基于串联空洞卷积神经网络的网球场景语义分割

室外网球场实景环境下的语义分割是开发网球运动服务机器人需要解决的一项关键技术.针对室外网球场环境由于光照、视角等因素的变化导致难以进行精细分割的问题,提出一种基于Deeplabv3架构的串联空洞卷积神经网络SACNet.该模型扩展了空洞卷积的多尺度模块,通过奇偶混合扩张率增大训练中特征图感受野,利用分组卷积减小SACNet模型时空复杂度.为克服小样本训练容易过拟合的问题,在经过预训练残差神经网络ResNet-50的基础上,通过冻结批量归一化(BN)层进行迁移学习.在自制小样本网球场数据集上进行实验结果表明,SACNet比现有Deeplabv3模型在测试精度提高了10.22%,且对于网球场边界和小目标语义分割结果更加准确.

Semantic Segmentation of Tennis Scene Based on Series Atrous Convolution Neural Network

Semantic segmentation of outdoor tennis scene is a key technology for developing mobile robot to service on tennis court.The outdoor tennis court environment contains dynamic illumination,perspective and other interference factors,which make it difficult to carry out semantic segmentation.To solve this problem series atrous convolution neural network(SACNet)is proposed based on Deeplabv3.SACNet extends the multi-scale module of atrous convolution to enlarge the receptive field of feature map in training by means of odd-even mixed dilation rate.SACNet reduces its space-time complexity by means of grouping convolution.To overcome the problem of over-fitting caused by limited training samples,the pre-trained residual neural network ResNet-50 is used for transfer learning by freezing the parameters of batch normalization layer.Experimental results on the self-made tennis court datasets show that SACNet improves the test mIOU compared with Deeplabv3 by 10.22%and is more accurate for tennis court boundaries and small targets.