用于草坪场景理解的轻量化图像分割算法

Lightweight Image Segmentation for Lawn Scene Understanding

为提高割草机器人作业过程中视觉感知模块的识别准确率,提出了使用图像分割算法进行草坪场景的理解识别。图像分割算法的计算量非常大,运行时依赖高性能的GPU,而割草机器人的硬件条件较差,因此设计了一种兼顾分割准确率和运行速度的轻量化深度卷积神经网络。网络采用编码-解码的结构,在编码网络部分,采用轻量化的特征提取模型,将深度可分离卷积的思想融入特征提取模型中,代替传统的卷积方式;在解码网络部分,基于RefineNet模块减少参数量,融合编码器的高分辨率特征和低分辨率特征。使用PASAL VOC2012分割数据集进行预训练,构建草坪场景数据集进行微调和测试评估。结果表明:提出的算法结构在保持较高准确率的前提下,网络的参数量有大幅度的减少,运行速率有大幅度提高,在机器人草坪场理解任务上有更好的综合性能。

In order to improve the recognition accuracy of visual perception module of the lawn mower,an image segmentation algorithm is proposed for recognizing and understanding lawn scenes.Image segmentation algorithm requires a large amount of computation,and its operation depends on high-performance GPU,while the hardware condition of g lawn mower is poor.Therefore,a lightweight depth convolutional neural network with both segmentation accuracy and running speed is designed.The network adopts a encoding-decoding structure.In the part of encoding network,a lightweight feature extraction model is adopted.The idea of deep separable convolution is integrated into feature extraction model to replace the traditional convolution method.In the part of decoding network,the parameters of RefineNet decoding module are reduced,fusing the high-resolution and low-resolution features of encoder.The PASAL VOC2012 segmentation dataset is used for pre-training,and the lawn scene dataset is constructed for fine-tuning and test evaluation.The results show that on the premise of maintaining high accuracy,the proposed algorithm structure can greatly reduce the number of network parameters and improve the speed of operation,which has better comprehensive performance in lawn scene understanding tasks of the lawn mower.