio are used.The design fails to meet the needs associated with detecting objects with different scales and aspect ratios,especially in terms of detecting small-scale objects with relatively weak texture information.In this study,we used a fully convolutional network to detect objects via per-pixel prediction and analogue to semantic segmentation.To detect the small objects remaining,we proposed a new method called the adaptive selection pyramid feature (ASPF) layer,which allows feature layers from the most suitable feature pyramid network (FPN) in each instance via defined settings and automatic feature selection in two ways.Specifically,each object was passed through the convolution network and inputted one FPN layer according to the scale setting.The former layer was changed using the automatic control to assign the optimal feature map.Using this strategy enables easy separation of the overlapping targets,which usually include many small objects that would otherwise be left out.The findings demonstrated the effectiveness of the method in a standard dataset.Compared with other detectors,the proposed method exhibited greater robustness for small object detection.
展开▼
