Channel Attention Mechanism Research Articles

ALTASN: A few-shot learning fault diagnosis method for rotating machinery of unmanned surface vehicles based on attention mechanism

Rotating machinery is one of the principal power equipment of unmanned surface vehicles (UVs). Considering the complex and harsh offshore working conditions of UVs, the health status of rotating machinery is highly to be affected, but it is often difficult to obtain enough fault samples. Accordingly, limited data fault diagnosis of rotating machinery holds great practical significance to increase the resilience and security of UVs. For limited data fault diagnosis, a novel few-shot learning model based on attention mechanism called adaptive long-term attention siamese network (ALTASN) is proposed. First, an efficient channel attention mechanism is combined with adaptive convolutional kernels to improve the spatial feature extraction capabilities of the convolutional neural network (CNN). To capture and assign higher weights to important long-term dependent information, long-term attention is introduced to improve the ability of long short-term memory networks (LSTM) temporal feature extraction. Finally, the siamese network is introduced to compare the features of different sample pairs to obtain the final fault type. In the case of limited data, the fault diagnosis performance and generalization ability of the proposed ALTASN are better compared with existing results. Experiments are carried out on the actual three-phase asynchronous motor experiment platform at the Zhejiang University of Technology to verify the effectiveness and generalization of the proposed method.

Ocean eddy detection based on YOLO deep learning algorithm by synthetic aperture radar data

Ocean eddies play a crucial role in the global energy cycle and significantly impact the transport of heat, salt, and nutrients in the global ocean. Spaceborne synthetic aperture radar (SAR), with its high spatial resolution (O(20 m)) and wide coverage, is an important tool for studying ocean eddies. In this paper, we propose a deep-learning network, named EOLO, to automatically detect ocean Eddies observed in C-band spaceborne SAR imagery, based on the You-Only-Look-Once (YOLO) deep learning algorithm. A Sentinel-1 (S1) SAR data-based ocean eddy dataset (named EddyDataset) is established to train the EOLO network. To effectively improve the performance of the EOLO network, we introduced a channel attention mechanism, adopted the up-sampling operator with the larger receptive field, and improved feature fusion method, anchor box size, and loss function. With the support of a high-performance detection network, the geographic information extraction module based on the affine geographic transformation model and a data preprocessing module were added, making EOLO an application-level framework. Our experiment results on EddyDataset demonstrate that EOLO achieves a high quality of eddy detection with 91.5% average precision. We further applied EOLO to entire scenes of S1 images to detect eddies in the Red Sea, the Baltic Sea, and the Western Mediterranean Sea, achieving 96.6%, 98.8% and 98.9% precision, respectively. Moreover, EOLO was used to extract size and location information of ocean eddies based on 6135 S1 scenes acquired in 2021 over the Western Mediterranean Sea. Their spatial characteristics were derived and compared with the eddies extracted from radar altimeter data, presenting interesting discrepancies between sub-mesoscale and mesoscale eddies in the ocean.

Enhancing Adversarial Learning-Based Change Detection in Imbalanced Datasets Using Artificial Image Generation and Attention Mechanism

Deep Learning (DL) has become a popular method for Remote Sensing (RS) Change Detection (CD) due to its superior performance compared to traditional methods. However, generating extensive labeled datasets for DL models is time-consuming and labor-intensive. Additionally, the imbalance between changed and unchanged areas in object CD datasets, such as buildings, poses a critical issue affecting DL model efficacy. To address this issue, this paper proposes a change detection enhancement method using artificial image generation and attention mechanism. Firstly, the content of the imbalanced CD dataset is enhanced using a data augmentation strategy that synthesizes effective building CD samples using artificial RS image generation and building label creation. The created building labels, which serve as new change maps, are fed into a generator model based on a conditional Generative Adversarial Network (c-GAN) to generate high-resolution RS images featuring building changes. The generated images with their corresponding change maps are then added to the CD dataset to create the balance between changed and unchanged samples. Secondly, a channel attention mechanism is added to the proposed Adversarial Change Detection Network (Adv-CDNet) to boost its performance when training on the imbalanced dataset. The study evaluates the Adv-CDNet using WHU-CD and LEVIR-CD datasets, with WHU-CD exhibiting a higher degree of sample imbalance compared to LEVIR-CD. Training the Adv-CDNet on the augmented dataset results in a significant 16.5% F1-Score improvement for the highly imbalanced WHU-CD. Moreover, comparative analysis showcases the superior performance of the Adv-CDNet when complemented with the attention module, achieving a 6.85% F1-Score enhancement.

