随着医学影像处理技术的发展，利用深度学习技术对血细胞图像中的白细胞进行识别具有较高的实用价值。然而，不同环境下显微镜采集的血细胞图像存在光照不均及白细胞特征不明显的问题，导致白细胞的自动化分割分类具有挑战性。同时，当卷积神经网络在传统的CPU或GPU上运行时有计算吞吐量低及速度慢等问题。因此，论文提出一种白细胞分割分类算法及ZYNQ加速的方法，其主要研究内容包括：

（1）针对不同环境下采集的血细胞图像光照不均及白细胞特征不明显的问题，提出一种基于AHE-MSRCR校正的图像预处理算法。该算法结合自适应直方图均衡化和带色彩恢复函数的多尺度视网膜增强算法对血细胞图像进行色彩校正，从而得到背景亮度趋于统一、白细胞特征明显的血细胞图像。

（2）针对血细胞图像背景复杂的问题，提出一种基于改进U-net的白细胞分割算法。该算法利用改进U-net卷积神经网络将白细胞从血细胞图像中分割出来，去除背景仅保留白细胞。在原始的U-net网络中，针对不同种类的白细胞之间存在一定相似性的问题，在4个跳跃连接中加入卷积注意力机制模块，将通道注意力机制和空间注意力机制串联来消除跳跃连接操作中的语义歧义；针对血细胞图像中类不平衡的问题，提出Focal loss和Dice loss相结合的损失函数。

（3）针对在CPU或GPU上运行卷积神经网络时计算吞吐量低及速度慢的问题，提出基于ZYNQ的YOLOv3目标检测算法的模型移植。采用Xilinx特有的Vitis-AI工具对YOLOv3网络进行量化，在损失极少准确率的情况下达到模型的压缩，再利用编译器完成部署。

论文的实验结果表明，所提的分割算法对于类不平衡、特征不明显的白细胞有较好的分割结果，平均交并比达到91.89%，准确率达到99.53%。YOLOv3网络的ZYNQ移植的平均精度均值为92.09%，且相较于GPU推断达到20倍的加速效果。

With the development of medical image processing technology, the identification of leukocytes in blood cell images using deep learning techniques has a high practical value. However, blood cell images captured by microscopes in different environments suffer from uneven illumination and inconspicuous leukocyte features, making automated segmentation and classification of leukocytes challenging. At the same time, convolutional neural networks have problems such as low computational throughput and slow speed when running on conventional CPUs or GPUs. Therefore, the thesis proposes a leukocyte segmentation and classification algorithm and ZYNQ acceleration method:

(1) An image pre-processing algorithm based on AHE-MSRCR correction is proposed to address the problems of uneven illumination and inconspicuous leukocyte features in blood cell images acquired under different environments. The algorithm combines Adaptive Histogram Equalization(AHE) and a multi-scale retinal enhancement algorithm with color recovery (MSRCR) to color correct the blood cell images, resulting in blood cell images with uniform background brightness and distinct leukocyte features.

(2) An improved U-net-based leukocyte segmentation algorithm is proposed to address the problem of complex backgrounds in blood cell images. The algorithm uses the improved U-net convolutional neural network to segment the leukocytes from the blood cell images, removing the background and keeping only the leukocytes. In the original U-net network, to address the problem of certain similarity between different types of leucocytes, the Convolutional Block Attention Module (CBAM) is added to the four skip connections, and the Channel Attention Module (CAM) and Spatial Attention Module (SAM) are connected in series to eliminate the semantic ambiguity in the skip connection operation; to address the problem of class imbalance in blood cell images, a combination of Focal loss and Dice loss is proposed for the loss function.

(3) To address the problems of low computational throughput, slow speed and high power consumption when running convolutional neural networks on CPUs or GPUs, a ZYNQ-based model port of the YOLOv3 target detection algorithm is proposed. Xilinx's unique Vitis-AI tool is used to quantize the YOLOv3 network to achieve model compression with minimal loss of accuracy, and then the compiler is used to complete the deployment.

The experimental results of the paper show that the proposed segmentation algorithm has better segmentation results for class-imbalanced and poorly characterized leukocytes, with a mean intersection over union of 91.89% and an accuracy of 99.53%. The mean average precision of the ZYNQ transplantation of the YOLOv3 network is 92.09% and achieves a 20-fold speedup compared to GPU inference.

[1]杨勇超. 基于卷积神经网络的白细胞分类识别方法研究[D]. 西安:西安科技大学,2021.

