Feature Fusion Strategy Research Articles

Coal Flow Foreign Body Classification Based on ESCBAM and Multi-Channel Feature Fusion.

Foreign bodies often cause belt scratching and tearing, coal stacking, and plugging during the transportation of coal via belt conveyors. To overcome the problems of large parameters, heavy computational complexity, low classification accuracy, and poor processing speed in current classification networks, a novel network based on ESCBAM and multichannel feature fusion is proposed in this paper. Firstly, to improve the utilization rate of features and the network's ability to learn detailed information, a multi-channel feature fusion strategy was designed to fully integrate the independent feature information between each channel. Then, to reduce the computational amount while maintaining excellent feature extraction capability, an information fusion network was constructed, which adopted the depthwise separable convolution and improved residual network structure as the basic feature extraction unit. Finally, to enhance the understanding ability of image context and improve the feature performance of the network, a novel ESCBAM attention mechanism with strong generalization and portability was constructed by integrating space and channel features. The experimental results demonstrate that the proposed method has the advantages of fewer parameters, low computational complexity, high accuracy, and fast processing speed, which can effectively classify foreign bodies on the belt conveyor.

A pan-sharpening network using multi-resolution transformer and two-stage feature fusion.

Pan-sharpening is a fundamental and crucial task in the remote sensing image processing field, which generates a high-resolution multi-spectral image by fusing a low-resolution multi-spectral image and a high-resolution panchromatic image. Recently, deep learning techniques have shown competitive results in pan-sharpening. However, diverse features in the multi-spectral and panchromatic images are not fully extracted and exploited in existing deep learning methods, which leads to information loss in the pan-sharpening process. To solve this problem, a novel pan-sharpening method based on multi-resolution transformer and two-stage feature fusion is proposed in this article. Specifically, a transformer-based multi-resolution feature extractor is designed to extract diverse image features. Then, to fully exploit features with different content and characteristics, a two-stage feature fusion strategy is adopted. In the first stage, a multi-resolution fusion module is proposed to fuse multi-spectral and panchromatic features at each scale. In the second stage, a shallow-deep fusion module is proposed to fuse shallow and deep features for detail generation. Experiments over QuickBird and WorldView-3 datasets demonstrate that the proposed method outperforms current state-of-the-art approaches visually and quantitatively with fewer parameters. Moreover, the ablation study and feature map analysis also prove the effectiveness of the transformer-based multi-resolution feature extractor and the two-stage fusion scheme.

HAFFseg: RGB-Thermal semantic segmentation network with hybrid adaptive feature fusion strategy

RGB-Thermal (RGB-T) semantic segmentation provides the pixel-level prediction of surrounding environments for autonomous vehicles and mobile robots in harsh conditions such as insufficient illumination and severe weather. Existing RGB-Thermal semantic segmentation networks with two modalities feature extraction branches and insufficient feature fusion strategy limit their segmentation performance. To this end, this study proposed a novel RGB-Thermal semantic segmentation network with hybrid adaptive feature fusion strategy (HAFFseg). The proposed HAFFseg contains three feature extraction branches to achieve outstanding feature extraction capacity. The frequency domain feature enhanced module is invented and inserted in the first feature extraction stage to enhance the low-level extracted features. A hybrid adaptive feature fusion strategy which consists of hybrid feature fusion architecture and feature enhanced adaptive fusion modules (FEAFM) to obtain excellent feature fusion performance is designed for the proposed network. The full-scale fusion connection decoder is developed to perform multi-stage feature integration and further boost segmentation accuracy. Extensive ablation studies and comparison experiments have been conducted on MF dataset and RoadScene-seg dataset. Experimental results proved the effectiveness of the proposed HAFFseg structure and demonstrated the superiority of HAFFseg over other state-of-the-art RGB-T semantic segmentation networks.