An Effective Image Classification Method for Plant Diseases with Improved Channel Attention Mechanism aECAnet Based on Deep Learning

Since plant diseases occurring during the growth process are a significant factor leading to the decline in both yield and quality, the classification and detection of plant leaf diseases, followed by timely prevention and control measures, are crucial for safeguarding plant productivity and quality. As the traditional convolutional neural network structure cannot effectively recognize similar plant leaf diseases, in order to more accurately identify the diseases on plant leaves, this paper proposes an effective plant disease image recognition method aECA-ResNet34. This method is based on ResNet34, and in the first and the last layers of this network, respectively, we add this paper’s improved aECAnet with the symmetric structure. aECA-ResNet34 is compared with different plant disease classification models on the peanut dataset constructed in this paper and the open-source PlantVillage dataset. The experimental results show that the aECA-ResNet34 model proposed in this paper has higher accuracy, better performance, and better robustness. The results show that the aECA-ResNet34 model proposed in this paper is able to recognize diseases of multiple plant leaves very accurately.

A study on expression recognition based on improved mobilenetV2 network

This paper proposes an improved strategy for the MobileNetV2 neural network(I-MobileNetV2) in response to problems such as large parameter quantities in existing deep convolutional neural networks and the shortcomings of the lightweight neural network MobileNetV2 such as easy loss of feature information, poor real-time performance, and low accuracy rate in facial emotion recognition tasks. The network inherits the characteristics of MobilenetV2 depthwise separated convolution, signifying a reduction in computational load while maintaining a lightweight profile. It utilizes a reverse fusion mechanism to retain negative features, which makes the information less likely to be lost. The SELU activation function is used to replace the RELU6 activation function to avoid gradient vanishing. Meanwhile, to improve the feature recognition capability, the channel attention mechanism (Squeeze-and-Excitation Networks (SE-Net)) is integrated into the MobilenetV2 network. Experiments conducted on the facial expression datasets FER2013 and CK + showed that the proposed network model achieved facial expression recognition accuracies of 68.62% and 95.96%, improving upon the MobileNetV2 model by 0.72% and 6.14% respectively, and the parameter count decreased by 83.8%. These results empirically verify the effectiveness of the improvements made to the network model.

DECTNet: Dual Encoder Network combined convolution and Transformer architecture for medical image segmentation.

Automatic and accurate segmentation of medical images plays an essential role in disease diagnosis and treatment planning. Convolution neural networks have achieved remarkable results in medical image segmentation in the past decade. Meanwhile, deep learning models based on Transformer architecture also succeeded tremendously in this domain. However, due to the ambiguity of the medical image boundary and the high complexity of physical organization structures, implementing effective structure extraction and accurate segmentation remains a problem requiring a solution. In this paper, we propose a novel Dual Encoder Network named DECTNet to alleviate this problem. Specifically, the DECTNet embraces four components, which are a convolution-based encoder, a Transformer-based encoder, a feature fusion decoder, and a deep supervision module. The convolutional structure encoder can extract fine spatial contextual details in images. Meanwhile, the Transformer structure encoder is designed using a hierarchical Swin Transformer architecture to model global contextual information. The novel feature fusion decoder integrates the multi-scale representation from two encoders and selects features that focus on segmentation tasks by channel attention mechanism. Further, a deep supervision module is used to accelerate the convergence of the proposed method. Extensive experiments demonstrate that, compared to the other seven models, the proposed method achieves state-of-the-art results on four segmentation tasks: skin lesion segmentation, polyp segmentation, Covid-19 lesion segmentation, and MRI cardiac segmentation.