[2]马力,帅仁俊,刘文佳,等. 基于改进的残差神经网络的白细胞分类[J]. 计算机工程与设计,2020,41(10):2982-2987.

[3]王静,孙紫雲,郭苹,等. 改进YOLOv5的白细胞检测算法[J]. 计算机工程与应用,2022,58(09):134-142.

[4]陈雪云,黄小巧,谢丽. 基于多尺度条件生成对抗网络血细胞图像分类检测方法[J]. 浙江大学学报(工学版),2021,55(09):1772-1781.

[5]徐晓涛,孙亚东,章军. 基于YOLO框架的血细胞自动计数研究[J]. 计算机工程与应用,2020,56(14):98-103.

[6]王亚品,曹益平,付光凯,等. 基于深度卷积神经网络的人体外周血白细胞显微图像分类[J]. 光电子·激光,2019,30(05):546-555.

[7]赵子豪. 基于卷积神经网络的白细胞图像检测及分类技术[D]. 西安:西北大学,2019.

[8]Lu Y,Qin X,Fan H,et al. WBC-Net: A white blood cell segmentation network based on UNet++ and ResNet[J]. Applied Soft Computing Journal,2021,101(2021):1-11.

[9]Bai H,Zhang T,Lu C, et al. Chromosome Extraction Based on U-net and YOLOv3[J]. IEEE ACCESS,2020,8(2020):178563-178569.

[10]Sapna S,Renuka A. Computer-aided system for Leukocyte nucleus segmentation and Leukocyte classification based on nucleus characteristics[J]. International Journal of Computers and Applications,2020,42(6):622-633.

[11]李创,刘宗林,刘胜,等. 快速卷积算法的综述研究[J].计算机工程与科学,2021,43(10):1711-1719.

[12]Wang Y,Cao Y. Human peripheral blood leukocyte classification method based on convolutional neural network and data augmentation[J]. Medical physics,2020,47(1): 142-151.

[13]Qin F,Gao N,Peng Y,et al. Fine-grained leukocyte classification with deep residual learning for microscopic images[J]. Computer Methods and Programs in Biomedicine, 2018,162(2018):243-252.

[14]周航,贺雅娟. 一种基于FPGA的感知量化卷积神经网络加速系统设计[J]. 电子世界,2021,11(2021):164-165.

[15]谢坤鹏,卢冶,靳宗明,等. FAQ-CNN:面向量化卷积神经网络的嵌入式FPGA可扩展加速框架[J].计算机研究与发展,2022,59(07):1409-1427.

[16]吴健,顾明剑,曾长紊,等. 基于ZYNQ的卷积神经网络加速器设计[J]. 计算机工程与设计,2022,43(06):1572-1581.

[17]许杰,张子恒,王新宇,等. 一种基于Zynq的CNN加速器设计与实现[J]. 计算机技术与发展,2021,31(11):108-113+121.

[18]张坤宁,赵烁,孙庆斌,等. 基于FPGA的多核可扩展卷积加速器设计[J]. 计算机工程与设计,2021,42(06):1592-1598.

[19]熊伟,黄鲁. 基于ZYNQ和CNN模型的服装识别系统[J]. 计算机系统应用,2019,28(11):101-106.

[20]Sledevič T,Serackis A. mNet2FPGA: a design flow for mapping a fixed-point CNN to Zynq SoC FPGA [J]. Electronics,2020,9(11):18-23.

[21]Hao L,Lv H,Zhang S et al. Research on Dynamic Reconfiguration Technology of Neural Network Accelerator Based on Zynq[C]//Journal of Physics. Conference Series. UK:IOP Publishing, 2020, 1650(3): 1-9.

[22]Al Koutayni M R,Rybalkin V,Malik J,et al. Real-time energy efficient hand pose estimation: A case study[J]. Sensors, 2020, 20(10): 28-40.

[23]Amara A B,Mohamed A,Pissaloux E,et al. Zynq FPGA based and Optimized Design of Points of Interest Detection and Tracking in Moving Images for Mobility System[J]. International Journal of Advanced Computer Science and Applications, 2018, 9(10):1-11.

[24]文海琼,李建成. 基于直方图均衡化的自适应阈值图像增强算法[J]. 中国集成电路,2022,31(03):38-42+71.

[25]Ketcham D J,Lowe R W,Weber J W. Image enhancement techniques for cockpit displays[R]. Culver City:Ca Display Systems Lab, 1974.

[26]陈翔. 基于HSV空间模型的阴影去除方法研究[J]. 信息技术与信息化,2022(4):84-87.

