Feature Refinement Research Articles

Progressive Frequency-Guided Depth Model with Adaptive Preprocessing for Casting Defect Detection

This article proposes a progressive frequency domain-guided depth model with adaptive preprocessing to solve the problem of defect detection with weak features based on X-ray images. In distinct intuitive surface defect detection tasks, non-destructive testing of castings using X-rays presents more complex and weak defect features, leading to lower accuracy and insufficient robustness on the part of current casting defect detection methods. To address these challenges, the proposed method establishes four specialized mechanisms to improve model accuracy. First, an adaptive image contrast enhancement method is proposed to enhance the features of defects in casting images to promote subsequent feature extraction and prediction. Second, a subtle clue mining module based on frequency domain attention is proposed to fully extract the discriminative features of casting defects. Third, a feature refinement module based on progressive learning is proposed to achieve a balance between feature resolution and semantic information. Finally, a refined deep regression supervision mechanism is designed to improve defect detection accuracy under strict intersection-to-union ratio standards. We established extensive ablation studies using casting defect images in GDXray, conducted detailed comparative experiments with other methods, and performed experiments to analyze the robustness of the resulting models. Compared with other X-ray defect detection methods, our framework achieves an average +4.6 AP. Compared to the baseline, our proposed refined deep regression supervision mechanism results in an improvement of 5.3 AP.

Research on Gangue Detection Algorithm Based on Cross-Scale Feature Fusion and Dynamic Pruning

The intelligent identification of coal gangue on industrial conveyor belts is a crucial technology for the precise sorting of coal gangue. To address the issues in coal gangue detection algorithms, such as high false negative rates, complex network structures, and substantial model weights, an optimized coal gangue detection algorithm based on YOLOv5s is proposed. In the backbone network, a feature refinement module is employed for feature extraction, enhancing the capability to extract features for coal and gangue. The improved BIFPN structure is employed as the feature pyramid, augmenting the model’s capability for cross-scale feature fusion. In the prediction layer, the ESIOU is utilized as the bounding box regression loss function to rectify the misalignment issue between predicted and actual box angles. This approach expedites the convergence speed of the network while concurrently enhancing the accuracy of coal gangue detection. Channel pruning is implemented on the network to diminish model computational complexity and weight, consequently augmenting detection speed. The experimental results demonstrate that the refined YOLOv5s coal gangue detection algorithm outperforms the original YOLOv5s algorithm, achieving a notable accuracy enhancement of 2.2% to reach 93.8%. Concurrently, a substantial reduction in model weight by 38.8% is observed, resulting in a notable 56.2% increase in inference speed. These advancements meet the detection requirements for scenarios involving mixed coal gangue.

Diverse branch feature refinement network for efficient multi‐scale super‐resolution

AbstractDespite the existence of various super‐resolution (SR) methods, most of them focus on designing models for specific upscaling factors rather than fully exploiting inter‐scale correlation to improve efficiency. In contrast, multi‐scale SR methods can effectively reduce the redundancy of network parameters by aggregating the feature extraction processes corresponding to multiple scales into a unified process. The aim of this study is to enhance the compactness and efficiency of the SR model. Thus, an efficient multi‐scale SR method called the diverse branch feature refinement network (DBFRN) is proposed. By decoupling the training process and inference process based on the idea of structural re‐parameterization, multi‐branch topology is adopted to enrich multi‐scale learning and merge branches to achieve efficient inference with equivalent effects. Specifically, two re‐parameterization strategies are designed and two corresponding feature refinement blocks for different feature levels in multi‐scale SR network. Extensive experiments demonstrate that the proposed multi‐scale SR method is effective and efficient, and it can outperform advanced single‐scale methods in terms of quantity and quality.

