| Image retrieval is a very challenging research direction in pattern recognition.Among them,feature extraction and compact feature description are important components of image retrieval technology.Traditional image retrieval technology is mainly composed of two parts:(1)Textbased image retrieval(TBIR);(2)Content-based image retrieval(CBIR).TBIR technology has limitations and it is difficult to accurately describe image content.Although CBIR can convey image information through low-level visual features,it still has many shortcomings in high-level semantic expression.In recent years,Convolutional Neural Networks(CNN)have achieved excellent performance in tasks such as image retrieval and image classification.In convolutional neural networks,the convolutional layer or pooling layer of the pre-trained CNN model is usually used to activate the high-level semantic information of the image.Although it is superior to traditional image retrieval technology in terms of semantic expression,the improvement of retrieval performance brought by it is very limited.On the basis of pre-trained convolutional neural network,targeted retraining can not only obtain efficient feature representation,but also has significant advantages in expressing image depth semantics.For different retrieval tasks,this article gives different solutions.The main contents are as follows:1.In CBIR,this paper proposes a multi-stage feature integration image retrieval method.Firstly,the input image is converted from the RGB color space to the HSV color space that conforms to human visual perception,and the image color and color difference are calculated.Secondly,the edge features of the image are obtained through simple color difference calculation.Finally,the multi-stage feature integration scheme is adopted.Combining low-level visual features represents image content.The multi-stage feature integration scheme can not only describe the image color and edge attributes,but also can represent the image area and spatial arrangement information well.The experimental results show that in the traditional image retrieval data sets(Corel-10 K,GHIM-10 K and Corel-5K),the multi-stage feature integration scheme proposed in this paper has excellent discrimination capabilities.2.Although the low-level visual features of the combined image can represent the visual content of the image well,the multi-stage feature integration scheme still belongs to the manual feature extraction method,and it is difficult to truly deal with the image semantic problem.In order to better alleviate the difference in image semantics,this paper proposes a deep feature image retrieval method based on end-to-end framework for fine-tuning and retraining.On the basis of pre-trained convolutional neural networks(Alex Net,VGGNet and Goog Le Net),the Siamese network architecture is used for contrastive loss,and the performance of different network benchmarks is compared.And through learning whitening parameters and weighted expansion query method to further improve the retrieval performance.At the same time,in the selection of training data set,this paper adopts training data set closer to the example image retrieval task,and the learning of network parameters is more targeted.Although the multi-stage feature integration scheme proposed in this paper has a good performance in image retrieval,the manual feature extraction method is not suitable for instance image retrieval.The performance on the instance image retrieval datasets(Oxford5k,Paris6 k and Holidays)proves that the deep feature image retrieval method based on fine-tuning retraining is not only superior to the traditional content-based image retrieval method,but also superior to the pre-training convolutional neural network,it can better deal with the problem of "semantic gap" at the same time. |