Research on Multi-Port Ship Traffic Prediction Method Based on Spatiotemporal Graph Neural Networks

The intelligent maritime transportation system has emerged as a pivotal component in port management, owing to the rapid advancements in artificial intelligence and big data technology. Its essence lies in the application of digital modeling techniques, which leverage extensive ship data to facilitate efficient operations. In this regard, effective modeling and accurate prediction of the fluctuation patterns of ship traffic in multiple port regions will provide data support for trade analysis, port construction planning, and traffic safety management. In order to better express the potential interdependencies between ports, inspired by graph neural networks, this paper proposes a data-driven approach to construct a multi-port network and designs a spatiotemporal graph neural network model. The model incorporates graph attention networks and a dilated causal convolutional architecture to capture the temporal and spatial dimensions of traffic variation patterns. It also employs a gated-mechanism-based spatiotemporal bi-dimensional feature fusion strategy to handle the potential unequal relationships between the two dimensions of features. Compared to existing methods for port traffic prediction, this model fully considers the network characteristics of the overall port and fills the research gap in multi-port scenarios. In the experiments, real port ship traffic datasets were constructed using data from the Automatic Identification System (AIS) and port geographical information data for model validation. The results demonstrate that the model exhibits outstanding robustness and performs well in predicting traffic in multiple sub-regional port clusters.

Raman spectral pattern recognition of breast cancer: A machine learning strategy based on feature fusion and adaptive hyperparameter optimization

Raman spectroscopy, as a kind of molecular vibration spectroscopy, provides abundant information for measuring components and molecular structure in the early detection and diagnosis of breast cancer. Currently, portable Raman spectrometers have simplified and made equipment application more affordable, albeit at the cost of sacrificing the signal-to-noise ratio (SNR). Consequently, this necessitates a higher recognition rate from pattern recognition algorithms. Our study employs a feature fusion strategy to reduce the dimensionality of high-dimensional Raman spectra and enhance the discriminative information between normal tissues and tumors. In the conducted random experiment, the classifier achieved a performance of over 96% for all three average metrics: accuracy, sensitivity, and specificity. Additionally, we propose a multi-parameter serial encoding evolutionary algorithm (MSEA) and integrate it into the Adaptive Local Hyperplane K-nearest Neighbor classification algorithm (ALHK) for adaptive hyperparameter optimization. The implementation of serial encoding tackles the predicament of parallel optimization in multi-hyperparameter vector problems. To bolster the convergence of the optimization algorithm towards a global optimal solution, an exponential viability function is devised for nonlinear processing. Moreover, an improved elitist strategy is employed for individual selection, effectively eliminating the influence of probability factors on the robustness of the optimization algorithm. This study further optimizes the hyperparameter space through sensitivity analysis of hyperparameters and cross-validation experiments, leading to superior performance compared to the ALHK algorithm with manual hyperparameter configuration.

Semantic segmentation-assisted instance feature fusion for multi-level 3D part instance segmentation

Recognizing 3D part instances from a 3D point cloud is crucial for 3D structure and scene understanding. Several learning-based approaches use semantic segmentation and instance center prediction as training tasks and fail to further exploit the inherent relationship between shape semantics and part instances. In this paper, we present a new method for 3D part instance segmentation. Our method exploits semantic segmentation to fuse nonlocal instance features, such as center prediction, and further enhances the fusion scheme in a multi- and cross-level way. We also propose a semantic region center prediction task to train and leverage the prediction results to improve the clustering of instance points. Our method outperforms existing methods with a large-margin improvement in the PartNet benchmark. We also demonstrate that our feature fusion scheme can be applied to other existing methods to improve their performance in indoor scene instance segmentation tasks.

A Feature Embedding Network with Multiscale Attention for Hyperspectral Image Classification

In recent years, convolutional neural networks (CNNs) have been widely used in the field of hyperspectral image (HSI) classification and achieved good classification results due to their excellent spectral–spatial feature extraction ability. However, most methods use the deep semantic features at the end of the network for classification, ignoring the spatial details contained in the shallow features. To solve the above problems, this article proposes a hyperspectral image classification method based on a Feature Embedding Network with Multiscale Attention (MAFEN). Firstly, a Multiscale Attention Module (MAM) is designed, which is able to not only learn multiscale information about features at different depths, but also extract effective information from them. Secondly, the deep semantic features can be embedded into the low-level features through the top-down channel, so that the features at all levels have rich semantic information. Finally, an Adaptive Spatial Feature Fusion (ASFF) strategy is introduced to adaptively fuse features from different levels. The experimental results show that the classification accuracies of MAFEN on four HSI datasets are better than those of the compared methods.

