Fault diagnosis under unseen working conditions remains a significant challenge in intelligent machinery health monitoring. Current domain generalization methods tend to overfit to domain-specific factors, limiting their robustness in unseen environments. This entanglement disrupts the class-level structure in the feature space, thereby diminishing the discriminative power of learned representations under unseen conditions. To address this limitation, we propose a novel domain generalization method for fault diagnosis, termed collaborative bi-stream hierarchical contrastive learning. First, a bi-stream factorization module separates features into task-invariant and domain-specific components. The task-invariant stream is regularized using an adversarial entropy objective to suppress domain-related cues and a reconstruction loss to preserve task-relevant information. Second, a hierarchical contrastive learning strategy enhances the discriminability of task-invariant features by jointly optimizing global prototype compactness, instance-to-prototype alignment, and domain-modulated instance-level contrast. Experimental results demonstrate that the proposed method achieves an average diagnostic accuracy of 94.4% across the tested datasets, representing an improvement of approximately 8.3% compared to the best-performing baseline, demonstrating strong generalization capability in industrial scenarios.

Reference-based super-resolution (RefSR) technology holds significant potential for enhancing the spatial resolution of remote sensing images. However, existing methods still face notable limitations in the effectiveness of large-scale texture transfer and struggle to guarantee the fidelity of re-constructed content when local variations exist between low-resolution (LR) images and reference (Ref) images. To address these problems, we propose a novel Mamba and U-Net-based method for RefSR, named Mamba-UNet. First, the LR image is spectrally decomposed into high-frequency (HF) and low-frequency (LF) components. The HF component is then enhanced by a Mamba module with long-sequence modeling capability, which can adaptively transfer textures from any location in the Ref image, thereby effectively resolving the blurring problem in large-scale texture transfer. Second, an adaptive multi-source data fusion module is designed to dynamically evaluate the land-cover consistency between the LR and Ref images through a channel attention gating mechanism. Once a local mismatch is detected, this mechanism automatically reduces the weight of Ref features and strengthens the dominance of the LR image’s own low-frequency content, thus ensuring the fidelity of content reconstruction. Finally, the gated fused features are fed into an attention-augmented U-Net for high-quality image reconstruction. Experimental results demonstrate that compared to state-of-the-art RefSR methods, the proposed method exhibits superior reconstruction performance and stronger robustness in quantitative metrics and visual quality assessments on the SECOND and CNAM-CD datasets. The code is available at https://github.com/wzh226/AM-UNET.

Long-term time series forecasting is widely applied in fields such as transportation, energy consumption, and disease control, yet existing Transformer-based methods often struggle to capture complex temporal dependencies and accurately model long-range patterns through a single branch. To address these issues, we propose a Dual-branch Aggregation Method (DAformer) for long-term time series forecasting, designed to enhance prediction accuracy through a modular framework. DAformer consists of three main components: a temporal feature extraction module, a correlation-based feature decomposition module, and a dual-branch aggregation module. The temporal feature extraction module utilizes a self-attention mechanism to capture both short-term and long-term dependencies between time points, regardless of their temporal distance, enabling effective modeling of complex temporal patterns. The correlation-based feature decomposition module decomposes time series into distinct seasonal and trend components and further analyzes key temporal features by evaluating interactions across various time intervals. Finally, the dual-branch aggregation module integrates the features extracted by the two branches, capturing feature similarities and interdependencies across multiple time series. Our method outperforms state-of-the-art (SOTA) methods in terms of predictive ability. It achieves the best forecasting performance across eight datasets. For prediction lengths of 96, 192, 336, and 720, the average MSE values are 0.258, 0.310, 0.375, and 0.499, respectively, with corresponding MAEs of 0.295, 0.334, 0.378, and 0.449.