A Lightweight Context-Aware Feature Transformer Network for Human Pose Estimation

We propose a Context-aware Feature Transformer Network (CaFTNet), a novel network for human pose estimation. To address the issue of limited modeling of global dependencies in convolutional neural networks, we design the Transformerneck to strengthen the expressive power of features. Transformerneck directly substitutes 3×3 convolution in the bottleneck of HRNet with a Contextual Transformer (CoT) block while reducing the complexity of the network. Specifically, the CoT first produces keys with static contextual information through 3×3 convolution. Then, relying on query and contextualization keys, dynamic contexts are generated through two concatenated 1×1 convolutions. Static and dynamic contexts are eventually fused as an output. Additionally, for multi-scale networks, in order to further refine the features of the fusion output, we propose an Attention Feature Aggregation Module (AFAM). Technically, given an intermediate input, the AFAM successively deduces attention maps along the channel and spatial dimensions. Then, an adaptive refinement module (ARM) is exploited to activate the obtained attention maps. Finally, the input undergoes adaptive feature refinement through multiplication with the activated attention maps. Through the above procedures, our lightweight network provides powerful clues for the detection of keypoints. Experiments are performed on the COCO and MPII datasets. The model achieves a 76.2 AP on the COCO val2017 dataset. Compared to other methods with a CNN as the backbone, CaFTNet has a 72.9% reduced number of parameters. On the MPII dataset, our method uses only 60.7% of the number of parameters, acquiring similar results to other methods with a CNN as the backbone.

Semantic Correlation Attention-Based Multiorder Multiscale Feature Fusion Network for Human Motion Prediction.

Human motion prediction is to predict future human states based on the observed human states. However, current research ignores the semantic correlations between body parts (joints and bones) in the observed human states and motion time; thus, the prediction accuracy is limited. To address this issue, we propose a novel semantic correlation attention-based multiorder multiscale feature fusion network (SCAFF), which includes an encoder and a decoder. In the encoder, a multiorder difference calculation module (MODC) is designed to calculate the multiorder difference information of joint and bone attributes in the observed human states. Then, multiple semantic correlation attention-based graph calculation operators (SCA-GCOs) are stacked to extract the multiscale features of the multiorder difference information. Each SCA-GCO captures joint and bone dependencies of the multiorder difference information, refines them with a semantic correlation attention module (SCAM), and captures temporal dynamics of the refined joint and bone dependencies as the output features. Note that SCAM learns a semantic attention mask describing the semantic correlations between body parts and motion time for feature refinement. Afterward, multiple multiorder feature fusion modules (MOFFs) and multiscale feature fusion modules (MSFFs) are designed to fuse the multiscale features of the multiorder difference information extracted by multiple SCA-GCOs, thus obtaining the motion features of the observed human states. Based on the obtained motion features, the decoder recurrently recruits a composite gated recurrent module (CGRM) and multilayer perceptrons (MLPs) to predict future human states. As far as we know, this is the first attempt to consider the semantic correlations between body parts and motion time in human motion prediction. The results on public datasets demonstrate that SCAFF outperforms existing models.

Hybrid Attention Feature Refinement Network for Lightweight Image Super-Resolution in Metaverse Immersive Display

Hybrid Attention Feature Refinement Network for Lightweight Image Super-Resolution in Metaverse Immersive Display

Swin-Net: A Swin-Transformer-Based Network Combing with Multi-Scale Features for Segmentation of Breast Tumor Ultrasound Images

Breast cancer is one of the most common cancers in the world, especially among women. Breast tumor segmentation is a key step in the identification and localization of the breast tumor region, which has important clinical significance. Inspired by the swin-transformer model with powerful global modeling ability, we propose a semantic segmentation framework named Swin-Net for breast ultrasound images, which combines Transformer and Convolutional Neural Networks (CNNs) to effectively improve the accuracy of breast ultrasound segmentation. Firstly, our model utilizes a swin-transformer encoder with stronger learning ability, which can extract features of images more precisely. In addition, two new modules are introduced in our method, including the feature refinement and enhancement module (RLM) and the hierarchical multi-scale feature fusion module (HFM), given that the influence of ultrasonic image acquisition methods and the characteristics of tumor lesions is difficult to capture. Among them, the RLM module is used to further refine and enhance the feature map learned by the transformer encoder. The HFM module is used to process multi-scale high-level semantic features and low-level details, so as to achieve effective cross-layer feature fusion, suppress noise, and improve model segmentation performance. Experimental results show that Swin-Net performs significantly better than the most advanced methods on the two public benchmark datasets. In particular, it achieves an absolute improvement of 1.4–1.8% on Dice. Additionally, we provide a new dataset of breast ultrasound images on which we test the effect of our model, further demonstrating the validity of our method. In summary, the proposed Swin-Net framework makes significant advancements in breast ultrasound image segmentation, providing valuable exploration for research and applications in this domain.

