作者:黄志添，谢怡宁，赵晶，何勇军
Authors:HUANGZhitian,XIEYining,ZHAOJing,HEYongjun摘要:针对病原微生物的检测研究较少且目标尺寸相差较大、背景复杂问题 ,提出了 一 种融合递归特征金字 塔的病原微生物多尺度检测方法 。考虑到有些病原微生物的特征与正常细胞有 一定的相似性 ,在 YOLOv5 特征提 取阶段引入混合注意力模块 CBAM ,可以增强病原微生物的显著度 , 降低了病原微生物的误诊率 。 由于尺寸较小 的病原微生物在下采样过程会丢失重要特征 ,在特征融合阶段使用递归特征金字塔 RFP 和增加小目标的检测层 , 可以对小目标的特征进行二次提取 , 丰富语义信息 ,提高了小目标的检出率 ,并在二 次特征提取模块中使用深度可 分离卷积 DSConv 替换普通卷积 Conv , 降低了模型的参数量 ,提升了性能 。实验表明 ,本文算法在病原微生物数据 集上的 mAP 达到了 74. 9% , 相比基线网络 YOLOv5 提升了 4. 9% , 小目标滴虫的 mAP 也提高了 6. 2% 。 同时在 PascalVOC2007、PascalVOC2012 和 CDetector 公开数据集上比基线网络 YOLOv5 分别提升了 1. 9% 、3. 1% 和 8% 。
Abstract:Aiming at the problems of less research on the detection of pathogenic microorganisms , large differences in target size and complex background , a multi-scale detection method of pathogenic microorganisms fused with recursive feature pyramid was proposed. Considering that the characteristics of some pathogenic microorganisms are similar to normal cells , introducing the mixed attention module CBAM in the feature extraction stage of YOLOv5 can enhance the significance of pathogenic microorganisms and reduce the misdiagnosis rate of pathogenic microorganisms. Since the pathogenic microorganisms with small size will lose important features in the downsampling process , the recursive feature pyramid RFP is used in the feature fusion stage and the detection layer of the small target is added. The detection rate of the target is improved , and the depthwise separable convolution DSConv is used in the secondary feature extraction module to replace the ordinary convolution Conv , which reduces the number of parameters of the model and improves the performance. Experiments show that the mAP of the algorithm in this paper on the pathogenic microorganism dataset has reached 74. 9% , which is 4. 9% higher than the baseline network YOLOv5 , and the mAP of the small target Trichomonas has alsoincreased by 6. 2% . At the same time , the public datasets of PascalVOC2007 , PascalVOC2012 and CDetector are improved by 1. 9% , 3. 1% and 8% respectively compared with the baseline network YOLOv5.

PDF全文下载地址:

可免费Download/下载PDF全文