面向小目标检测的改进RetinaNet模型及其应用

Small Object Detection Oriented Improved-RetinaNet Model and Its Application

基于深度学习的目标检测算法广泛应用于工业检测,RetinaNet算法因兼具速度与精度两方面的优势而备受关注,但对于小于32×32像素的小目标,该算法的检测精度不能满足工业检测的要求。为此,文中以增强小目标的训练为基本思路,针对RetinaNet算法进行了如下改进:在采样阶段,将低层特征图P2添加到FPN中,以确保小目标能被充分采样,同时引入自适应训练样本选择策略,以保证增加特征层之后仍能保持足够快的检测速度;在训练后期采用了损失权重调整策略,用于提高小目标中困难样本的拟合度。针对公共数据集MS COCO 2017及实际应用中的LED点胶工业数据集,改进后的方法使小于32×32目标的检测精度分别提高了4.1%和10.7%,这表明改进后的方法能显著提升小目标检测的水平。

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.