Automatic temporal clothing deformation prediction: applicable to variable shape and motion

Simulating temporal three-dimensional (3D) deformations of clothing worn by the human body is a core technology in computer graphics that plays a vital role in various fields, such as computer games, animation, and movies. Physics-based simulation and data-driven methods are two mainstream technologies used to generate clothing deformation. However, when it is necessary to quickly generate clothing animations of different body shapes and motions, the existing methods cannot balance efficiency and effectiveness. In this paper, we present a learning-based method, given human body shape and motion, which automatically synthesizes temporal 3D deformations of clothing in real-time. A temporal framework based on the transformer is designed to capture the correlation between clothing deformation and shape features of the moving human body, as well as the frame-level dependency. A feature fusion strategy is an innovation used to fuse the two features of shape and motion. We also perform post-processing on the penetration of clothing and the human body, generating collision-free cloth deformation sequences. To evaluate the method, we build a human motion dataset based on the large-scale public human body dataset AMASS, and further develop a clothing deformation dataset. We qualitatively and quantitatively demonstrate that our approach outperforms existing methods in terms of temporal clothing deformation with variable shape and motion, as well as producing realistic deformation at interactive rates.

A surrogate evolutionary neural architecture search algorithm for graph neural networks

Due to the unique construction module and design of graph neural networks (GNNs), neural architecture search (NAS) methods specifically for GNNs have become a promising research hotspot in recent years. Among the existing methods, one class of methods microscopically searches for the constituent components of network layers. However, most of them ignore the topology connections between network layers or the feature fusion strategies. Another class of methods, called differentiable architecture search methods, has the advantage of searching topology connections and feature fusion strategies. However, constrained by the requirement of predefining all candidate operations, these methods can only sample a limited number of network layers. In this paper, we propose a surrogate evolutionary graph neural architecture search (GNAS) algorithm whose search space contains not only the microscopic network layer components but also topology connections and feature fusion strategies (called CTFGNAS). The GNN sampled in CTFGNAS is represented by a simple one-dimensional vector and does not fix the network depth. To address the problem that traditional crossover and mutation operators applied to GNAS may produce illegal solutions, we design a repair operation to guarantee the legitimacy of the solutions. The network depth is also increased with a large probability in the mutation operation to alleviate the oversmoothing problem. In addition, to cope with the challenge of computational resources due to the increased search space, we form a surrogate model with three classical regression models, where only a small number of solutions are truly evaluated for their fitness, and the remaining large number of solutions are predicted for their fitness by the surrogate model. Finally, experiments are executed on six widely used real-world datasets. The experimental results illustrate that CTFGNAS obtains more effective results than the state-of-the-art handcrafted GNNs and GNAS methods on all datasets. CTFGNAS is now available on the following website: https://github.com/chnyliu/CTFGNAS.

Depth and Video Segmentation Based Visual Attention for Embodied Question Answering.

Embodied Question Answering (EQA) is a newly defined research area where an agent is required to answer the users questions by exploring the real-world environment. It has attracted increasing research interests due to its broad applications in personal assistants and in-home robots. Most of the existing methods perform poorly in terms of answering and navigation accuracy due to the absence of fine-level semantic information, stability to the ambiguity, and 3D spatial information of the virtual environment. To tackle these problems, we propose a depth and segmentation based visual attention mechanism for Embodied Question Answering. Firstly, we extract local semantic features by introducing a novel high-speed video segmentation framework. Then guided by the extracted semantic features, a depth and segmentation based visual attention mechanism is proposed for the Visual Question Answering (VQA) sub-task. Further, a feature fusion strategy is designed to guide the navigators training process without much additional computational cost. The ablation experiments show that our method effectively boosts the performance of the VQA module and navigation module, leading to 4.9% and 5.6% overall improvement in EQA accuracy on House3D and Matterport3D datasets respectively.

