Research On Image Retrieval Based On Statistical Deep Texture Features

Image retrieval is an important research direction in the field of computer vision.Image retrieval technology can be divided into text-based image retrieval and content-based image retrieval.As text-based methods have certain limitations,content-based image retrieval technology has been widely concerned by researchers.According to the image features used,content-based image retrieval techniques can be divided into two categories: image retrieval based on low-level features and image retrieval based on deep features.With the rapid development of deep learning,image retrieval based on deep features has become a trend in the academic community,but it is undeniable that both techniques have their advantages,among which texture features have always been the focus of low-level features.Therefore,in order to explore how to use low-level and deep features to construct image descriptors with higher discriminability,this paper conducts research on image retrieval based on statistical deep texture features and proposes two methods.The main contents are as follows:(1)In view of the fact that different image features have similar visual content,this paper proposes an image retrieval method based on double-weighted deep feature descriptor.The main contributions of this method are as follows: Firstly,by studying the similarity between deep features,the ability of deep features to represent image content is improved by means of multiple aggregation of similar features.Secondly,Gabor filters are used to extract the texture features of the original image,and the data is normalized to reduce the numerical difference between texture features and deep features.Finally,by calculating the distance between the texture feature and the deep feature,the deep texture feature is further counted and the final image descriptor is generated.This method is applied to different convolutional neural networks,and all of them achieve good retrieval results and have good universality.(2)In response to the problem that the target content in deep features is not prominent enough and non-target content affects the representation ability of the features,as well as the low-frequency features having a relatively low impact on retrieval performance,this paper proposes an image retrieval method based on deep statistical feature descriptors.The main contributions of this method are as follows: Firstly,by selecting the key feature maps to construct filters,the representation ability of feature maps for image content is improved.Secondly,a spatial weighting scheme is proposed to highlight the target feature area and balance the importance of different regions.Finally,the gray level co-occurrence matrix is used to extract the texture information from the deep feature map,combined with the deep features for statistical calculations,to generate descriptors for image retrieval.Comparative experiments show that this method can effectively improve the retrieval accuracy.The method proposed in this paper has shown good retrieval performance on several popular datasets,making contributions to addressing the problem of combining texture features and deep features for image retrieval.