Modulation Of Aggregation Research Articles

HRD-Net: High resolution segmentation network with adaptive learning ability of retinal vessel features

Retinal segmentation is a crucial step in the early warning of human health conditions. However, retinal blood vessels possess complex curvature, irregular distribution, and contain multi-scale fine structures, which make the limited receptive field of regular convolution challenging to process their vascular details efficiently. Additionally, the encoder-decoder based network leads to irreversible spatial information loss because of multiple downsampling, resulting in over-segmentation and missed segmentation of the vessels. For this reason, we develop a high-resolution network based on Deformable Convolution v3, called HRD-Net. By constructing a high-resolution representation, the network allows special attention to be paid to the details of tiny blood vessels. The proposed feature enhancement cascade module based on Deformable Convolution v3 can flexibly adapt and capture the ever-changing morphology and intricate connections of retinal blood vessels, ensuring the continuity of vessel segmentation. In the output phase of the network, the proposed global aggregation module integrates full-resolution feature maps while suppressing redundant features, achieving an effective fusion of high-level semantic information and spatial detail information. In addition, we have re-examined the selection criteria for activation and normalization methods, and also refine the network architectures from a spatial domain perspective to release redundant computational loads. Testing on the DRIVE, STARE, and CHASE_DB1 datasets indicates that HRD-Net, with fewer parameters, outperforms existing segmentation methods on several evaluation metrics such as F1, ACC, SE, SP, AUC, and IOU.

A new segmentation algorithm for peripapillary atrophy and optic disk from ultra-widefield Photographs

Background and objectiveThe prevalence of myopia and high myopia is increasing globally, underscoring the growing importance of diagnosing high myopia-related pathologies. While existing image segmentation models, such as U-Net, UNet++, ResU-Net, and TransUNet, have achieved significant success in medical image segmentation, they still face challenges when dealing with ultra-widefield (UWF) fundus images. This study introduces a novel automatic segmentation algorithm for the optic disc and peripapillary atrophy (PPA) based on UWF fundus images, aimed at assisting ophthalmologists in more accurately diagnosing high myopia-related diseases. MethodsIn this study, we developed a segmentation model leveraging a Transformer-based network structure, complemented by atrous convolution and selective boundary aggregation modules, to elevate the accuracy of segmenting the optic disc and PPA in UWF photography. The atrous convolution module adeptly manages multi-scale features, catering to the variances in target sizes and expanding the deep network's receptive field. Concurrently, the incorporation of the selective boundary aggregation module in the skip connections of the model significantly improves the differentiation of boundary information between segmentation targets. Moreover, the comparison of our proposed algorithm with classical segmentation models like U-Net, UNet++, ResU-Net, and TransUNet highlights its considerable advantages in processing UWF photographs. ResultsThe experimental results show that, compared to the other four models, our algorithm demonstrates substantial improvements in segmenting the optic disc and PPA in UWF photographs. In PPA segmentation, our algorithm improves by 0.8% in Dice, 1.8% in sensitivity, and 1.3% in intersection over union (IOU). In optic disc segmentation, our algorithm improves by 0.3% in Dice, 0.6% in precision, and 0.4% in IOU. ConclusionOur proposed method improves the segmentation accuracy of PPA and optic disks based on UWF photographs, which is valuable for diagnosing high myopia-related diseases in ophthalmology clinics.

Integrating image and gene-data with a semi-supervised attention model for prediction of KRAS gene mutation status in non-small cell lung cancer.

KRAS is a pathogenic gene frequently implicated in non-small cell lung cancer (NSCLC). However, biopsy as a diagnostic method has practical limitations. Therefore, it is important to accurately determine the mutation status of the KRAS gene non-invasively by combining NSCLC CT images and genetic data for early diagnosis and subsequent targeted therapy of patients. This paper proposes a Semi-supervised Multimodal Multiscale Attention Model (S2MMAM). S2MMAM comprises a Supervised Multilevel Fusion Segmentation Network (SMF-SN) and a Semi-supervised Multimodal Fusion Classification Network (S2MF-CN). S2MMAM facilitates the execution of the classification task by transferring the useful information captured in SMF-SN to the S2MF-CN to improve the model prediction accuracy. In SMF-SN, we propose a Triple Attention-guided Feature Aggregation module for obtaining segmentation features that incorporate high-level semantic abstract features and low-level semantic detail features. Segmentation features provide pre-guidance and key information expansion for S2MF-CN. S2MF-CN shares the encoder and decoder parameters of SMF-SN, which enables S2MF-CN to obtain rich classification features. S2MF-CN uses the proposed Intra and Inter Mutual Guidance Attention Fusion (I2MGAF) module to first guide segmentation and classification feature fusion to extract hidden multi-scale contextual information. I2MGAF then guides the multidimensional fusion of genetic data and CT image data to compensate for the lack of information in single modality data. S2MMAM achieved 83.27% AUC and 81.67% accuracy in predicting KRAS gene mutation status in NSCLC. This method uses medical image CT and genetic data to effectively improve the accuracy of predicting KRAS gene mutation status in NSCLC.