Adaptive decoder-block selection with filter reweighting for medical image segmentation

Image segmentation models are mostly designed for specific tasks and trained on large-scale datasets. Medical image segmentation tasks, however, are highly diverse and usually have a small training dataset, making it difficult for a segmentation model to generalize well on other tasks. This architecture bias motivates a lot of research on designing a task-adaptive segmentation network with the most suitable settings including the network’s depth and width and feature fusion strategy. In this paper, we propose the adaptive Decoder-block Selection with Filter Reweighting (DeSFiR) algorithm to generate an efficient and effective medical image segmentation model for each specific dataset. This algorithm is a two-step iterative process. In the adaptive decoder-block selection step, we develop an auxiliary indicator function layer to bypass the decoder block whose individual responses are lower than a pre-defined threshold, aiming to generate a lightweight model. In the filter reweighing step, we replace invalid layers in the segmentation network with the external information from a simultaneously trained twin network. To this end, we utilize an entropy-based criterion to measure the informativeness of each layer and an adaptive weighting strategy to balance the replaced information between two networks. We have evaluated our algorithm on three tasks, including the segmentation of neural structures, nuclei, and liver. Our results indicate that the medical image segmentation model generated by the proposed DeSFiR algorithm has a significantly reduced complexity and comparable accuracy.

A Lightweight Radar Ship Detection Framework with Hybrid Attentions

One of the current research areas in the synthetic aperture radar (SAR) processing fields is deep learning-based ship detection in SAR imagery. Recently, ship detection in SAR images has achieved continuous breakthroughs in detection precision. However, determining how to strike a better balance between the precision and complexity of the algorithm is very meaningful for real-time object detection in real SAR application scenarios, and has attracted extensive attention from scholars. In this paper, a lightweight object detection framework for radar ship detection named multiple hybrid attentions ship detector (MHASD) with multiple hybrid attention mechanisms is proposed. It aims to reduce the complexity without loss of detection precision. First, considering that the ship features in SAR images are not inconspicuous compared with other images, a hybrid attention residual module (HARM) is developed in the deep-level layer to obtain features rapidly and effectively via the local channel attention and the parallel self-attentions. Meanwhile, it is also capable of ensuring high detection precision of the model. Second, an attention-based feature fusion scheme (AFFS) is proposed in the model neck to further heighten the features of the object. Meanwhile, AFFS constructs and develops a fresh hybrid attention feature fusion module (HAFFM) upon the local channel and spatial attentions to guarantee the applicability of the detection model. The Large-Scale SAR Ship Detection Dataset-v1.0 (LS-SSDD-v1.0) experimental results demonstrate that MHASD can balance detection speed and precision (improving average precision by 1.2% and achieving 13.7 GFLOPS). More importantly, extensive experiments on the SAR Ship Detection Dataset (SSDD) demonstrate that the proposed method is less affected by the background such as ports and rocks.

Chinese–Vietnamese Pseudo-Parallel Sentences Extraction Based on Image Information Fusion

Parallel sentences play a crucial role in various NLP tasks, particularly for cross-lingual tasks such as machine translation. However, due to the time-consuming and laborious nature of manual construction, many low-resource languages still suffer from a lack of large-scale parallel data. The objective of pseudo-parallel sentence extraction is to automatically identify sentence pairs in different languages that convey similar meanings. Earlier methods heavily relied on parallel data, which is unsuitable for low-resource scenarios. The current mainstream research direction is to use transfer learning or unsupervised learning based on cross-lingual word embeddings and multilingual pre-trained models; however, these methods are ineffective for languages with substantial differences. To address this issue, we propose a sentence extraction method that leverages image information fusion to extract Chinese–Vietnamese pseudo-parallel sentences from collections of bilingual texts. Our method first employs an adaptive image and text feature fusion strategy to efficiently extract the bilingual parallel sentence pair, and then, a multimodal fusion method is presented to balance the information between the image and text modalities. The experiments on multiple benchmarks show that our method achieves promising results compared to a competitive baseline by infusing additional external image information.

Local to non-local: Multi-scale progressive attention network for image restoration

Image restoration (IR) tasks aim to form a balance between complex textures and spatial details. To this end, the combination of local and non-local attention mechanisms has been well studied in recent years. However, existing local attention-based modules ignore the interaction between channel and spatial attention, while non-local attention operations solely focus on short-range or long-range dependency. To overcome these problems, a novel multi-scale progressive attention network is proposed in this paper, which is termed as MPANet. The proposed MPANet is composed of two parts, a local multi-scale feature extractor and a window-dilation self-attention module. In the local multi-scale feature extractor, a single-scale feature enhancement strategy is designed to model the correlation between channel and spatial dimensions, and a multi-scale feature fusion strategy is applied to further exchange contextual information across all the scales. Furthermore, the window-dilation self-attention is introduced to establish global representation while preserving local details. Experimental results on four IR tasks demonstrate that the proposed MPANet outperforms the state-of-the-art methods in both quantitative results and visual perception.