Weighted Guided Optional Fusion Network for RGB-T Salient Object Detection

There is no doubt that the rational and effective use of visible and thermal infrared image data information to achieve cross-modal complementary fusion is the key to improving the performance of RGB-T salient object detection (SOD). A meticulous analysis of the RGB-T SOD data reveals that it mainly consists of three scenarios in which both modalities (RGB and T) have a significant foreground and only a single modality (RGB or T) is disturbed. However, existing methods are obsessed with pursuing more effective cross-modal fusion based on treating both modalities equally. Obviously, the subjective use of equivalence has two significant limitations. Firstly, it does not allow for practical discrimination of which modality makes the dominant contribution to performance. While both modalities may have visually significant foregrounds, differences in their imaging properties will result in distinct performance contributions. Secondly, in a specific acquisition scenario, a pair of images with two modalities will contribute differently to the final detection performance due to their varying sensitivity to the same background interference. Intelligibly, for the RGB-T saliency detection task, it would be more reasonable to generate exclusive weights for the two modalities and select specific fusion mechanisms based on different weight configurations to perform cross-modal complementary integration. Consequently, we propose a weighted guided optional fusion network (WGOFNet) for RGB-T SOD. Specifically, a feature refinement module is first used to perform an initial refinement of the extracted multilevel features. Subsequently, a weight generation module (WGM) will generate exclusive network performance contribution weights for each of the two modalities, and an optional fusion module (OFM) will rely on this weight to perform particular integration of cross-modal information. Simple cross-level fusion is finally utilized to obtain the final saliency prediction map. Comprehensive experiments on three publicly available benchmark datasets demonstrate the proposed WGOFNet achieves superior performance compared with the state-of-the-art RGB-T SOD methods. The source code is available at: https://github.com/WJ-CV/WGOFNet .

Structured Life Narratives: Building Life Story Hierarchies with Graph-Enhanced Event Feature Refinement

The life stories of older adults encapsulate an array of personal experiences that reflect their care needs. However, due to inherent fuzzy features, fragmented natures, repetition, and redundancies, the practical application of the life story approach poses challenges for caregivers in acquiring and comprehending these narratives. Addressing this challenge, our study introduces a novel approach called Life Story Hierarchies with Graph-Enhanced Event Feature Refinement (LSH-GEFR). LSH-GEFR constructs a bilayer graph. Firstly, the event element map leverages intricate relationships between event elements to extract environmental features, providing a detailed context for understanding each event element. Secondly, the event map explores the complex web of relationships between the events themselves, allowing LSH-GEFR to generate a comprehensive understanding of each event and enhance its representation. Subsequently, we conducted experiments on different datasets and found that, in comparison with four advanced event tree generation methods, the proposed LSH-GEFR method outperformed them in terms of path coherence, branch reasonableness, and overall readability when generating life story hierarchies. Over 84.91% of the structured life narratives achieved readability, marking a 5.96% increase over the best-performing approach at the baseline.

Defocus blur detection via adaptive cross-level feature fusion and refinement

Defocus blur detection via adaptive cross-level feature fusion and refinement

Boosting adversarial robustness via feature refinement, suppression, and alignment

AbstractDeep neural networks are vulnerable to adversarial attacks, bringing high risk to numerous security-critical applications. Existing adversarial defense algorithms primarily concentrate on optimizing adversarial training strategies to improve the robustness of neural networks, but ignore that the misguided decisions are essentially made by the activation values. Besides, such conventional strategies normally result in a great decline in clean accuracy. To address the above issues, we propose a novel RSA algorithm to counteract adversarial perturbations while maintaining clean accuracy. Specifically, RSA comprises three distinct modules: feature refinement, activation suppression, and alignment modules. First, the feature refinement module refines malicious activation values in the feature space. Subsequently, the feature activation suppression module mitigates redundant activation values induced by adversarial perturbations across both channel and spatial dimensions. Finally, to avoid an excessive performance drop on clean samples, RSA incorporates a consistency constraint and a knowledge distillation constraint for feature alignment. Extensive experiments on five public datasets and three backbone networks demonstrate that our proposed algorithm achieves consistently superior performance in both adversarial robustness and clean accuracy over the state-of-the-art.

