查看更多>>摘要：Change detection (CD) is a crucial task in various real-world applications, aiming to identify regions of change between two images captured at different times. However, existing approaches mainly focus on designing advanced network architectures that map feature differences to change maps, overlooking the impact of feature difference quality. In this paper, we approach CD from a different perspective by exploring how to optimize feature differences to effectively highlight changes and suppress background regions. To achieve this, we propose a novel module called the iterative difference-enhanced transformers (IDET). IDET consists of three transformers: two for extracting long-range information from the bi-temporal images, and one for enhancing the feature difference. Unlike previous transformers, the third transformer utilizes the outputs of the first two transformers to guide iterative and dynamic enhancement of the feature difference. To further enhance refinement, we introduce the multi-scale IDET-based change detection approach, which utilizes multi-scale representations of the images to refine the feature difference at multiple scales. Additionally, we propose a coarse-to-fine fusion strategy to combine all refinements. Our final CD method surpasses nine state-of-the-art methods on six large-scale datasets across different application scenarios. This highlights the significance of feature difference enhancement and demonstrates the effectiveness of IDET. Furthermore, we demonstrate that our IDET can be seamlessly integrated into other existing CD methods, resulting in a substantial improvement in detection accuracy.
查看更多>>摘要：Inspired by the great success of the Transformer in computer vision, some works have started to explore the use of the Transformer for super-resolution (SR). However, with regard to stereoscopic SR, which aims to recover details from input pairs, how to efficiently integrate cross-view interactions into the Transformer architecture is still an ongoing development. Additionally, most existing stereoscopic SR methods only adopt a parallax mechanism in the middle of the network, and another issue is that the feature correlation from different viewpoints inevitably weakens as the network depth increases. To address these issues, we first utilize an efficient residual transformer block (ERTB) as the backbone for long-range intra-view feature extraction. Subsequently, we propose a novel multi-Dconv cross attentive block (MCAB) to enhance the cross-view interactions at the rear part of the Transformer architecture. Notably, the proposed MCAB promotes feature fusion from two viewpoints by employing bidirectional cross-attention, as opposed to an unidirectional flow from left to right or vice versa. This approach results in an efficient cross-view interaction from both branches. By leveraging the advantages of the proposed ERTB and MCAB, we introduce an efficient cross-view interaction Transformer (CVIformer) for stereoscopic SR. This architecture is capable of incorporating long-range intra-view and cross-view information with an acceptable computational overhead. Without excessive complexity, extensive experiments conducted on four public datasets demonstrate that our model achieves state-of-the-art results using only 1.17 million parameters, with approximately a 40% reduction in parameters compared to leading methods like iPASSR.
查看更多>>摘要：Deep learning (DL) has successfully automated dose distribution prediction for radiotherapy planning, increasing both efficiency and quality. However, existing methods commonly utilize ${{\mathbf{L}}_1}$ or ${{\mathbf{L}}_2}$ loss to calculate the posterior average, thus heavily suffering from the over-smoothing problem. To address this, we propose a diffusion model-based method, named DiffDose, to automatically predict radiotherapy dose distribution for cancer patients. Specifically, our DiffDose model contains a forward process and a reverse process. In the forward process, DiffDose gradually adds small noise to dose distribution maps via multiple steps until converting them to pure Gaussian noise, and a noise predictor is simultaneously trained to estimate the noise added in each step. In the reverse process, DiffDose iteratively removes the noise from a pure Gaussian noise leveraging the well-trained noise predictor and finally outputs the predicted dose distribution maps. Concretely, to provide the model with essential structure information, we design a structure encoder to learn the anatomical knowledge from patients’ anatomy images, guiding the noise predictor to generate dose distribution maps that are aware of personalized structures. Considering the latent continuity and similarity among sliced anatomy images, an inter-slice interaction transformer (I2T) block is embedded in the structure encoder to capture such long-range dependency. Extensive experiments on an in-house dataset involving 130 rectum cancer cases validate the superiority of our method.
查看更多>>摘要：In knowledge graph construction, a challenging issue is how to extract complex (e.g., overlapping) entities and relationships from a small amount of unstructured historical data. The traditional pipeline methods are to divide the extraction into two separate subtasks, which misses the potential interactio between the two subtasks and may lead to error propagation. In this work, we propose an effective cascade dual-decoder method to extract overlapping relational triples, which includes a text-specific relation decoder and a relation-corresponded entity decoder. Our approach is straightforward and it includes a text-specific relation decoder and a relation-corresponded entity decoder. The text-specific relation decoder detects relations from a sentence at the text level. That is, it does this according to the semantic information of the whole sentence. For each extracted relation, which is with trainable embedding, the relation-corresponded entity decoder detects the corresponding head and tail entities using a span-based tagging scheme. In this way, the overlapping triple problem can be tackled naturally. We conducted experiments on a real-world open-pit mine dataset and two public datasets to verify the method's generalizability. The experimental results demonstrate the effectiveness and competitiveness of our proposed method and achieve better F1 scores under strict evaluation metrics.
查看更多>>摘要：Although model based $H_{\infty }$ control scheme for nonlinear continuous-time (CT) systems with unknown system dynamics has been extensively studied, model-free $H_{\infty }$ control of nonlinear CT systems via Q-learning is still a challenging problem. This paper develops a novel Q-learning based model-free $H_{\infty }$ control scheme for nonlinear CT systems, where the adaptive critic and actor continuously and simultaneously update each other, eliminating the need for iterative steps. As a result, a hybrid structure is avoided and there is no longer a requirement for an initial stabilizing control policy. To obtain the $H_{\infty }$ control of the CT nonlinear system, the Q-learning strategy is introduced to online resolve the $H_{\infty }$ control problem in a non-iterative approach, where the system dynamics are not required. In addition, a new learning law is further developed by utilizing a sliding mode scheme to online update the critic neural network (NN) weights. Due to the strong convergence of critic NN weights, the actor NN used in most $H_{\infty }$ control algorithms is removed. Finally, numerical simulation and experimental results of an adaptive cruise control (ACC) system based on a real vehicle effectively demonstrate the feasibility of the presented control method and learning algorithm.
查看更多>>摘要：Supervised classification problems from the real world typically face a challenge characterized by the scarcity of samples in one or more target classes compared to the rest of the majority classes. In response to such class imbalance, we propose an oversampling technique based on clustering, aiming to populate the minority class with synthetic samples. This approach capitalizes on the notion of “Cluster Cores,” representing locally dense regions within clusters. These Cluster Cores act as central, densely crowded areas that capture intricate topological properties of the corresponding clusters, especially in complex datasets with a non-convex spatial orientation in the feature space. By concentrating on these high-density regions, our clustering-based oversampling technique generates synthetic samples within the convex hull region of minority class instances in the formed clusters. This strategy ensures the creation of points that align with the data space and considers each minority instance within a specific cluster, thereby averting the problems encountered due to the generation of artificial samples by mere linear combination of the minority class data points, as is encountered in SMOTE (Synthetic Minority Oversampling Technique)-based algorithms. To assess the efficacy of our proposal, we conducted experimental comparisons against several cutting-edge algorithms, considering an array of evaluation metrics on well-known datasets used in the literature for both binary and multi-class classification. Additionally, we undertook a detailed ablation study, scrutinized existing algorithms in our context, delineated their strengths and limitations, and contemplated potential research directions in this domain.
NETL
NSTL
IEEE