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

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%.