Improved YOLOv5 Network for Detection of Peach Blossom Quantity

In agricultural production, rapid and accurate detection of peach blossom bloom plays a crucial role in yield prediction, and is the foundation for automatic thinning. The currently available manual operation-based detection and counting methods are extremely time-consuming and labor-intensive, and are prone to human error. In response to the above issues, this paper proposes a natural environment peach blossom detection model based on the YOLOv5 model. First, a cascaded network is used to add an output layer specifically for small target detection on the basis of the original three output layers. Second, a combined context extraction module (CAM) and feature refinement module (FSM) are added. Finally, the network clusters and statistically analyzes the range of multi-scale channel elements using the K-means++ algorithm, obtaining candidate box sizes that are suitable for the dataset. A novel bounding box regression loss function (SIoU) is used to fuse the directional information between the real box and the predicted box to improve detection accuracy. The experimental results show that, compared with the original YOLOv5s model, our model has correspondingly improved AP values for identifying three different peach blossom shapes, namely, bud, flower, and falling flower, by 7.8%, 10.1%, and 3.4%, respectively, while the final mAP value for peach blossom recognition increases by 7.1%. Good results are achieved in the detection of peach blossom flowering volume. The proposed model provides an effective method for obtaining more intuitive and accurate data sources during the process of peach yield prediction, and lays a theoretical foundation for the development of thinning robots.

Supervised-unsupervised combined transformer for spectral compressive imaging reconstruction

To solve the low spatial and/or temporal resolution problem which the conventional hyperspectral cameras often suffer from, spectral compressive imaging systems (SCI) have attracted more attention recently. Recovering a hyperspectral image (HSI) from its corresponding 2D coded image is an ill-posed inverse problem, and learning accurate prior from HSI and 2D coded image is essential to solve this inverse problem. Existing methods only use supervised networks that focus on learning generalized prior from training datasets, or only use unsupervised networks that focus on learning specific prior from 2D coded image, resulting in the inability to learn both generalized and specific priors. Also, when learning the priors, existing methods cannot simultaneously give consideration to both global and local scales, as well as both spatial and spectral dimensions. To cope with this problem, in this paper, we propose a Supervised-Unsupervised Combined Transformer Network (SUCTNet) composed by a supervised Spatio-spectral Transformer network (SSTNet) and an Unsupervised Multi-level Feature Refinement network (UMFRNet). Specifically, we first develop the SSTNet to learn generalized prior and obtain a preliminary HSI. In SSTNet, the proposed spatial encoding and spectral decoding network architecture enables it to simultaneously consider both spatial and spectral dimensions, and a proposed Global and Local Multi head Self Attention block (GL-MSA) enables it simultaneously to consider both global and local scales. Then, the preliminary HSI is fed into the proposed UMFRNet to learn specific prior and obtain the target HSI. In UMFRNet, a proposed multi-level feature refinement mechanism and the physical imaging model of SCI are used to improve reconstruction accuracy and generalization performance. Extensive experiments show that our method significantly outperforms state-of-the-art (SOTA) methods on simulated and real datasets. Codes will be available at https://github.com/Vzhouhan/SUCTNet.

CSSNet: Cascaded spatial shift network for multi-organ segmentation

Multi-organ segmentation is vital for clinical diagnosis and treatment. Although CNN and its extensions are popular in organ segmentation, they suffer from the local receptive field. In contrast, MultiLayer-Perceptron-based models (e.g., MLP-Mixer) have a global receptive field. However, these MLP-based models employ fully connected layers with many parameters and tend to overfit on sample-deficient medical image datasets. Therefore, we propose a Cascaded Spatial Shift Network, CSSNet, for multi-organ segmentation. Specifically, we design a novel cascaded spatial shift block to reduce the number of model parameters and aggregate feature segments in a cascaded way for efficient and effective feature extraction. Then, we propose a feature refinement network to aggregate multi-scale features with location information, and enhance the multi-scale features along the channel and spatial axis to obtain a high-quality feature map. Finally, we employ a self-attention-based fusion strategy to focus on the discriminative feature information for better multi-organ segmentation performance. Experimental results on the Synapse (multiply organs) and LiTS (liver & tumor) datasets demonstrate that our CSSNet achieves promising segmentation performance compared with CNN, MLP, and Transformer models. The source code will be available at https://github.com/zkyseu/CSSNet.

