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。

Aimed at the problem-poor result of image retrieval arising from the complexity of image background, a kind of image retrieval method combined with subject detection was put forward. This method has initially trained the deep Convolutional Neural Network (CNN) model used in object detection and used the model detection well trained to inquiry the object class, class probability and the coordinate and feature of region where it was placed in the image. After the image subject estimated in accordance with the objects class probability and coordinate of region where it was placed, the image similar to the subject in the database was found. The cosine distance of region feature between the image inquired and similar image retrieved was caculated, combined with the class probability to carry out grading and sorting for all images retrieved and returned the top 10 images with the highest scores to be as the retrieved result. Finally verification of algorithm was conducted on VCO2007 dataset and paper dataset collected by myself. The experiment result shows that the total accuracy for retrieved result of 1 000 test images is 96.5%, which has raised 6.6 percent points than the existing method. The proposed method can effectively exclude the disturbance of image background and get more accurate retrieved result and location accuracy.