基于融合边缘优化的煤矿图像语义分割方法

由于煤矿井下环境恶劣,使得煤矿井下视频获取的图像严重降质,而现有的基于深度学习的语义分割模型在图片清晰化后存在边缘分割模糊的问题,提出了一种采用融合边缘优化模块处理边界信息并运用门控卷积层连接传统特征提取模块并行处理信息的方法;为监督学习轮廓信息,采用二元交叉熵损失函数提高学习效果,并与常规分支的损失函数共同优化模型分割效果。试验结果表明:对已完成清晰化的煤矿井下图像进行语义分割任务时,基于融合边缘优化模块的方法与其他方法相比整体语义分割精度得到提升并且边缘分割精度更高。

Images semantic segmentation method based on fusion edge optimization

Due to the poor underground environment, which makes the images of underground video in coal mines seriously degraded. However, the existing semantic segmentation model based on deep learning has the problem of fuzzy edge segmentation after image sharpness. A new method is proposed, which uses the fusion edge optimization module to process the boundary information and uses the gated convolution layer to connect the traditional feature extraction module to process the information in parallel. In order to supervise the learning of contour information, the binary cross entropy loss function was used to improve the learning effect, and the segmentation effect was optimized together with the loss function of conventional branches. The experimental results show that compared with other methods, the method based on the fusion edge optimization module improves the overall semantic segmentation accuracy and edge segmentation accuracy when the semantic segmentation task of the clarified coal mine image is performed.