基于权重和BN层剪枝的晶界检测模型压缩算法探析  

Analysis of Grain Boundary Detection Model Compression Algorithm Based on Weight Pruning and BN Layer Pruning

作　　者：李静[1] 卯福启[2,3] LI Jing;MAO Fuqi(North China University of Technology,Beijing 100144,China;China University of Geosciences (Beijing),Beijing 100083,China)

机构地区：[1]北方工业大学信息学院,北京100144 [2]北方工业大学,北京100144 [3]中国地质大学(北京),北京100083

出　　处：《北京工业职业技术学院学报》2022年第3期27-34,共8页Journal of Beijing Polytechnic College

摘　　要：性能优越的晶界缺陷检测模型往往存在网络参数量过多、结构冗余及推理时间慢的问题,导致模型部署过程成本高、时延长。针对上述问题,通过稀疏化权重和通道剪枝对晶界检测算法EfficientDet网络进行模型压缩。由数据验证可知,剪枝后的晶界模型检测时间约降为1/2,参数量降低70%,网络模型计算量FLOPs降低60%左右。The superior performance of grain boundary defect detection models often suffers from excessive number of network parameters,structural redundancy and slow inference time,resulting in high cost and prolonged model deployment process.To solve the above problems,the model compression of the grain boundary detection algorithm EfficientDet network is carried out by sparse weights and channel pruning.It can be seen from the data verification that the detection time of the grain boundary model after pruning is reduced by about 1/2,the amount of parameters is reduced by 70%,and the amount of FLOPs calculated by the network model is reduced by about 60%.

关 键 词：晶界缺陷检测模型 模型压缩 EfficientDet网络 权重剪枝 BN层通道剪枝 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]