基于结构化剪枝的遥感飞机检测算法

Remote Sensing Aircraft Detection Algorithm Based on Structural Pruning

针对在遥感飞机目标检测场景中轻量级算法难以兼顾准确性与实时性的问题,提出了一种基于YOLOv4结构化剪枝的模型压缩方法。为了使锚框参数更加符合遥感数据集并发挥网络多尺度检测的优势,采用K-means++算法对数据集进行锚框聚类分析,并设计尺度自适应调整,抑制小目标过多以及目标大小接近造成的锚框冗余。此外,为了减少模型参数量,利用归一化层中的缩放因子γ进行L1稀疏正则化,重新评估滤波器及卷积核权重,对特征信息较少的通道进行迭代剪枝,然后微调剪枝模型恢复精度。实验结果表明,剪枝后模型参数量压缩了93.1%,检测速度比原模型提升2.46倍,能够在保证检测准确性的同时有效提升检测实时性。

Regarding the problem that lightweight algorithm in remote sensing aircraft target detection is difficult to balance accuracy and real-time performance,a model compression method based on YOLOv4 structured pruning is presented.In order to make the anchor frame parameters more suitable for remote sensing datasets and take advantage of network multi-scale detection,K-means++ algorithm is used to cluster the datasets, and scale adaptive adjustment is designed to restrain the redundancy of the anchor frame caused by too many small targets and close target sizes.In addition,in order to reduce the parameters of the model,the scaling factor γ in the normalization layer is used for L1 sparse regularization,the filter and convolution kernel weights are re-evaluated,channels with less feature information are iteratively pruned,and then the pruning model is fine-tuned to recover accuracy.The experimental results show that after pruning,the model parameters are compressed by 93.1%,and the detection speed is 2.46 times faster than that of the original model,which can effectively improve the detection accuracy and real-time performance.