基于改进Mask R−CNN算法对堆叠汽车零件的实例分割方法

Instance segmentation method based on improved Mask R−CNN for the stacked automobile parts

针对堆叠汽车零件识别检测与分割速度慢、精度低及鲁棒性差等问题,提出一种基于改进Mask R−CNN算法对堆叠汽车零件快速检测与实例分割的方法.首先,对Mask R−CNN中的特征提取网络进行优化,将ResNet+特征金字塔网络(Feature Pyramid Networks,FPN)替换成MobileNets+FPN作为骨干网络,有效减少网络参数并压缩模型体积,提高模型检测的速度;然后,通过在Mask R−CNN的ROI Align结构后加入空间变换网络(Spatial Transformer Networks,STN)模块,保证模型的检测精度.试验结果表明,改进后压缩了模型的尺寸,识别检测速度提升了1倍;模型的平均精度均值(Mean Average Precision,mAP)较改进前也有所提升.对未经训练的新样本进行检测,结果表明该模型速度上优于Mask R−CNN,且更轻量和精准,能够快速准确地实现对堆叠汽车零件检测与分割,验证了改进模型的实际可行性.

Aiming at the problems of slow speed,low accuracy and poor robustness in recognition,detection and segmentation of stacked automobile parts,a fast detection and instance segmentation method based on improved Mask R−CNN algorithm was proposed.Firstly,the feature extraction network of Mask R-CNN was optimized,and ResNet+Feature Pyramid Networks(FPN)was replaced by MobileNets+FPN as the backbone network,which effectively reduced network parameters,compressed model volume and improved model detection speed.Then,Spatial Transformer Networks(STN)module was added after the ROI Align structure of Mask R-CNN to ensure the detection accuracy of the model.The experimental results show that the size of the model is compressed and the detection speed is doubled.The mean Average Precision(mAP)of the model is also improved.The detection of untrained new samples shows that the model is better than Mask R−CNN in speed,lighter and more accurate,and can quickly and accurately detect and segment stacked automobile parts,which verifies the practical feasibility of the improved model.