基于深度学习的水面目标检测模型压缩方法

Compression method of water surface target detection model based on deep learning

针对目标检测模型过大且计算复杂而导致其无法应用于无图形处理器嵌入式终端的问题,通过改进YOLO算法,提出一种基于深度学习的水面目标检测模型压缩方法.采用带有深度可分离卷积和轻量级注意力模型的改进网络替代特征提取网络DarkNet,通过多尺度特征融合进行模型压缩,引入k-means++算法与Mish激活函数,保证模型压缩后的准确度.试验结果表明,YOLOv3-MobileNetV3网络模型较YOLOv3网络模型的参数量减少61.35%,模型大小减少144 MB,模型平均精度均值较YOLOv3-MobileNetV1网络模型提升5.55%,满足嵌入式设备水面目标检测实时性和准确性的要求.

Aiming at the problem that the target detection model is too large and the calculation is too complex,which makes it unable to be applied to embedded terminals without GPU(graphic processing unit,GPU),a deep learning-based model compression method is proposed by improving the YOLO(you only look once)algorithm.The feature extraction network DarkNet is replaced by an improved network with deep separable convolution and lightweight attention model(squeeze-and-excitation,SE).To ensure the accuracy of the compressed model,multi-scale feature fusion is adopted in model compression,and the k-means++algorithm is introduced,together with the Mish activation Function.The experimental results show that,compared with the YOLOv3 network model,the YOLOv3-MobileNetV3 network model reduces the number of parameters by 61.35%,the model size is reduced by 144 MB,and the mAP(mean average precision,MAP)of the YOLOv3-MobileNetV3 model is 5.55%higher than YOVOv3-mobilenetv1 network model,which meets the accuracy and real-time requirements of water surface target detection on embedded devices.