Research On Deep Hashing Method For Remote Sensing Image Retrieval

In forestry remote sensing,remote sensing images contain a wealth of information on forestry-related land features and terrain elements,such as forests,shrubs,herbaceous plants,and roads.Retrieving the required images quickly and accurately from massive remote sensing image datasets using hash techniques has become an urgent challenge in the forestry remote sensing field.Currently,most deep hash methods are supervised learning-based.However,due to the relatively small number of labeled remote sensing images and the significant manpower and material resources required for annotation,the retrieval performance of supervised hash methods is poor in scenarios with insufficient labeled data.Compared with natural images,remote sensing images exhibit more pronounced intra-class diversity and inter-class similarity.Even if Image Net pre-trained models are used to extract features from remote sensing images,the retrieval performance cannot be significantly improved.Although deep unsupervised hash methods do not require the use of real label information,due to the local constraints of convolutional neural networks,the feature extraction of remote sensing images is not obvious enough,resulting in unreliable pseudo labels.Moreover,existing unsupervised hash research still faces issues such as the same training weights for image pairs.To address these problems,this thesis proposes two improved deep hash methods for the retrieval of remote sensing images with few or no labels.The main contributions of this thesis are as follows:(1)This thesis proposed a deep self-supervised contrastive hashing(DSCH)method for retrieving remote sensing images with few labels.Based on the self-supervised contrastive learning framework,a multi-level dual-branch fusion network is constructed as a new encoder.The parallel structure extracts and fuses global and local features at different levels and semantic scales,enhancing the encoder’s feature extraction ability.The accuracy of the feature representation is further improved through feature comparison between pairs of augmented images of the same remote sensing image.After self-supervised training,the encoder is finetuned in a downstream retrieval task based on a deep supervised hashing model with a small amount of labeled data,which better preserves the semantic information of the data and further improves the retrieval accuracy of remote sensing images.Experimental results show that the proposed DSCH method outperforms existing self-supervised learning methods in retrieving low-labeled remote sensing images.(2)This thesis proposed a deep unsupervised multi-similarity hashing(DMSH)method for unlabeled remote sensing image retrieval.The method employs an adaptive pseudo-label module based on k-nearest neighbors and kernel similarity to evaluate the similarity between images,and to generate and correct pseudo-labels online to improve their reliability.To measure the importance of different image pairs,a pairwise structural information module is designed to map the multi-scale structural similarity of image pairs to training attention,and allocate different training weights to optimize deep hashing learning.The network is trained by alternating optimization on two modules based on different similarities,fully exploring multiple similarity information between images to generate high-precision hash codes,and improving the accuracy of remote sensing image retrieval.Experimental results show that the proposed DMSH method achieves better retrieval performance than existing traditional unsupervised hashing and deep unsupervised hashing methods for unlabeled remote sensing image retrieval.