结合主体检测的图像检索方法

发布时间：2018-09-10 07:25

【摘要】：为解决图像背景复杂造成图像检索效果差的问题,提出了一种结合主体检测的图像检索方法。该方法首先训练用于目标检测的深度卷积神经网络模型,利用训练好的模型检测查询图像中的物体类别、类别概率和其所在区域坐标及特征。根据物体的类别概率和其所在区域的坐标判断图像主体后,在数据库中查找和主体类别相同的图像。计算查询图像与检索的同类别图像之间区域特征的余弦距离,结合类别概率对所有检索图像进行打分排序,返回分值最高的前10幅图像作为检索结果。最后在VCO2007数据集和自己收集的书页数据集上进行算法验证。实验结果表明,在随机选取的1 000幅测试图片检索结果的全正确率为96.5%,比现有方法提升了6.6个百分点。本文方法可有效排除图像背景的干扰,得到更加准确的检索结果和定位精度。
[Abstract]:In order to solve the problem of poor image retrieval effect caused by complex image background, an image retrieval method combining subject detection is proposed. The method first trains the deep convolution neural network model used for target detection, and uses the trained model to detect the class of objects, the probability of category, the coordinates and features of the region in which the objects are queried. After judging the subject of the image according to the probability of the object's category and the coordinates of the region in which it is located, the image subject is found in the database which is the same as the subject's category. The cosine distance between the queried image and the same category image is calculated, and all the retrieval images are sorted with the category probability, and the top 10 images with the highest score are returned as the retrieval results. Finally, the algorithm is validated on the VCO2007 dataset and the book page data set collected by ourselves. The experimental results show that the total correct rate of the retrieval results of 1 000 images selected at random is 96. 5%, which is 6.6 percentage points higher than that of the existing methods. This method can effectively eliminate the interference of image background and obtain more accurate retrieval results and location accuracy.
【作者单位】：城市道路交通智能控制技术北京市重点实验室;北方工业大学理学院;
【基金】：北京市属高等学校青年拔尖人才培育计划资助项目(No.CIT&TCD201404009) 科技创新服务能力建设—科技成果转化—提升计划项目(PXM2016_014212_000036)
【分类号】：TP391.41

【相似文献】