LiteRevNet:a lightweight algorithm for industrial image instance segmentation
To address the low accuracy in instance segmentation algorithms and the high number of parameters and computational demands in industrial laser cutting seam visual inspection,we propose a lightweight industrial image instance segmentation algorithm(LiteRevNet).First,coordinate information is incorporated into convolution operations,combined with multi-scale convolution kernels to build efficient space-aware convolutions,enhancing the model's feature extraction capabilities.Then,a lightweight backbone network is designed using reversible column networks and efficient space-aware convolutions,maintaining detection accuracy while reducing the amount of computation and parameters.Next,a pyramid coordinate channel attention mechanism with both spatial and channel awareness is designed to increase the network model's focus on target areas.Finally,a lightweight trilateral prototype mask branch is built,significantly reducing the model's computational load.Our results on a self-built laser cutting seam dataset show our proposed algorithm achieves 96.7%in bounding box mAP50 and 95.1%in mask mAP50 with a detection speed of up to 200 FPS.The number of parameters and computational load are down by 15.4%and 38.7%respectively compared with those of YOLOv8s-Seg.
国家科技期刊平台
NSTL
万方数据