Structural insights into the modulation Of SOD1 aggregation By a fungal metabolite Phialomustin-B: Therapeutic potential in ALS.

Amyotrophic lateral sclerosis (ALS) is a fatal human motor neuron disease leading to muscle atrophy and paralysis. Mutations in superoxide dismutase 1 (SOD1) are associated with familial ALS (fALS). The SOD1 mutants in ALS have a toxic-gain of function by destabilizing the functional SOD1 homodimer, consequently inducing fibril-like aggregation with a cytotoxic non-native trimer intermediate. Therefore, reducing SOD1 oligomerization via chemical modulators is an optimal therapy in ALS. Here, we report the discovery of Phialomustin-B, an unsaturated secondary metabolite from the endophytic fungus Phialophora mustea, as a modulator of SOD1 aggregation. The crystal structure of the SOD1-Phialomustin complex refined to 1.90 Å resolution demonstrated for the first time that the ligand binds to the dimer interface and the lateral region near the electrostatic loop. The aggregation analyses of SOD1WT and the disease mutant SOD1A4V revealed that Phialomustin-B reduces cytotoxic trimerization. We propose that Phialomustin-B is a potent lead molecule with therapeutic potential in fALS.

HyperHAR

Human activity recognition (HAR) has emerged as a prominent research field in recent years. Current HAR models are only able to model bilateral correlations between two sensing devices for feature extraction. However, for some activities, exploiting correlations among more than two sensing devices, which we call hyper-correlations in this paper, is essential for extracting discriminatory features. In this work, we propose a novel HyperHAR framework that automatically models both bilateral and hyper-correlations among sensing devices. The HyperHAR consists of three modules. The Intra-sensing Device Feature Extraction Module generates latent representation across the data of each sensing device, based on which the Inter-sensing Device Multi-order Correlations Learning Module simultaneously learns both bilateral correlations and hyper-correlations. Lastly, the Information Aggregation Module generates a representation for an individual sensing device by aggregating the bilateral correlations and hyper-correlations it involves in. It also generates the representation for a pair of sensing devices by aggregating the hyper-correlations between the pair and other different individual sensing devices. We also propose a computationally more efficient HyperHAR-Lite framework, a lightweight variant of the HyperHAR framework, at a small cost of accuracy. Both the HyperHAR and HyperHAR-Lite outperform SOTA models across three commonly used benchmark datasets with significant margins. We validate the efficiency and effectiveness of the proposed frameworks through an ablation study and quantitative and qualitative analysis.

Modulation of Aggregation and Immunogenicity of a Protein: Based on the Study of Hen Lysozyme

This study focuses on the modulation of protein aggregation and immunogenicity. As a starting point for investigating long-range interactions within a non-native protein, the effects of perturbing denatured protein states on their aggregation, including the formation of amyloid fibrils, were evaluated. The effects of adducts, sugar modifications, and stabilization on protein aggregation were then examined. We also investigated how protein immunogenicity was affected by enhancing protein conformational stability and other factors.

Dynamic multiple-graph spatial-temporal synchronous aggregation framework for traffic prediction in intelligent transportation systems.

Accurate traffic prediction contributes significantly to the success of intelligent transportation systems (ITS), which enables ITS to rationally deploy road resources and enhance the utilization efficiency of road networks. Improvements in prediction performance are evident by utilizing synchronized rather than stepwise components to model spatial-temporal correlations. Some existing studies have designed graph structures containing spatial and temporal attributes to achieve spatial-temporal synchronous learning. However, two challenges remain due to the intricate dynamics: (a) Accounting for the impact of external factors in spatial-temporal synchronous modeling. (b) Multiple perspectives in constructing spatial-temporal synchronous graphs. To address the mentioned limitations, a novel model named dynamic multiple-graph spatial-temporal synchronous aggregation framework (DMSTSAF) for traffic prediction is proposed. Specifically, DMSTSAF utilizes a feature augmentation module (FAM) to adaptively incorporate traffic data with external factors and generate fused features as inputs to subsequent modules. Moreover, DMSTSAF introduces diverse spatial and temporal graphs according to different spatial-temporal relationships. Based on this, two types of spatial-temporal synchronous graphs and the corresponding synchronous aggregation modules are designed to simultaneously extract hidden features from various aspects. Extensive experiments constructed on four real-world datasets indicate that our model improves by 3.68-8.54% compared to the state-of-the-art baseline.