BiCFormer: Swin Transformer based model for classification of benign and malignant pulmonary nodules

Pulmonary cancer is one of the most common and deadliest cancers worldwide, and the detection of benign and malignant nodules in the lungs can be an important aid in the early diagnosis of lung cancer. Existing convolutional neural networks inherit their limitations by extracting global contextual information, and in most cases prove to be less efficient in obtaining satisfactory results. Transformer-based deep learning methods have obtained good performance in different computer vision tasks, and this study attempts to introduce them into the task of computed tomography (CT) image classification of lung nodules. However, the problems of sample scarcity and difficulty of local feature extraction in this field. To this end, we are inspired by Swin Transformer to propose a model named BiCFormer for the task of classifying and diagnosing CT scan images of lung nodules. Specifically, first we introduce a multi-layer discriminator generative adversarial network module for data augmentation to assist the model in extracting features more accurately. Second, unlike the encoder of traditional Transformer, we divide the encoder part of BiCFormer into two parts: bi-level coordinate (BiC) and fast-partial-window (FPW). The BiC module has a part similar to the traditional channel attention mechanism is able to enhance the performance of the model, and is more able to enhance the representation of attention object features by aggregating features along two spatial directions. The BiC module also has a dynamic sparse attention mechanism that filters out irrelevant key-value pairs in rough regions, allowing the model to focus more on features of interest. The FPW module is mainly used to reduce computational redundancy and minimize feature loss. We conducted extensive experiments on the LIDC-IDRI dataset. The experimental results show that our model achieves an accuracy of 97.4% compared to other studies using this dataset for lung nodule classification, making it an effective and competitive method.

Improved metric-based meta learning with attention mechanism for few-shot cross-domain train bearing fault diagnosis

Traction motor bearings, as a crucial component of subway trains, play a pivotal role in ensuring the safety of train operations. Therefore, intelligent diagnosis of train bearings holds significant importance. However, due to the complex and dynamic nature of bearing conditions coupled with limited fault data availability, traditional diagnostic methods fail to yield satisfactory results. To address this issue, we propose an improved metrics-based meta-learning approach for accurate few-shot cross-domain fault diagnosis of train bearings. Firstly, we introduce a 1D-signal channel attention mechanism that effectively extracts latent features and enhances recognition accuracy. Secondly, by incorporating the Adabound algorithm into our model framework, we further enhance its classification performance. Finally, through several case studies, we demonstrate the effectiveness of our proposed method in comparison to other approaches within similar settings.

A Data Augmentation Method for Insulators Based on Cycle GAN

Data augmentation is an important task of using existing data to expand data sets. Using generative countermeasure network technology to realize data augmentation has the advantages of high-quality generated samples, simple training, and fewer restrictions on the number of generated samples. However, in the field of transmission line insulator images, the freely synthesized samples are prone to produce fuzzy backgrounds and disordered samples of the main insulator features. To solve the above problems, this paper uses cycle generative adversarial network (Cycle-GAN) used for domain conversion in the generation countermeasure network as the initial framework and uses the self-attention mechanism and channel attention mechanism to assist the conversion to realize the mutual conversion of different insulator samples. The attention module with prior knowledge is used to build the generation countermeasure network, and the GAN model with local controllable generation is built to realize the directional generation of insulator belt defect samples. The experimental results show that the samples obtained by this method are improved in a number of quality indicators, and the quality effect of the samples obtained is excellent, which has a reference value for the data expansion of insulator images.

Ternary Precursor Centrifuge Rolling Bearing Fault Diagnosis Based on Adaptive Sample Length Adjustment of 1DCNN-SeNet

To address the issues of uneven sample lengths in the centrifuge machine bearings of the ternary precursor, inaccurate fault feature extraction, and insensitivity of important feature channels in rolling bearings, a rolling bearing fault diagnosis method based on adaptive sample length adjustment of one-dimensional convolutional neural network (1DCNN) and squeeze-and-excitation network (SeNet) is proposed. Firstly, by controlling the cumulative variance contribution rate in the principal component analysis algorithm, adaptive adjustment of sample length is achieved, reducing data with uneven sample lengths to the same dimensionality for various classes. Then, the 1DCNN extracts local features from bearing signals through one-dimensional convolution-pooling operations, while the SeNet network introduces a channel attention mechanism which can adaptively adjust the importance between different channels. Finally, the 1DCNN-SeNet model is compared with four classic models through experimental analysis on the CWRU bearing dataset. The experimental results indicate that the proposed method exhibits high diagnostic accuracy in rolling bearings, demonstrating good adaptability and generalization capabilities.