[27]田宁,程莉,元海文,等. 基于Retinex模型的水下图像增强方法[J]. 中国科技论文,2022,17(11):1281-1288.

[28]胡其霖. 基于Retinex算法的去雾型车牌识别系统的设计与实现[D]. 武汉:华中科技大学,2020.

[29]李灿林,朱金娟,刘金华,等. 一种自适应SSR的雾天低照度图像增强方法[J]. 计算机应用与软件,2022,39(9):233-239+268.

[30]王艳霞. 多尺度Retinex算法的园林景观图像优化研究[J]. 信息技术,2022,46(9):25-29.

[31]云海姣,夏洋. 结合自适应Gamma变换和MSRCR算法的低光照图像增强方法[J]. 中国科技论文,2022,17(11):1245-1253.

[32]Mnih V,Heess N,Graves A. Recurrent models of visual attention[J]. Advances in neural information processing systems, 2014, 27(1):1-7.

[33]Vaswani A,Shazeer N,Parmar N,et al. Attention is all you need[J]. Advances in neural information processing systems, 2017, 30(1):1-5.

[34]Wang X,Girshick R,Gupta A,et al. Non-local neural networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. USA:IEEE, 2018: 7794-7803.

[35]Hu J,Shen L,Sun G. Squeeze-and-excitation networks[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. USA:IEEE, 2018: 7132-7141.

[36]Li X,Wang W,Hu X,et al. Selective kernel networks[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. USA:IEEE, 2019: 510-519.

[37]Woo S,Park J,Lee J Y,et al. Cbam: Convolutional block attention module[C]//Proceedings of the European conference on computer vision (ECCV). USA:IEEE, 2018: 3-19.

[38]朱磊,仝超,董亮,等. 基于残差网络和注意力机制的步态识别算法[J]. 电讯技术,2022,62(12):1722-1728.

[39]刘世川,王庆,唐晴,等. 基于多特征提取与注意力机制深度学习的高分辨率影像松材线虫病树识别[J]. 林业工程学报,2022,7(01):177-184.

[40]Lin T Y,Goyal P,Girshick R,et al. Focal loss for dense object detection[C]//Proceedings of the IEEE international conference on computer vision. USA:IEEE, 2017: 2980-2988.

[41]曹建荣,张玉婷,朱亚琴,等. 基于改进MDNet的视频目标跟踪算法[J]. 计算机系统应用,2022,31(05):277-284.

[42]Milletari F,Navab N,Ahmadi S A. V-net: Fully convolutional neural networks for volumetric medical image segmentation[C]//2016 fourth international conference on 3D vision (3DV). USA:IEEE, 2016: 565-571.

[43]Ronneberger O,Fischer P,Brox T. U-net: Convolutional networks for biomedical image segmentation[C]//Medical Image Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18. Germany:Springer International Publishing, 2015: 234-241.

[44]陈卓,邓兴广,缪晓强,等. 基于动态DR的肺区域智能分割方法[J]. 中国医疗器械信息,2022,28(13):32-37+136.

[45]Redmon J,Farhadi A. YOLOv3: An Incremental Improvement[C]//IEEE Conference on Computer Vision and Pattern Recognition. USA:IEEE, 2018:2311-2314.

[46]Simonyan K,Zisserman A. Very deep convolutional networks for large-scale image recognition[C]//International Conference on Learning Representations (ICLR 2015). Ethiopia:OpenReview.net, 2015: 1-14.

[47]後山晃彦, 渡邊誠也, 名古屋彰, et al. Vitis AIを用いたCNN実装[J]. 電子情報通信学会技術研究報告. 情報システムソサイエティ各誌,2021,121(175):13-18.

[48]Redmon J,Farhadi A. YOLO9000: better, faster, stronger[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. USA:IEEE, 2017: 7263-7271.

[49]Foody G M. Status of land cover classification accuracy assessment[J]. Remote sensing of environment, 2002, 80(1): 185-201.

[50]Chen L C,Zhu Y,Papandreou G, et al. Encoder-decoder with atrous separable convolution for semantic image segmentation[C]//Proceedings of the European conference on computer vision (ECCV). USA:IEEE, 2018: 801-818.

[51]Zhao H,Shi J,Qi X,et al. Pyramid scene parsing network[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. USA:IEEE, 2017: 2881-2890.

[52]Liu Z,Li J,Shen Z,et al. Learning efficient convolutional networks through network slimming[C]//Proceedings of the IEEE international conference on computer vision. USA:IEEE, 2017: 2736-2744.