Integrating Contextual Information and Attention Mechanisms with Sparse Convolution for the Extraction of Internal Objects within Buildings from Three-Dimensional Point Clouds

Deep learning-based point cloud semantic segmentation has gained popularity over time, with sparse convolution being the most prominent example. Although sparse convolution is more efficient than regular convolution, it comes with the drawback of sacrificing global context information. To solve this problem, this paper proposes the OcspareNet network, which uses sparse convolution as the backbone and captures global contextual information using the offset attention module and context aggregation module. The offset attention module improves the network’s capacity to obtain global contextual information about the point cloud. The context aggregation module utilizes contextual information in the training and testing phases, which increases the network’s capacity to discern the overall structure and successfully improves the network’s capacity and the accuracy of the difficult-scene segmentation category. Compared to the state-of-the-art (SOTA) models, our model has a smaller parameter count and achieves higher accuracy on challenging segmentation categories such as ‘pictures’, ‘counters’, and ‘desks’ in the ScanNetV2 dataset, with IoU scores of 41.1%, 70.3%, and 72.5%, respectively. Furthermore, ablation experiments confirmed the efficacy of our designed modules.

Convolutional transformer-driven robust electrocardiogram signal denoising framework with adaptive parametric ReLU.

The electrocardiogram (ECG) is a widely used diagnostic tool for cardiovascular diseases. However, ECG recording is often subject to various noises, which can limit its clinical evaluation. To address this issue, we propose a novel Transformer-based convolutional neural network framework with adaptively parametric ReLU (APtrans-CNN) for ECG signal denoising. The proposed APtrans-CNN architecture combines the strengths of transformers in global feature learning and CNNs in local feature learning to address the inadequacy of learning with long sequence time-series features. By fully exploiting the global features of ECG signals, our framework can effectively extract critical information that is necessary for signal denoising. We also introduce an adaptively parametric ReLU that can assign a value to the negative information contained in the ECG signal, thereby overcoming the limitation of ReLU to retain negative information. Additionally, we introduce a dynamic feature aggregation module that enables automatic learning and retention of valuable features while discarding useless noise information. Results obtained from two datasets demonstrate that our proposed APtrans-CNN can accurately extract pure ECG signals from noisy datasets and is adaptable to various applications. Specifically, when the input consists of ECG signals with a signal-to-noise ratio (SNR) of -4 dB, APtrans-CNN successfully increases the SNR to more than 6 dB, resulting in the diagnostic model's accuracy exceeding 96%.

Attention guided multi-level feature aggregation network for camouflaged object detection

Camouflaged object detection (COD) aims to identify objects that are visually blended into their highly similar surroundings, which is an extremely complex and challenging visual task in real-world scenarios, and has recently attracted increasing research interest in the field of computer vision due to its valuable applications. The existing deep learning based methods of COD have the following problems: 1) the ambiguous boundary of the camouflaged objects in prediction map, 2) an inaccurate detection of the camouflaged object with accurate and complete structure details. To this end, an attention guided multi-level feature aggregation network is proposed for this task, which is based on three key designs. First, by embedding spatial pyramid attention (SPA) in ResNet-like backbone network, better multi-level features are extracted to identify the camouflaged objects with complete internal details. Second, an edge context module is designed to make full use of edge information, which highlights camouflaged object structure and generates accurate edge localization of COD. Third, the feature aggregation module based on local attention is used to fuse these enhanced multi-level features and edge context cues, which consists of two major components: the iterative Attentional Feature Fusion (iAFF) module and the Dual-branch Global Context Module (DGCM). Compared with the existing 16 state-of-the-art methods, extensive experiments on four widely-used benchmark datasets under four authoritative evaluation metrics illustrate that the proposed method is very beneficial to the COD task.

Atom-ProteinQA: Atom-level protein model quality assessment through fine-grained joint learning

MotivationProtein model quality assessment (ProteinQA) is a fundamental task that is essential for biologically relevant applications, i.e., protein structure refinement, protein design, etc. Previous works aimed to conduct ProteinQA only on the global structure or per-residue level, ignoring potentially usable and precise cues from a fine-grained per-atom perspective. In this study, we propose an atom-level ProteinQA model, named Atom-ProteinQA, in which two innovative modules are designed to extract geometric and topological atom-level relationships respectively. Specifically, on the one hand, a geometric perception module exploits 3D sparse convolution to capture the geometric features of the input protein, generating fine-grained atom-level predictions. On the other hand, natural chemical bonds are utilized to construct an atom-level graph, then message passing from a topological perception module is applied to output residue-level predictions in parallel. Eventually, through a cross-model aggregation module, features from different modules mutually interact, enhancing performance on both the atom and residue levels. ResultsExtensive experiments show that our proposed Atom-ProteinQA outperforms previous methods by a large margin, regardless of residue-level or atom-level assessment. Concretely, we achieved state-of-the-art performance on CATH-2084, Decoy-8000, public benchmarks CASP13 & CASP14, and the CAMEO. AvailabilityThe repository of this project is released on: https://github.com/luyfcandy/Atom_ProteinQA.