Deep neural networks have achieved strong performance in 3D point cloud understanding, yet their decision process remains difficult to interpret. Existing explainability methods are mainly designed for images and are not well suited to irregular, sparse point clouds. In point cloud networks, Farthest Point Sampling (FPS) breaks the correspondence between deep features and original inputs, leading to spatial inconsistency. Some existing methods alter the original point cloud by removing points during explanation, which alters the input itself and undermines input consistency. We propose Hierarchical Progressive Fusion (HPF), which progressively fuses multi-level gradients and activations and projects them back to the original input space to achieve spatial consistency. Furthermore, HPF only requires a single forward and backward propagation to generate heatmaps, achieving input consistency. Moreover, HPF does not modify the point cloud network architecture as a plug-and-play tool. Experiments on ShapeNet-Part and SUN RGB-D datasets show that HPF yields faithful and effective explanations for point cloud models. Code is available at https://github.com/perfect979/HPF.

The existing cervical whole slide image classification methods ignore the influence of image quality, resulting in low classification accuracy. To address this, we propose a dual-path multiple instance learning classification method guided by image quality assessment. Specifically, a pre-trained quality assessment model assigns quality scores to patches, splitting them into high- and low-quality paths. In the high-quality path, patch features are weighted by their quality scores to emphasize reliable diagnostic regions. In the low-quality path, a key instance is selected using clustering and feature distance matching. Finally, a cross-attention module fuses features across quality levels. Our method achieves 94.64% accuracy and 91.74% AUC on a dataset of 2,434 WSIs collected from five medical centers, outperforming state-of-the-art methods.

Instance segmentation plays an important role in the automatic diagnosis of cervical cancer. Although deep learning-based instance segmentation methods can achieve outstanding performance, they need large amounts of labeled data. This results in a huge consumption of manpower and material resources. To solve this problem, we propose an unsupervised cervical cell instance segmentation method based on human visual simulation, named HVS-Unsup. Our method simulates the process of human cell recognition and incorporates prior knowledge of cervical cells. Specifically, firstly, we utilize prior knowledge to generate three types of pseudo labels for cervical cells. In this way, the unsupervised instance segmentation is transformed to a supervised task. Secondly, we design a Nucleus Enhanced Module (NEM) and a Mask-Assisted Segmentation module (MAS) to address problems of cell overlapping, adhesion, and even scenarios involving visually indistinguishable cases. NEM can accurately locate the nuclei by the nuclei attention feature maps generated by point-level pseudo labels, and MAS can reduce the interference from impurities by updating the weight of the shallow network through the dice loss. Next, we propose a Category-Wise droploss (CW-droploss) to reduce cell omissions in lower-contrast images. Finally, we employ an iterative self-training strategy to rectify mislabeled instances. Experimental results on our dataset MS-cellSeg, the public datasets Cx22 and ISBI2015 demonstrate that HVS-Unsup outperforms existing mainstream unsupervised cervical cell segmentation methods.

Automatic abnormal cervical cell detection provides cytopathologists with reliable evidence, rapiddiagnosis and reduced workload. However, this task is often prone to interference factors such as impurities,microorganisms, and neutrophils. The morphological diversity of abnormal cells also increasesthe complexity of detection. Unfortunately, existing abnormal cervical cell detection methods havenot fully utilized diagnostic knowledge from cytopathologists to handle these challenges.We proposean abnormal cervical cell detection method that leverages image and textual diagnostic knowledgein The Bethesda System (TBS) to interpret cervical cytology results, boosting the performance ofthe foundation model for target detection. First, we build a TBS diagnostic knowledge base througha TDK-Bert module. In TDK-Bert, TBS diagnostic knowledge can be efficiently extracted throughBert to capture textual features, which can then be aligned and fused with image features through anattention mechanism. In addition, we propose an ODC-SPPF module to improve the Spatial PyramidPooling-Fast (SPPF) module in a YOLO-based foundation model. The ODC-SPPF employs omnidimensionaldynamic convolution to improve the ability of feature extraction, particularly for subtlecells, and alleviates the interference of complex backgrounds in Thinprep cytologic test images. Experimentalresults on a cervical cell detection dataset demonstrate that our method effectively combinesknowledge bases and foundation models for abnormal cervical cell detection and outperforms thestate-of-the-art methods.