Hybrid Spatial-Channel Attention Mechanism for Cross-Age Face Recognition

Face recognition techniques have been widely employed in real-world biomimetics applications. However, traditional approaches have limitations in recognizing faces correctly with large age differences because of significant changes over age in the same person, leading to unsatisfactory recognition performance. To address this, previous studies propose to decompose and identify age and identity features independently in facial images across diverse age groups when optimizing the discriminative model so as to improve the age-invariant face recognition accuracy. Nevertheless, the interrelationships between these features make it difficult for the decomposition to disentangle them properly, thus compromising the recognition accuracy due to the interactive impacts on both features. To this end, this paper proposes a novel deep framework that incorporates a novel Hybrid Spatial-Channel Attention Module to facilitate the cross-age face recognition task. Particularly, the proposed module enables better decomposition of the facial features in both spatial and channel dimensions with attention mechanisms simultaneously while mitigating the impact of age variation on the recognition performance. Beyond this, diverse pooling strategies are also combined when applying those spatial and channel attention mechanisms, which allows the module to generate discriminative face representations while preserving complete information within the original features, further yielding sounder recognition accuracy. The proposed model is extensively validated through experiments on public face datasets such as CACD-VS, AgeDB-30, and FGNET, where the results show significant performance improvements compared to competitive baselines.

AC-CfC: An attention-based convolutional closed-form continuous-time neural network for raw multi-channel EEG-based emotion recognition

Emotion recognition based on electroencephalogram (EEG) is a critical task in the field of affective brain-computer interfaces. However, due to the non-stationarity and individual variability of EEG, hand-designed features cannot adequately capture the nonlinear and high-dimensional properties of raw EEG. Currently, spatial–temporal models have been verified to effectively capture the spatial–temporal information for EEG. However, these models are characterized by complex structures, large parameter counts, and the need for extensive training data. To overcome these disadvantages, this paper propose an end-to-end spatial–temporal model for emotion recognition based on raw EEG, called attention-based convolutional closed-form continuous-time neural network (AC-CfC). This model employs a channel attention mechanism to weight and encode EEG, capturing channel dependencies. Subsequently, one-dimensional convolutional neural networks and closed-form continuous-time neural networks are used to extract deep spatial–temporal features. Additionally, an adaptive loss-controlling mechanism is designed to enhance the model's decision-making ability between classes that are easily confused. To verify the effectiveness of the proposed model, experiments are conducted on the DEAP and DREAMER datasets. The average accuracies of the proposed model reach to 94.76% and 93.01% for valence and arousal in subject-independent experiments on the DEAP dataset, with an improvement of 11.8% and 8.73% respectively compared with the state-of-art model ACRNN. On the DREAMER dataset, the average accuracies reach to 81.83%, 81.22%, and 80.63% for valence, arousal and dominance, with an improvement of 2.55%, 6.64%, and 6.98% respectively over ACRNN. These results show that our proposed model exhibits better performance than some state-of-art models in the same category.

A Study for Design of Lightweight Dense Connection Network on Hyperspectral Image Classification

The characteristics of hyperspectral remote sensing images such as inconspicuous feature representativeness, single feature level, and complex information content, can lead to unstable classification results. We propose a lightweight dense network model that injects channel attention in the form of dense connections between network layers (DSE-DN) for the classification of hyperspectral images. In the DSE-DN network, principal component analysis (PCA) is applied to reduce redundancy in the hyperspectral images. Subsequently, a densely connected network is constructed, incorporating channel attention mechanisms through dense connections to enhance the analysis of spectral image features. Finally, the processed hyperspectral images are classified using a fully interconnected layer. We assess two classical hyperspectral datasets and construct 2DCNN, 3DCNN, ResNet, and the network that injects channel attention layer by layer to compare with DSE-DN. The experimental results indicate the utility of the DSE-DN network in hyperspectral image classification and its superiority over other networks.

Research on PointPillars Algorithm Based on Feature-Enhanced Backbone Network

In the industrial field, the 3D target detection algorithm PointPillars has gained popularity. Improving target detection accuracy while maintaining high efficiency has been a significant challenge. To address the issue of low target detection accuracy in the PointPillars 3D target detection algorithm, this paper proposes an algorithm based on feature enhancement to improve the backbone network. The algorithm enhances preliminary feature information of the backbone network by modifying it based on PointPillars with the aid of channel attention and spatial attention mechanisms. To address the inefficiency caused by the excessive number of subsampled parameters in PointPillars, FasterNet (a lightweight and efficient feature extraction network) is utilized for down-sampling and forming different scale feature maps. To prevent the loss and blurring of extracted features resulting from the use of inverse convolution, we utilize the lightweight and efficient up-sampling modules Carafe and Dysample for adjusting resolution. Experimental results indicate improved accuracy under all difficulties of the KITTI dataset, demonstrating the superiority of the algorithm over PointPillars.

