摘要：目的 针对基于内容的图像检索存在着低层视觉特征与用户对图像理解的高层语义不一致、图像检索的精度较低以及传统的分类方法准确度低等问题，提出了一种基于卷积神经网络和相关反馈支持向量机的遥感图像检索方法。方法 该方法通过对比度受限直方图均衡化算法对遥感图像进行预处理，限制遥感图像噪声的放大，采用自学习能力良好的卷积神经网络对遥感图像进行多层神经网络的监督学习提取丰富的图像特征，并将支持向量机作为基分类器，根据测试样本数据到分类超平面的距离进行排序得到检索结果，最后采用相关反馈策略对检索结果进行重新调整。结果 在UC Merced Land-Use Data Set遥感图像数据集上进行图像检索实验，在mAP(mean average precision)精度指标上，当检索返回图像数为100时，本文方法相对于LSH(Locality Sensitive Hashing)方法提高了29.4%，相对于DSH(Density Sensitive Hashing)方法提高了34.7%，相对于EMR(Efficient Manifold Ranking)提高了60.8%，相对于未添加反馈和训练集筛选的SVM(Support Vector Machine)方法提高了3.5%，对于平均检索速度，本文方法相对于对比方法中mAP精度最高的方法提高了4倍，针对复杂的遥感图像数据，本文方法的检索效果较其他方法表现出色。结论 本文提出了一种以距离评价标准为核心的反馈策略提高检索精度；并采用多距离结合的Top-k排序方法合理筛选训练集提高检索速度，本文方法可以广泛的应用于人脸识别、目标跟踪等领域，对提升检索性能具有重要意义。
Remote Sensing Image Retrieval based on Deep Learning and Relevance Feedback
pengyanfei,songxiaonan,wuhong,zilingling(School of Electronic and Information Engineering, Liaoning Technical University)
Abstract: Objective The traditional content-based image retrieval method is only retrieving and analyzing the features of the low layers, such as color, texture and shape, which exist in the image. Therefore, there is a low level of visual features which are inconsistent with the high-level semantic meaning of the user''s understanding of the image, resulting in the production of the "semantic gap" phenomenon, which further leads to the low accuracy of image retrieval. What''s more,it can not meet the user''s demand for high accuracy retrieval,and remote sensing images have rich information, complex content, and high dimensionality. Only analyzing the low-level features has greatly refused the accuracy of image retrieval. Therefore, choosing an appropriate image feature extraction is the key step to achieve high-accuracy retrieval.At the same time, the traditional classification method is less accurate in image classification. How to select a high accuracy image classification method is also essential. A remote sensing image retrieval method based on convolution neural network and relevance feedback support vector machine is proposed. Method This method can preprocess remote sensing image by contrasting limited histogram equalization algorithm, limiting the noise magnification of remote sensing image, avoiding the influence of noise interference on the retrieval precision, and on the basis of the GoogLeNet convolution neural network model with good self-learning ability, the multi-layer neural network of remote sensing image is supervised and studied, and the rich features of remote sensing images are extracted, and the problem of "semantic gap" in the content based image retrieval method is solved, The original dataset is divided into training set and test set, and selecting the training set reasonably is the basis for the best classification. If there are too many samples in other categories in the training set, the determination of the classification Hyperplane will be greatly affected. A multi-distance combined Top-k sorting method is proposed to rationally screen the original training set. The image closest to the query one will be used as the training set. On the one hand, the method saves a lot of time for the subsequent determination of the optimal hyperplane, on the other hand, most dissimilar images are filtered out to avoid the influence of more dissimilar images on the classification results, Then the support vector machine is used as the basic classifier, and the optimal hyperplane is trained according to the training set samples. The retrieval results are sorted according to the distance between the test sample data and the classified hyperplane. Finally, a feedback of the distance evaluation standard is proposed to update the retrieval results with the distance evaluation standard. The strategy readjusts the results of the experiment. For one thing, the method uses a small sample marking method to mark the counterexample images to avoid too many markers and losing the meaning of the retrieval. For another, there is no need to retrain the optimal Hyperplane of the support vector machine to avoid unnecessary time waste. Only multiple iterations are used to update the retrieval results, and one feedback can achieve the desired results.Result The image retrieval experiments are performed on the UC Merced Land-Use Data Set remote sensing image dataset .Experiments show that the mAP(mean average precision) of the proposed method increased by 29.4 compared with the LSH(Locality Sensitive Hashing) method, is 34.7% higher than the DSH (Density Sensitive Hashing) method, which is 60.8% higher than the EMR(Efficient Manifold Ranking), and is 3.5% higher than the SVM(Support Vector Machine) method without feedback and training set screening,whereas the number of retrieved images are 100. For the average retrieval speed, this method is 4 times higher than the method with the highest mAP accuracy in the comparison method.What’s more, for the average recall rate and the average precision rate, this method is also higher than the comparison method, which shows that this method can not only improve the retrieval accuracy, but also improve the retrieval speed. For complex remote sensing image data, the retrieval effect of this method is better than those of other methods. Conclusion In order to improve the retrieval accuracy, a new feedback strategy is proposed in this paper. We use small sample markers for the poor retrieval results and use the distance evaluation standard as the core to carry out the many iterations, and one time feedback can achieve better retrieval results. And in terms of speed increase, this paper proposes a multi-distance combined Top-k sorting method, which reduces the time of supporting vector machines to train the optimal Hyperplane by rationally selecting the training sample set, and then improves the retrieval speed. This method can be widely applied to face recognition, target tracking and other fields, and it is of great significance to improve retrieval performance.