基于深度学习的结晶过程原位图像分割方法

Deep learning based in⁃situ image segmentation method for crystallization process

针对结晶过程中晶体原位图像存在的目标像素低、晶体重叠以及背景干扰等导致的分割检测困难等现象,提出一种改进YOLOv8的晶体原位图像分割方法。为了提高模型的分割检测性能,首先引入高效多尺度注意力机制(EMA),增强模型的感知能力;其次使用空间到深度非跨步卷积(SPD-Conv)对原卷积块进行改进,在提升对低像素、小目标晶体分割精度的同时降低了模型的计算量;最后采用高效交互比(EIoU)损失函数优化对遮挡和重叠目标的检测效果。实验结果表明,文中提出的算法晶体检测精度(mAP)达到71.3%,精度提高了5.3%,浮点运算量降低了1.9 GFLOPs。此外,该方法对改善结晶图像质量较差以及存在晶体重叠的工况也具有明显的优势。

A crystal in⁃situ image segmentation method based on improved YOLOv8 is proposed to address the crystal segmentation difficulties caused by low pixels,overlapped crystals and background interference of the in⁃situ images in the process of crystallization.To improve the segmentation detection performance of the model,the efficient multi⁃scale attention(EMA)mechanism is introduced to enhance the perception ability of the model.Subsequently,the original convolutional block is improved by the space⁃to⁃depth non⁃strided convolution(SPD⁃Conv)method,so as to enhance the segmentation accuracy of crystals(the objects)with low pixel and small size while reducing the computational effort of the model.Finally,the efficient intersection over union(EIoU)loss function is used to optimize the detection results of the occluded and overlapped crystals(the objects).The experimental results show that the crystal detection accuracy(mAP)of the proposed algorithm reaches 71.3%,its accuracy is improved by 5.3%,and its floating⁃point calculation burden is reduced by 1.9 GFLOPs.In addition,the proposed method has advantages in improving the quality of crystal image and eliminating the crystal overlap.