Abstract:

To realize online morphology information measurement of HMX molding powders, the particle imaging probe and acquisition system were developed. Experimental studies were conducted to capture the images of suspended particles of molding powders. Using Mask RCNN as framework, an image processing method based on improved Swin Transformer was proposed, and the CA-Swin Transformer structure was proposed by connecting channel attention module (CAM) and the window-based multi-head self-attention (W-MSA) in parallel, which was utilized to reasonably allocate the attention degree of image channels. The particle recognition network (PRNet) was further established by combining the proposed feature enhancement module (FEM) with CA-Swin Transformer, effectively improving the recognition accuracy of particle images. The PRNet was trained and tested with the labeled image dataset of HMX molding powders. The obtained results showed that the AP, AP₅₀ and AP₇₅ of PRNet reached 62.3%, 84.4% and 72.5%, respectively. The relative errors of recognized feature particle sizes D₁₀, D₅₀, D₉₀ and D_max relative to manual labeling were -4.788%, -0.770%, -0.272% and 0.313%, respectively, outperforming baseline network Mask RCNN and its variant with the Swin Transformer backbone. Moreover, PRNet exhibited a better recovery ability for the occluding part of overlapping particles. The absolute relative errors of circularity, Feret diameter and aspect ratio of overlapping particles relative to manual labeling were less than 8%, 4% and 5%, respectively.

Key words: two-phase flow, particles, Octogen (HMX) molding powders, in-situ measurement, deep learning

摘要：

针对原位、在线获取奥克托今（HMX）造型粉形貌信息的需求，设计了颗粒图像探针及采集系统，获取了造型粉悬浮颗粒图像。以掩模区域卷积神经网络（Mask RCNN）为框架，提出了一种基于改进Swin Transformer的图像处理方法，通过并联通道注意力模块和窗口自注意力机制提出CA-Swin Transformer结构以合理分配图像通道的关注度，并进一步结合特征增强模块建立了一种颗粒识别网络（particle recognition network，PRNet），有效提高了颗粒识别精度。以标注的HMX造型粉图像数据集对PRNet进行训练与测试。结果表明，PRNet的AP、AP₅₀和AP₇₅分别达到了62.3%、84.4%和72.5%；识别特征粒径D₁₀、D₅₀、D₉₀和D_max与人工标注值的相对误差分别为-4.788%、-0.770%、-0.272%和0.313%，均优于基准网络Mask RCNN及其以Swin Transformer为主干网络的变体。此外，对重叠颗粒的遮挡部分进行复原，重叠颗粒圆形度、Feret直径和长宽比与人工标注绝对相对误差分别小于8%、4%和5%。

关键词: 两相流, 颗粒, 奥克托今造型粉, 原位测量, 深度学习