Quantitative determination of wheat moisture content based on microwave detection technique combined with multivariate data analysis

In this study, a new method is proposed for the quantitative detection of wheat moisture content based on double heterodyne mixing microwave transmission technique. First of all, the microwave detection device received the command from the computer and detected the wheat in the frequency domain of 2.5 GHz–11.5 GHz, while acquiring the transmission index corresponding to each frequency point. Then, competitive adaptive reweighted sampling (CARS) was used for coarse feature selection of the transmission index after least partial squares filtering, and Genetic Algorithm (GA) was used for secondary feature refinement of the coarse selected feature variables, and they were sorted and combined according to their weights. Finally, Support vector regression (SVR) was used to establish quantitative models with different combinations of feature variables to achieve rapid detection of wheat moisture content. The results showed that the CARS-GA-SVR model had better detection performance for wheat moisture content determination compared to the CARS-SVR model. Among them, the CARS-GA-SVR model had the best detection effect when the combination of feature variables is 6-dimensional, with coefficient of determination (RP2) of 0.9756 and relative prediction deviation (RPD) of 6.3234. Compared to the SVR model established by the full transmission index, the RPD improved from 3.7089 to 6.3234. The results reveal that rapid and high-precision detection of wheat moisture content can be achieved by using a self-made miniaturized microwave device combined with a suitable chemometric method.

A New Framework Based on Supervised Joint Distribution Adaptation for Bearing Fault Diagnosis across Diverse Working Conditions

To address the degradation of diagnostic performance due to data distribution differences and the scarcity of labeled fault data, this study has focused on transfer learning-based cross-domain fault diagnosis, which attracts considerable attention. However, deep transfer learning-based methods often present a challenge due to their time-consuming and costly nature, particularly in tuning hyperparameters. For this issue, on the basis of classical features-based transfer learning method, this study introduces a new framework for bearing fault diagnosis based on supervised joint distribution adaptation and feature refinement. It first utilizes ensemble empirical mode decomposition to process raw signals, and statistical features extraction is implemented. Then, a new feature refinement module is designed to refine domain adaptation features from high-dimensional feature set by evaluating the fault distinguishability and working-condition invariance of feature data. Next, it proposes a supervised joint distribution adaptation method to conduct improved joint distribution alignment that preserves neighborhood relationships within a manifold subspace. Finally, an adaptive classifier is trained to predict fault labels of feature data across varying working conditions. To prove the cross-domain fault diagnosis performance and superiority of the proposed methods, two bearing datasets are applied for experiments, and the experimental results verify that the model built by the proposed framework can achieve desirable diagnosis performance under different working conditions and that it apparently outperforms comparative models.

A generic plug & play diffusion-based denosing module for medical image segmentation

Medical image segmentation faces challenges because of the small sample size of the dataset and the fact that images often have noise and artifacts. In recent years, diffusion models have proven very effective in image generation and have been used widely in computer vision. This paper presents a new feature map denoising module (FMD) based on the diffusion model for feature refinement, which is plug-and-play, allowing flexible integration into popular used segmentation networks for seamless end-to-end training. We evaluate the performance of the FMD module on four models, UNet, UNeXt, TransUNet, and IB-TransUNet, by conducting experiments on four datasets. The experimental data analysis shows that adding the FMD module significantly positively impacts the model performance. Furthermore, especially for small lesion areas and minor organs, adding the FMD module allows users to obtain more accurate segmentation results than the original model.

Colposcopic Image Segmentation Based on Feature Refinement and Attention

Colposcopic Image Segmentation Based on Feature Refinement and Attention

Accelerated Feature Extraction and Refinement for Improved Aerial Scene Categorization

Accelerated Feature Extraction and Refinement for Improved Aerial Scene Categorization

ER-Swin: Feature Enhancement and Refinement Network Based on Swin Transformer for Semantic Segmentation of Remote Sensing Images

ER-Swin: Feature Enhancement and Refinement Network Based on Swin Transformer for Semantic Segmentation of Remote Sensing Images