Small Object Detection Oriented Improved-RetinaNet Model and Its Application

Abstract

Object detection algorithms based on deep learning are widely used in industrial detection.The RetinaNet algorithm has attracted much attention because of its advantages in both speed and accuracy.However,for small objects smaller than 32*32 pixels,the detection accuracy of this algorithm cannot meet the requirements of industrial detection.To this end,this article takes the enhancement of small object training as the basic idea,and makes the following improvements to the RetinaNet algorithm:in the sampling phase,the low-level feature map P2 is added to the FPN to ensure that the small object can be fully sampled,and adaptive training sample selection(ATSS) strategy is introduced to ensure that the detection speed is still fast enough after the feature layer is increased;the loss weight adjustment strategy is adopted in the later training stage to improve the fit of difficult samples in small objects.For the public data set MS COCO 2017 and the LED dispensing industrial data set in practical applications,the detection accuracy of this method for objects smaller than 32*32 increases by 4.1% and 10.7%,respectively,indicating that this method can significantly improve the detection ability of small objects.

Keywords

• deep learning|small object detection|retinanet|adaptive sampling