Cervical cell segmentation plays an important role in intelligent cervical cancer diagnosis. However, the performance of current methods is still limited because cervical cells are overlapping, adhesion, and even visual inseparability. In addition, deep learning-based instance segmentation methods need large amounts of labeled data, leading to significant consumption of manpower and material resources. To solve these problems, we propose a multi-branch instance segmentation network for cervical cells, named MBSeg. This method selects cervical cell images and automatically annotates them at the pixel level using active learning. Next, we leverage CondInst, a one-stage instance segmentation network, and improve it by adding a new over-lapping segmentation branch. This branch generates overlapping mask maps, which are integrated into the original mask branch to enhance attention on the receptive fields of the overlapping areas. Furthermore, we employ a Nucleus Enhanced Module (NEM) to accurately locate nuclei and a Mask-Assisted Segmentation (MAS) module to minimize interference from impurities. Experiments on our dataset MS-cellSeg, the public Cx22, and ISBI2015 datasets demonstrate the superiority of our method in segmenting overlapping cervical cells.

Automated cervical cancer screening through computer-assisted diagnosis has shown considerable potential to improve screening accessibility and reduce associated costs and errors. However, classification performance on whole slide images (WSIs) remains suboptimal due to patient-specific variations. To improve the precision of the screening, pathologists not only analyze the characteristics of suspected abnormal cells, but also compare them with normal cells. Motivated by this practice, we propose a novel cervical cell comparative learning method that leverages pathologist knowledge to learn the differences between normal and suspected abnormal cells within the same WSI. Our method employs two pre-trained YOLOX models to detect suspected abnormal and normal cells in a given WSI. A self-supervised model then extracts features for the detected cells. Subsequently, a tailored Transformer encoder fuses the cell features to obtain WSI instance embeddings. Finally, attention-based multi-instance learning is applied to achieve classification. The experimental results show an AUC of 0.9319 for our proposed method. Moreover, the method achieved professional pathologist-level performance, indicating its potential for clinical applications.

It is important to extract both global and local features for view-based 3D shape classification. Therefore, we propose a 3D shape classification method based on global and local features extraction with collaborative learning. This method consists of a patch-level transformer sub-network (PTS) and a view-level transformer sub-network (VTS). In the PTS, a single view is divided into multiple patches. And a multi-layer transformer encoder is employed to accurately highlight discriminative patches and capture correlations among patches in a view, which can efficiently filter out the meaningless information and enhance meaningful information. The PTS can aggregate patch features into a 3D shape representation with rich local details. In the VTS, a multi-layer transformer encoder is employed to assign different attention to each view and obtain the contextual relationship among views, which can highlight the discriminative views among all the views of the same 3D shape and efficiently aggregate view features into a 3D shape representation. A collaborative loss is applied to encourage the two branches to learn collaboratively and teach each other in training. Experiments on two 3D benchmark datasets show that our proposed method outperforms current methods.

目前的三维模型分类方法均是对初始姿态已经对齐的数据集进行分类,但是在实际应用中,三维模型的姿态是未知的,非对齐的三维模型将导致分类准确率急剧下降.本文提出了一种新的三维模型分类方法,适用于模型姿态对齐和非对齐两种情况.该方法采用图卷积神经网络(Graph Convolutional neural Network,GCN)学习视图间的空间关系,将预先设置好的相机位置作为图结构中的顶点,并通过时序特征提取网络以及注意力网络进一步提升GCN的运算效果,从而完成三维模型的分类.实验表明,该方法在ModelNet10和ModelNet40数据集上进行实验,在三维模型姿态对齐的情况下,分类准确率分别高达99.3％和97.4％,远高于现有方法.在三维模型姿态非对齐的情况下,也有较高的分类准确率.