Multi‐scale feature fusion pyramid attention network for single image dehazing

AbstractTexture and color distortion are common in existing learning‐based dehazing algorithms, and it is argued that one of the major reasons is that the shallow features of fog images are underutilized, and the deep features of fog images are insufficient for single image dehazing. In order to provide more texture and color information for image restoration, more shallow features need to be added in the process of image decoding. Therefore, a multi‐scale feature fusion pyramid attention network (PAN) for single image dehazing is proposed. In PAN, combined with the attention mechanism, a shallow and deep feature fusion (SDF) strategy is designed. SDF considers multi‐scale as well as channel‐level fusion to provide feature information under different receptive fields while also highlighting important channels, such as texture and color information. DC is designed as a latent space mapping module to learn a mapping relationship between the latent space representation of the hazy image at low resolution and the corresponding latent space representation of the haze‐free image. Additionally, network deconvolution (ND) and deformed convolution network (DCN) are introduced into PAN. The ND module can remove pixel‐wise and channel‐wise correlation of features, reduce data redundancy to obtain sparse representation of features, and speed up network convergence. The DCN module can use its adaptive receptive field to focus on the area of interest for calculation and play a role in texture feature enhancement. Finally, the perceptual loss is chosen as the regularization item of the loss function, which makes style features of the restored image closer to the real fog‐free image. Extensive experiments reveal that the proposed PAN outperforms other existing dehazing methods on real‐world and synthetic datasets.

Salient object detection for RGBD video via spatial interaction and depth-based boundary refinement

Recently proposed state-of-the-art saliency detection models rely heavily on labeled datasets and rarely focus on perfect RGBD feature fusion, which lowers their generalization ability. In this paper, we propose a depth-based interaction and refinement network (DIR-Net) to fully leverage the depth information provided with RGB images to generate and refine the corresponding saliency segmentation maps. In total, three modules are included in our framework. A depth-based refinement module (DRM) and an RGB module work in parallel while coordinating via interactive spatial guidance modules (ISGMs), which utilize spatial and channel attention computed from both depth features and RGB features. In each layer, the features in both modules are refined and guided by the spatial information obtained from the other module through ISGMs. In the RGB module, before sending the depth-guided feature map to the decoder, a convolutional gated recurrent unit (ConvGRU)-based block is introduced to handle temporal information. Thinking about the clear movement information in RGB features, the block also guides temporal information in DRM. By merging the results from both the DRM and RGB modules, a segmentation map with distinct boundaries is generated. Considering the lack of depth images in popular public datasets, we utilize a depth estimation network that incorporates manual postprocessing-based correction to generate depth images on the DAVIS and UVSD datasets. The state-of-the-art performance achieved on both the original and new datasets illustrates the advantage of our RGBD feature fusion strategy, with a real-time speed of 19 fps on a single GPU.

A Deep Feature Fusion Method for Complex Ground Object Classification in the Land Cover Ecosystem Using ZY1-02D and Sentinel-1A

Despite the successful application of multimodal deep learning (MDL) methods for land use/land cover (LULC) classification tasks, their fusion capacity has not yet been substantially examined for hyperspectral and synthetic aperture radar (SAR) data. Hyperspectral and SAR data have recently been widely used in land cover classification. However, the speckle noise of SAR and the heterogeneity with the imaging mechanism of hyperspectral data have hindered the application of MDL methods for integrating hyperspectral and SAR data. Accordingly, we proposed a deep feature fusion method called Refine-EndNet that combines a dynamic filter network (DFN), an attention mechanism (AM), and an encoder–decoder framework (EndNet). The proposed method is specifically designed for hyperspectral and SAR data and adopts an intra-group and inter-group feature fusion strategy. In intra-group feature fusion, the spectral information of hyperspectral data is integrated by fully connected neural networks in the feature dimension. The fusion filter generation network (FFGN) suppresses the presence of speckle noise and the influence of heterogeneity between multimodal data. In inter-group feature fusion, the fusion weight generation network (FWGN) further optimizes complementary information and improves fusion capacity. Experimental results from ZY1-02D satellite hyperspectral data and Sentinel-1A dual-polarimetric SAR data illustrate that the proposed method outperforms the conventional feature-level image fusion (FLIF) and MDL methods, such as S2ENet, FusAtNet, and EndNets, both visually and numerically. We first attempt to investigate the potentials of ZY1-02D satellite hyperspectral data affected by thick clouds, combined with SAR data for complex ground object classification in the land cover ecosystem.

