Detect small object based on FCOS and adaptive feature fusion

Abstract

Object identification has always been a difficult issue in the field of computer vision, where the objective is to identify the location and kind of particular objects in images. Convolutional neural networks have achieved significant advancements in accuracy and speed, and this is mostly due to their quick evolution. However, in practical applications, the detection of small objects is still an unresolved problem. Due to the sparse pixel information of small target, it is relatively difficult to extract small target features effectively. Although many methods have made significant progress, the detection rate and recall rate both have space for improvement under various conditions. This study suggests a small object detection technique that utilizes feature map weight self allocation and single-stage detection model FCOS to further increase the recall and accuracy of small object detection., which is used in aerial photography dataset and airport foreign object dataset respectively. The average detection accuracy and recall of the technique are roughly 10% higher than the baseline, as shown by the quantitative and qualitative trial findings, demonstrating the effectiveness of the adaptive feature fusion approach.