DECIDE: A decoupled semantic and boundary learning network for precise osteosarcoma segmentation by integrating multi-modality MRI

Automated Osteosarcoma Segmentation in Multi-modality MRI (AOSMM) holds clinical significance for effective tumor evaluation and treatment planning. However, the precision of AOSMM is challenged by the diverse characteristics of multi-modality MRI and the inherent heterogeneity and boundary ambiguity of osteosarcoma. While numerous methods have made significant strides in automated osteosarcoma segmentation, they primarily focused on the use of a single MRI modality and overlooked the potential benefits of integrating complementary information from other MRI modalities. Furthermore, they did not adequately model the long-range dependencies of complex tumor features, which may lead to insufficiently discriminative feature representations. To this end, we propose a decoupled semantic and boundary learning network (DECIDE) to achieve precise AOSMM with three functional modules. The Multi-modality Feature Fusion and Recalibration (MFR) module adaptively fuses and recalibrates multi-modality features by exploiting their channel-wise dependencies to compute low-rank attention weights for effectively aggregating useful information from different MRI modalities, which promotes complementary learning between multi-modality MRI and enables a more comprehensive tumor characterization. The Lesion Attention Enhancement (LAE) module employs spatial and channel attention mechanisms to capture global contextual dependencies over local features, significantly enhancing the discriminability and representational capacity of intricate tumor features. The Boundary Context Aggregation (BCA) module further enhances semantic representations by utilizing boundary information for effective context aggregation while also ensuring intra-class consistency in cases of boundary ambiguity. Substantial experiments demonstrate that DECIDE achieves exceptional performance in osteosarcoma segmentation, surpassing state-of-the-art methods in terms of accuracy and stability.

The Detection of Ear Tag Dropout in Breeding Pigs Using a Fused Attention Mechanism in a Complex Environment

The utilization of ear tags for identifying breeding pigs is a widely used technique in the field of animal production. Ear tag dropout can lead to the loss of pig identity information, resulting in missing data and ambiguity in production management and genetic breeding data. Therefore, the identification of ear tag dropout is crucial for intelligent breeding in pig farms. In the production environment, promptly detecting breeding pigs with missing ear tags is challenging due to clustering overlap, small tag targets, and uneven sample distributions. This study proposes a method for detecting the dropout of breeding pigs’ ear tags in a complex environment by integrating an attention mechanism. Firstly, the approach involves designing a lightweight feature extraction module called IRDSC using depthwise separable convolution and an inverted residual structure; secondly, the SENet channel attention mechanism is integrated for enhancing deep semantic features; and finally, the IRDSC and SENet modules are incorporated into the backbone network of Cascade Mask R-CNN and the loss function is optimized with Focal Loss. The proposed algorithm, Cascade-TagLossDetector, achieves an accuracy of 90.02% in detecting ear tag dropout in breeding pigs, with a detection speed of 25.33 frames per second (fps), representing a 2.95% improvement in accuracy, and a 3.69 fps increase in speed compared to the previous method. The model size is reduced to 443.03 MB, a decrease of 72.90 MB, which enables real-time and accurate dropout detection while minimizing the storage requirements and providing technical support for the intelligent breeding of pigs.

A Three-branch Jointed Feature and Topology Decoder guided by game-theoretic interactions for temporomandibular joint segmentation

Segmentation of the temporomandibular joint (TMJ) disc and condyle from magnetic resonance imaging (MRI) is a crucial task in TMJ internal derangement research. The automatic segmentation of the disc structure presents challenges due to its intricate and variable shapes, low contrast, and unclear boundaries. Existing TMJ segmentation methods often overlook spatial and channel information in features and neglect overall topological considerations, with few studies exploring the interaction between segmentation and topology preservation. To address these challenges, we propose a Three-Branch Jointed Feature and Topology Decoder (TFTD) for the segmentation of TMJ disc and condyle in MRI. This structure effectively preserves the topological information of the disc structure and enhances features. We introduce a cross-dimensional spatial and channel attention mechanism (SCIA) to enhance features. This mechanism captures spatial, channel, and cross-dimensional information of the decoded features, leading to improved segmentation performance. Moreover, we explore the interaction between topology preservation and segmentation from the perspective of game theory. Based on this interaction, we design the Joint Loss Function (JLF) to fully leverage the features of segmentation, topology preservation, and joint interaction branches. Results on the TMJ MRI dataset demonstrate the superior performance of our TFTD compared to existing methods.

