摘要
针对肺炎图像中病灶组织与正常组织难以区分,导致的肺炎检测准确率低的问题,提出一种基于改进Faster R-CNN的肺炎目标检测算法。通过CRP-CLAHELS的流程在增强图像对比度的同时提取图像的边缘特征,提出IN-ResNet50网络作为特征提取主干网络,提取更丰富的图像特征。在此基础上,引入Soft-NMS改进候选框合并策略,提高网络在多个目标肺炎区域下的检测准确率。在RSNA数据集上的实验结果表明,该算法相比Faster R-CNN平均精度均值提高7.26%,与其它目标检测主流算法SSD、YOLOv3相比平均精度均值分别提高8.83%、7.02%,验证了其有效性。
Aiming at the problem that the focus tissue and normal tissue in the pneumonia image are difficult to distinguish,which results in the low accuracy of pneumonia detection,a pneumonia target detection algorithm based on improved Faster R-CNN was proposed.The CRP-CLAHELS process was used to enhance the contrast of the image while extracting the edge features of the image,and the IN-ResNet50 network was proposed as the feature extraction backbone network to obtain more abundant image features.On this basis,Soft-NMS was introduced to improve the candidate frame merging strategy to improve the detection accuracy of the network in multiple target pneumonia regions.Experimental results on the RSNA data set show that the average accuracy of this algorithm is increased by 7.26%compared with that of Faster R-CNN.Compared with other mainstream target detection algorithms SSD,YOLOv3,the average accuracy is increased by 8.83%and 7.02%respectively,which verifies the effectiveness of the algorithm.
作者
宋雯琦
赵荣彩
姜旭
刘艳青
SONG Wen-qi;ZHAO Rong-cai;JIANG Xu;LIU Yan-qing(School of Information Engineering,Zhengzhou University,Zhengzhou 450001,China;National Supercomputing Center in Zhengzhou,Zhengzhou University,Zhengzhou 450052,China)
出处
《计算机工程与设计》
北大核心
2023年第7期2087-2092,共6页
Computer Engineering and Design
基金
河南省重大科技专项基金项目(201400211300、201400210600)。