3D model retrieval with sketches as input is convenient for users to express their retrieval needs, and has become a research hotspot. However, a sketch which is only the rough description of a 3D model from a single viewpoint has high degree of abstraction. In addition, there are large domain differences between sketches and the 3D models. These lead to low accuracy of the current sketch retrieval systems. To solve these problems, we propose a new sketch-based 3D model retrieval method. In this method, a 3D model is firstly projected to a set of 2D views with aligned viewpoints, thereby obtaining one-to-many sketch-views pairs. Secondly , the Canny edge detection algorithm is used to extract the multi-level pseudo-sketches from each view, which aims at reducing the domain differences between the views and the sketches. Thirdly, sample selection and data enhancement are adopted to alleviate unbalanced training data. Finally, the prepared data are used to train a Triplet Network which maps the pseudo-sketches and sketches to the common feature embedding space. The Triplet Network is finally used to extract features and build indexes for sketch retrieval. Experiments show that the retrieval accuracy of the proposed method on SHREC ’13 and SHREC ’14 can reach up to 70. 0% and 63. 6% , respectively.

Zero-shot learning (ZSL) strives to classify unseen categories for which no data is available during training.At present, among generative methods, zero-shot learning based on joint generative model VAEGAN is a research hotspot.On this basis, we propose a zero-shot image classification method based on Discriminator Feedback VAEGAN (DF-VAEGAN).This method introduces a feedback module in the discriminator part, which can improve the overall performance of the model in the training stage.In the feature generation stage, it can be combined with the generator to jointly improve the quality of feature generation.Finally, the classifier is trained through high quality synthetic features to improve classification accuracy.The method also reconstructs attribute features through the decoder and uses a cycle consistency loss to ensure semantic consistency of the generated feature.Experiments on ZSL and generalized zero-shot learning (GZSL) show that our method outperforms existing methods on five classical datasets, effectively enhancing the quality of feature synthesis and reducing the goal of between categories in the zero-shot image classification task.

Zero-shot learning (ZSL) aims to classify samples of unseen categories for which no training data is available. At present, the VAEGAN framework which combines Generative Adversarial Networks (GAN) with Variational Auto-Encoder (VAE) has achieved good performance in zero-shot image classification. Based on the VAEGAN, we propose a new zero-shot image classification method named Enhanced VAEGAN (E-VAEGAN). Firstly, we design a feature alignment module to align visual features and attribute features. Then, the aligned features are fused with the hidden layer features of the encoder to improve output features of the encoder. Secondly, the triplet loss is applied during the encoder training, which further increases the discriminability of features. Finally, the hidden layer features of the discriminator are input into a transform module and then fed back to the generator, which improves the quality of the generated fake samples. The originality of this paper is that we design a new E-VAEGAN which employs the feature alignment module, triplet loss and transform module to reduce the ambiguity between categories and make the generated fake features similar to the real features. Experiments show that our method outperforms the compared methods on five zero-shot learning benchmarks.

The integration of view-based 3D model classification and deep learning can effectively improve the classification accuracy. However, current methods consider that the views from different viewpoints of 3D model with same category belong to the same category and ignore the view differences, which makes it difficult for the classifier to learn a reasonable classification surface. To solve this problem, a 3D model classification method based on deep neural network is proposed. The multiple viewpoint groups are set evenly around the 3D model in this method, and the view classifier for each viewpoint group is trained for fully mining the deep information of the 3D model in different viewpoint groups. These classifiers share a feature extraction network, but have their own classification network. In order to extract the discriminative view features, the attention mechanism is added to the feature extraction network; In order to model the views of the non-viewpoint group, additional classes are added to the classification network. In the classification stage, a view selection strategy is first proposed, which can use a small number of views to classify the 3D model and improve classification efficiency. Then a classification strategy is proposed to achieve reliable 3D model classification through classification view. Experimental results on ModelNet10 and ModelNet40 show that the classification accuracy can reach up to 93.6% and 91.0% with only 3 views.

