基于改进DeepLabV3+的煤尘图像分割方法

Coal dust image segmentation method based on improved DeepLabV3+

采用传统的语义分割网络对煤尘颗粒这种较小的目标进行分割时存在深层信息易丢失、细节提取不明显等问题。针对该问题,提出了一种基于改进DeepLabV3+的煤尘图像分割方法。从3个方面对DeepLabV3+网络模型进行改进:(1)在编码器中,用CA-MobileNetV3轻量化模块代替Xception实现特征提取,确保特征提取更加细致、准确。(2)在空洞空间卷积池化金字塔(ASPP)模块中对空洞率进行改进,使其更适合小颗粒煤尘提取。(3)在解码器中引入全局注意力上采样(GAU)模块,在计算量较小时对低层特征信息进行加权,用高层特征信息指导低层特征信息,实现特征融合。GAU模块用全局上采样机制代替解码器的上采样机制,使煤尘颗粒的特征信息经过长距离传输后不衰减,更加有利于捕捉煤尘颗粒的边缘细节信息。实验结果表明,改进DeepLabV3+网络模型在煤尘数据集上的召回率为90.26%,准确度为89.23%,相比于其他网络模型,改进DeepLabV3+对煤尘特征的学习能力更强,能获取更多细节信息,并大幅缩短训练时间,对小目标的分割效果更优。

When the traditional semantic segmentation network is used to segment the small coal dust particles,there are some problems such as easy loss of deep information and unclear detail extraction.In order to solve this problem,a coal dust image segmentation method based on improved DeepLabV3+is proposed.DeepLabV3+network model is improved in three aspects.(1)In the encoder,the CA-MobileNetV3 lightweight module is used to replace Xception to achieve characteristic extraction and ensure more detailed and accurate characteristic extraction.(2)The atrous rate is improved in the atrous spatial pyramid pooling(ASPP)module to make it more suitable for extracting small coal dust particles.(3)A global attention up-sample(GAU)module is introduced into the decoder to weight the low-level characteristic information when the calculation amount is small.And the high-level characteristic information is used to guide the low-level characteristic information to realize characteristic fusion.The GAU module uses a global up-sampling mechanism to replace the up-sampling mechanism of the decoder.The characteristic information of the coal dust particles is not attenuated after longdistance transmission.And the method is more conducive to capture the edge detail information of the coal dust particles.The experimental results show that the recall rate of the improved DeepLabV3+network model on the coal dust data set is 90.26%,and the accuracy is 89.23%.Compared with other network models,the improved DeepLabV3+network model can effectively enhance the learning ability of coal dust characteristics,obtain more detailed information,greatly shorten the training time,and has better segmentation effect on small targets.