Unsupervised stepwise extraction of offshore aquaculture ponds using super-resolution hyperspectral images

Despite offshore aquaculture brings great economic benefits, it also has destructive effects on the ecological environment of coastal regions. Therefore, the accurate monitoring of offshore aquaculture areas is vital. Existing methods for extracting aquaculture ponds are still limited by the lack of high-resolution hyperspectral imagery and inadequate feature extraction. In this paper, we proposed an unsupervised aquaculture ponds extraction method based on hyperspectral imagery super-resolution, feature fusion and stepwise extraction strategy. First, the resolution of the original hyperspectral imagery is enhanced via a deep learning-based super-resolution method. Then we introduce a feature fusion method by combining multi-dimensional features including elevation, reflectance, biochemistry, principal component analysis, and fast-Fourier transform features to enhance the feature sensitivity to the aquaculture ponds. The aquaculture ponds are finally extracted via a step-wise extraction strategy. Experiments on ZH-1 and GF-7 imageries are conducted, and the experimental results show that our method achieves an OA of 97.9% on the aquaculture pond extraction and that the method can be generalized for the object extraction on multispectral datasets. Ablation experiments show that all three of our innovative modules can effectively improve the extraction accuracy and confirm the generalization of the method.

IS-Net: Automatic Ischemic Stroke Lesion Segmentation on CT Images

Ischemic stroke is an acute cerebral vascular disease and makes up about 80% of all stroke cases. Non-contrast computed tomography (NCCT) is a widely applied imaging technique for ischemic stroke assessment. However, it is challenging to identify ischemic lesion on NCCT images due to its high variability in location, contrast, and geometry. In this work, we propose IS-Net, an encoder-decoder convolutional neural network for automatic ischemic stroke lesion segmentation on NCCT images. The proposed IS-Net takes hierarchical network as backbone while pyramid feature aggregation (PFA) module is designed to aggregate features from multi stages of backbone, and reasonable feature fusion strategy is considered in PFA to enhance multi-level propagation. To fully mine the boundary cues, edge constraint scheme is introduced by deep supervision which broadcasts low-level features to different modules. In addition, to overcome the limitation of fixed geometric structure of convolution for multi-range dependency exploitation, non-local parallel decoder is introduced with deformable convolution and self-attention. The proposed IS-Net is evaluated on manually labeled follow-up NCCT dataset composed of 1,004 cases (totally 9,020 images). The proposed IS-Net is compared with the state-of-the-art segmentation models and illustrates the highest score on segmentation criteria and sensitivity.

DHUnet: Dual-branch hierarchical global–local fusion network for whole slide image segmentation

Hematoxylin and eosin stained whole slide images (WSIs) are the gold standard for pathologists and medical professionals for tumor diagnosis, surgery planning, and postoperative examinations. In recent years, due to the rapidly emerging field of deep learning, there have been many convolutional neural networks (CNNs) and Transformer based models applied to computational pathology for accurate segmentation. However, the generalization ability and robustness of models often affect the diagnosis and prognosis of cancer, and we attempt to effectively combine the advantage of CNN which excels in sparse WSI segmentation while Transformer excels in dense cases. In this paper, We propose a novel feature fusion strategy, DHUnet, which utilizes Swin Transformer and ConvNeXt modules with a dual-branch hierarchical U-shaped architecture to fuse global and local features for WSI segmentation. Firstly, a WSI is divided into small patches, which are put into the global and local encoders to generate hierarchical features in parallel. Then, with the help of global–local fusion modules and skip connections, the decoder can fully obtain the global coarse and local fine-grained information during the upsampling process. The proposed Cross-scale Expand Layer can make the patches with the same center but different scales recover the input resolution gradually at each stage. Finally, all the projected pixel-level patch masks are merged to restore the final WSI tumor segmentation. Extensive experiments demonstrate that DHUnet has excellent performance and generalization ability for WSI segmentation, achieving the best segmentation results on three datasets with different types of cancer, densities, and target sizes. The code and pre-processed datasets will be publicly available at https://github.com/pengsl-lab/DHUnet.