At present, view-based 3D model classification is a research hotspot. However, current methods produce many redundant views, and all views are treated equally, ignoring their differences and importance. To solve the above problems, we propose a multi-view fusion 3D model classification method. This method first extracts view features using the view feature extraction network with mixed domain attention, and then fuses these view features and inputs the fused features into the view weight learning network with channel attention, giving different weights to different views according to their importance to the 3D model, and forming representative feature descriptors for 3D model classification. Experimental results shows that the classification accuracy rates in the rigid 3D model data sets ModelNet10 and ModelNet40 reached 98.3% and 95.5%.

Nuclei segmentation is a key technique for automatic pathological screening. Although many methods have been proposed, it remains a challenge because of numerous nuclei clusters, high variability of object appearances and complex backgrounds. To address these issues, we propose a novel multi-task region-enhanced nuclei segmentation network (REU-Net). It stacks three U-shaped structures by combining serial and parallel approaches to construct a multi-task architecture. The model employs two auxiliary tasks, i.e., contour extraction and rough segmentation to help the main task of fine segmentation. The saliency regions are enhanced by the prediction results of the auxiliary tasks, and the enhanced images are further segmented through the main task. In addition, the spatial and texture features in auxiliary tasks are aggregated by attention gates, helping the main task to refine the details of nuclei and contours. Extensive experiments are conducted to evaluate the proposed method qualitatively and quantitatively. Experimental results show that REU-Net outperforms the state-of-the-art methods on HUSTS, MoNuSeg, CoNSep and CPM-17 datasets.

The cells under the microscope are not in the same focal plane due to adhesion and overlap，and image fusion is required to obtain a clear image. The existing method is inefficient to fuse the whole image. In order to solve this problem，this paper proposes a method of cell image fusion based on content analysis. Firstly，the cell images at different depth positions under the same field of view were obtained by fast asynchronous image capture， and the images for fusion were screened out. Then the image was segmented，and the image content was divided into single cell，group cell and impurity through the VGG-16 model classifier，and then fused with different methods. Do not deal with garbage; Single cell image fusion was selected by contrast definition. Guided filtering was used to fuse the cluster cells. Finally，the fused cell contents are stitched together into a clear image as the final result. This method only performs complex fusion operation on the group cells，and effectively reduces the time complexity of the operation. The experimental results show that the proposed method has a 219% higher average fusion time and higher fusion efficiency than the existing methods.

With the development of artificial intelligence technology, the automatic reading system plays an increasingly important role in assisting the diagnosis of pathologists, improving the accuracy of pathology diagnosis and reducing labor intensity. Accurate segmentation of the nucleus is the primary factor affecting the performance of the automated reading system. Because the boundary between the nucleus, the cytoplasm and the background is unclear, and the color difference between the cells is large, the nuclear segmentation is challenged. In order to solve this problem, a method of cervical nucleus segmentation based on optimal maximum stability regions(Maximally Stable Extremal Regions, MSER) algorithm is proposed. This method first converts the image to the HSV (Hue, Saturation, Value) color space. Then, after weighted combination of S and V channels, the optimized MSER algorithm is used to obtain a coarse segmentation region with uniform gray values. The parameter segmentation method is used to perform fine segmentation. Finally, the feature extraction technique is used to extract various features from the nuclear image, and the artificial neural network classifier is trained to judge whether the result obtained after segmentation is the nucleus. Experiments show that the method can accurately segment the cervical nucleus

3D model retrieval based on representative views was proposed. On the view representation of the 3D model, in order to fully represent the model and reduce redundant information, we firstly adopt Light Field Descriptor (LFD) to generate 2D views, and then use K-MEANS to get representative views from the 2D views. Next, a Convolution Neural Network (CNN) is adopted to extract the view feature and classify. At the same time, a similarity metrics supporting multiple query method is proposed to realize model retrieval with sketches, pictures or 3D models as input. Results on ModelNet40 showed that the proposed method could achieve an accuracy of 100% for part of models with distinct features