压缩SSD低空小目标检测算法研究

Research on Detection Algorithm of Small Target at Low Altitude Based on Compressed SSD

采用SSD模型对低空飞行目标进行检测,为提高模型检测速度,进行SSD模型压缩。首先,在SSD模型的激活层后添加批量归一化(BN)层,为各通道引入比例因子;然后,联合训练比例因子,使其数值与通道的重要性相关联,再采用正则化方法对比例因子进行稀疏化处理;最后,通过衡量比例因子的大小,剪除重要性低的通道,以70%的剪枝率对检测模型进行通道剪枝。实验结果显示,模型压缩率达到18.4%,检测精度值提高了0.3%,检测速度由原来的每秒28帧提升至每秒61帧。

The SSD model is used to detect the low-altitude flying target.Therefore,In order to improve the model detection spead,the SSD model should be compressed.Firstly,BN layer was added after the activation layer of SSD model to introduce scaling factor for each channel.Then,the scale factors are jointly trained to correlate their values with the importance of channels.Next,the regularization method is used to sparse the scale factors.Finally,by measuring the scale factor,the channels with low importance are cut off.At a 70%pruning rate,channel pruning was carried out on the detection model.Experimental results showed that the compression rate of the model reached 18.4%,the detection accuracy increased by 0.2%,and the detection speed increased from the original 28 frames per second to 61 frames per second.