基于改进UNet网络的烧结矿粘连图像分割研究

Research on image segmentation of sinter adhesion based on improved UNet network

为了解决烧结矿粒度识别过程中颗粒形状不规则、边缘模糊、高度粘连,难以识别的问题,提出了一种基于多尺度特征融合的改进UNet网络对烧结矿图像进行分割。该网络通过引入残差和密集连接,并加入SE模块(squeeze excitation block)进行特征融合,提取低层次精确的轮廓信息,解决烧结矿轮廓特征利用率低的问题;每层利用多尺度密集连接,完成非对称信息交换,有效缓解编码特征和解码特征之间的语义鸿沟问题;在解码器部分添加两个解码器,引入烧结矿轮廓特征,进行多任务学习并形成新的损失函数,利用轮廓信息来约束烧结矿对象掩码分割,提高粘连烧结矿轮廓分割能力。试验结果表明,改进的UNet网络与其他网络相比,显著提高了粘连烧结矿图像的分割准确度,为烧结矿粒度的准确识别奠定了基础。

In order to solve such problems as irregular particle shape,fuzzy edges,high adhesion and difficult to identify in the process of sinter particle size recognition,an improved UNet network based on multi-scale feature fusion to segment sinter images is proposed.By introducing residuals and dense connections and adding the SE module for feature fusion,a low-level accurate profile information is extracted and the problem of low utilization rate of sinter profile features is solved.Multi-scale dense connection is used by each layer to complete asymmetric information exchange,which effectively alleviates the semantic gap between encoded features and decoded features.Two decoders are added to the decoder part,the sinter profile feature is introduced,the multi-task learning is carried out,a new loss function is formed,and the profile information is used to constrain the mask segmentation of the sinter object,so as to improve the ability of the adhesion sinter profile segmentation.The results show that the improved UNet network significantly improves the segmentation accuracy of the adhesion sinter image compared with other networks,and lays a foundation for the accurate identification of sinter particle size.