基于共现流增强双向金字塔卷积网络的密集液滴识别

基于深度学习的数字聚合酶链式反应（PCR）液滴识别对PCR图像中的目标进行高阶语义建模，能够减少人工参与特征设计和筛选带来的误差，但忽略了目标的低层物理结构和几何外观细节信息，且在特征建模的过程中重复使用了大量冗余信息，对特征的表征能力有待改善。提出一种共现流增强双向金字塔卷积网络（CoF-BiPCN）框架用于PCR液滴识别和统计。为增强金字塔的内部和层间相关性，设计具有时空分支的双向金字塔卷积网络，从正反2个方向对金字塔卷积网络得到的多尺度特征进行聚合，模拟PCR图像中液滴的上下文语义以及不同层级的细节信息，以捕获液滴的物理外观等低层信息。同时，设计切片的共现注意力（SCo-AN）模块，将不同尺度的高低层信息在不同的切片子空间中进行共享聚合，并交叉传递到不同分支的BiPCN中，强化高低层特征信息的交互和依赖关系，进一步增强信息流对PCR图像上液滴的表征，实现低层和高阶信息流的共享与交叉聚合。实验结果表明，CoF-BiPCN具备良好的识别性能，准确率和平均精度均值分别达到84.74%和45.09%，与Cascade RCNN模型相比分别提高4.3和3.12个百分点。

Dense Droplet Identification Based on Co-occurrence Flow Enhanced Bidirectional Pyramidal Convolution Network

A pyramid network faces a range of problems in recognizing digital Polymerase Chain Reaction(PCR) droplets, such as ignoring the physical appearance of droplets, and the internal structure and interlayer correlation of pyramids.To address these issues, this study proposes a co-occurrence flow enhanced Bidirectional Pyramid Convolution Network(BiPCN) framework for PCR droplet recognition and statistics framework.First, to enhance the internal and inter-layer correlation of the pyramid, a BiPCN with spatio-temporal branches is designed to capture the low-level information such as the physical appearance of droplets, so as to model the high-order semantics and context information of droplets.Second, Slice Co-occurrence Attention(SCo-AN) module is designed to further enhance the characterization of droplets on PCR images by information flow, so as to realize the sharing and cross polymerization of low-level and high-order information flow.The experimental results show that the proposed method has good recognition performance, that is, the accuracy and mean Average Precision(mAP) reach 84.74% and 45.09% respectively, which is more accurate than that of Cascade RCNN model, the accuracy and mAP are increased by 4.3 and 3.12 percentage points, respectively.