Aiming at the problems that the existing automatic instrument reading algorithm occupies a large space,the reasoning speed is slow,and it cannot effectively segment the dense and small objects in the image,an improved DeepLabV3+ pointer instrument segmentation algorithm is proposed.Firstly,MobileNetV2 is used to build the network backbone to reduce the amount of parameters and inference weight,and improve the detection speed.Secondly,the CSP-ASPP structure is designed through the block merge strategy to reduce the amount of parameters while ensuring the network performance.Then,the improved SKFF module is used to fuse multi-scale features in a non-linear manner through the self-attention mechanism,and the two-scale feature fusion in the original network decoder is changed to four-scale feature fusion.Finally,the Dice Loss jointly weighted by cross-entropy loss is used as the total loss function of the network to solve the problem of uneven distribution of pixels in each category in instrument segmentation.Finally,it is proved by experiments that the improved DeepLabV3+ average intersection ratio(mIoU)and mean pixel accuracy(mPA)reached 89.3%and 94.8%,respectively,increased by 0.7%and 0.6%compared with the original network,but the amount of parameters and inference weight is only about 7%of the original network,while the inference speed on GPU and CPU reaches 91 and 16 frames/s,respectively.Meet the requirements of real-time detection,which solves the problem of difficult deployment of embedded devices and improves the efficiency of automatic instrument reading.
semantic segmentation of pointer metersDeepLabV3+lightweightblock-wise aggregationmulti-scale feature fusionDice Loss
NETL
NSTL
万方数据