PETformer network enables ultra-low-dose total-body PET imaging without structural prior

Objective. Positron emission tomography (PET) is essential for non-invasive imaging of metabolic processes in healthcare applications. However, the use of radiolabeled tracers exposes patients to ionizing radiation, raising concerns about carcinogenic potential, and warranting efforts to minimize doses without sacrificing diagnostic quality. Approach. In this work, we present a novel neural network architecture, PETformer, designed for denoising ultra-low-dose PET images without requiring structural priors such as computed tomography (CT) or magnetic resonance imaging. The architecture utilizes a U-net backbone, synergistically combining multi-headed transposed attention blocks with kernel-basis attention and channel attention mechanisms for both short- and long-range dependencies and enhanced feature extraction. PETformer is trained and validated on a dataset of 317 patients imaged on a total-body uEXPLORER PET/CT scanner. Main results. Quantitative evaluations using structural similarity index measure and liver signal-to-noise ratio showed PETformer’s significant superiority over other established denoising algorithms across different dose-reduction factors. Significance. Its ability to identify and recover intrinsic anatomical details from background noise with dose reductions as low as 2% and its capacity in maintaining high target-to-background ratios while preserving the integrity of uptake values of small lesions enables PET-only fast and accurate disease diagnosis. Furthermore, PETformer exhibits computational efficiency with only 37 M trainable parameters, making it well-suited for commercial integration.

A novel deep Siamese framework for burned area mapping Leveraging mixture of experts

Due to the complexity of the areas and the diversity of the objects, traditional Burned Area Mapping (BAM) methods cannot provide promising results. Moreover, these methods focus on additional processing to improve their results, which is time-consuming and complex. Therefore, an advanced framework is needed to achieve accurate results in burned area mapping. In this context, this study proposes a novel Siamese-based mixture of expert networks for burned area mapping, which takes advantage of a mixture of experts (MoE) and position and channel attention mechanisms for deep feature generation. The proposed framework has two deep feature extractor channels to account for the bi-temporal multispectral pre- and post-fire input datasets. To evaluate the performance of the SMoE model, three multispectral Sentinel-2 were used in different countries. The results were compared with other advanced machine and deep learning models, including Light Gradient Boosting Machine (LGBM), Extreme Gradient Boosting (XGBoost), 2D-Siamse Convolutional Neural Network (2D-SCNN), and 3D-Siamse Convolutional Neural Network (3D-SCNN). The result of the burned mapping shows that the proposed model has a high effectiveness compared to other models, as it provides an average accuracy of more than 98% and 0.91 b y overall accuracy (OA) and kappa coefficient (KC) indices, respectively.

Omni-Kernel Network for Image Restoration

Image restoration aims to reconstruct a high-quality image from a degraded low-quality observation. Recently, Transformer models have achieved promising performance on image restoration tasks due to their powerful ability to model long-range dependencies. However, the quadratically growing complexity with respect to the input size makes them inapplicable to practical applications. In this paper, we develop an efficient convolutional network for image restoration by enhancing multi-scale representation learning. To this end, we propose an omni-kernel module that consists of three branches, i.e., global, large, and local branches, to learn global-to-local feature representations efficiently. Specifically, the global branch achieves a global perceptive field via the dual-domain channel attention and frequency-gated mechanism. Furthermore, to provide multi-grained receptive fields, the large branch is formulated via different shapes of depth-wise convolutions with unusually large kernel sizes. Moreover, we complement local information using a point-wise depth-wise convolution. Finally, the proposed network, dubbed OKNet, is established by inserting the omni-kernel module into the bottleneck position for efficiency. Extensive experiments demonstrate that our network achieves state-of-the-art performance on 11 benchmark datasets for three representative image restoration tasks, including image dehazing, image desnowing, and image defocus deblurring. The code is available at https://github.com/c-yn/OKNet.