英文摘要 | As an important research field in remote sensing technology, change detection can help analyze and determine changes in surface objects, and is widely used in land and resource surveys, agricultural production monitoring, urban expansion analysis, and natural disaster assessment. Due to the influence of illumination changes, changes in imaging conditions, and non-physical changes, it has always been a difficult research problem to develop high-resolution remote sensing image change detection methods with high precision and multi-scene generalization. At present, the rapid development of artificial intelligence technology represented by deep learning has set off a new wave in the field of remote sensing image processing. Deep learning technology can efficiently and adaptively extract hierarchical representative features from input data, which is very suitable for processing high-resolution remote sensing images. However, deep learning-based methods in the field of change detection still face difficulties and challenges, including low accuracy, poor generalization, and limited applicability to specific scenes. This thesis takes this opportunity, based on full investigation and analysis of existing methods, summarizes shortcomings and experiences, and conducts in-depth research from multiple perspectives around the application of depth learning technology in multi-class change detection and single-class change detection in high-resolution remote sensing images. The main research contributions are as follows:
(1) Aiming at the change detection problem of multi-category remote sensing images, a processing framework of multi-task learning is proposed. Multi-task learning is reflected in the fact that two tasks of semantic segmentation and change detection can be completed at the same time. Among them, the change detection branch identifies the changed area, thus limiting the scope of semantic segmentation, so that the network's identification of the change type is only for the changed area. In addition, considering that in the multi-category change detection task, there is often a certain category that dominates, and the resulting category imbalance will cause the network to tend to multi-category while ignoring few categories. The proposed semantic constraint function for the non-changing area can make the network focus on the semantic comparison of the changing area, constrain the semantics of the non-changing area to be consistent as much as possible, and improve the feature extraction ability of the network and the ability to distinguish changes. Experiments show that the proposed method outperforms other multi-class change detection methods in detection accuracy.
(2) For the change detection problem of a single category, this paper mainly focuses on the change detection of buildings, and proposes a change detection model based on double cross attention. First, starting from the essence of change, we propose cross-attention. The cross-attention mechanism is inspired by changes in human visual observation, and detects changes through alternate comparison from left to right and continuous matching from front to back. Second, in order to fully mine the high-frequency and low-frequency information in images, a mixer combining high-frequency and low-frequency information is proposed to replace the traditional self-attention mixer. The high-frequency and low-frequency branches are divided by the frequency ramp structure, and the convolution-based and Transformer-based cross-attention hybrids are coupled, enabling the model to expand its perceptual capabilities and capture more rich features from the input paired data. In addition, the pure twin structure is abandoned in the feature extraction stage, and the interaction of information is achieved through cross-attention connections, and the protection and restoration of detailed information is achieved through hierarchical feature extraction and fusion. A large number of experiments show that the method proposed in this paper has achieved a balance between change detection accuracy and efficiency on the four datasets, and its comprehensive performance exceeds similar methods, and has the potential to be a general backbone for change detection. |
修改评论