HFENet: Hybrid feature encoder network for detecting salient objects in RGB-thermal images

Deep convolutional neural networks (CNNs) have gained prominence in computer vision applications, including RGB salient object detection (SOD), owing to the advancements in deep learning. Nevertheless, the majority of deep CNNs employ either VGGNet or ResNet as their backbone architecture for extracting image information. This approach may lead to the following problems. 1) Variations between imaging modalities during feature extraction across layers. Cross-modal features across layers are often fused in a single step, resulting in inadequate cross-modal feature extraction. 2) Feature long-range dependence problem in multilayer feature decoding. 3) Image boundary blurring. To address these issues, we initially leverage the advantages offered by the VGGNet and ResNet architectures. Additionally, we present a novel hybrid VGG–ResNet feature encoder for RGB-T SOD. Specifically, we introduce a geometry information aggregation module that effectively combines and enhances the VGGNet spatial features of the RGB-T modalities from the bottom to the top. Moreover, we propose a innovative global saliency perception module that progressively refines the ResNet semantic features from the top to the bottom by integrating both local and global information. Furthermore, we introduce a Pearson-gated module to tackle the challenge of long-range dependence between features. This module utilizes gating to merge features by calculating the Pearson correlation coefficients of the fused features at multiple levels. Lastly, we devise an edge-aware module to precisely learn the contours of salient objects, thereby enhancing the clarity of the object boundaries. Extensive experiments conducted on three RGB-T SOD benchmarks demonstrate that our proposed network surpasses the performance of state-of-the-art methods for SOD.

Motion-guided and occlusion-aware multi-object tracking with hierarchical matching

In the field of multi-target tracking, the widely embraced tracking-by-detection paradigm has rapidly progressed with the refinement of detectors and matching techniques. However, the paradigm of joint detection and tracking is relatively limited, and it is difficult to model complex scenes, such as the complexities introduced by camera motion and occlusion. In this work, a hierarchical joint detection and tracking framework is proposed, namely MSPNet. From a temporal concern, a motion-guided feature aggregation module is proposed to address the complexities of multi-frame variations. From a spatial concern, an occlusion-aware head and hierarchical spatial association are proposed to handle the challenges of occlusion. Extensive experiments on MOT challenging benchmarks demonstrate that the MSPNet can effectively reduce false negatives and improve the accuracy of tracking while outperforming a wide range of existing methods.

ALIKE-APPLE: A Lightweight Method for the Detection and Description of Minute and Similar Feature Points in Apples

Current image feature extraction methods fail to adapt to the fine features of apple image texture, resulting in image matching errors and degraded image processing accuracy. A multi-view orthogonal image acquisition system was constructed with apples as the research object. The system consists of four industrial cameras placed around the apple at different angles and one camera placed on top. Following the image acquisition through the system, synthetic image pairs—both before and after transformation—were generated as the input dataset. This generation process involved each image being subjected to random transformations. Through learning to extract more distinctive and descriptive features, the deep learning-based keypoint detection method surpasses traditional techniques by broadening the application range and enhancing detection accuracy. Therefore, a lightweight network called ALIKE-APPLE was proposed for surface feature point detection. The baseline model for ALIKE-APPLE is ALIKE, upon which improvements have been made to the image feature encoder and feature aggregation modules. It comprises an Improved Convolutional Attention Module (ICBAM) and a Boosting Resolution Sampling Module (BRSM). The proposed ICBAM replaced max pooling in the original image feature encoder for downsampling. It enhanced the feature fusion capability of the model by utilizing spatial contextual information and learning region associations in the image. The proposed BRSM replaced the bilinear interpolation in the original feature aggregator for upsampling, overcoming the apple side image’s geometric distortion and effectively preserving the texture details and edge information. The model size was shrunk by optimizing the number of downsampling operations from the image encoder of the original model. The experimental results showed that the average number of observed keypoints and the average matching accuracy were improved by 166.41% and 37.07%, respectively, compared with the baseline model. The feature detection model of ALIKE-APPLE was found to perform better than the optimal SuperPoint. The feature point distribution of ALIKE-APPLE showed an improvement of 10.29% in average standard deviation (Std), 8.62% in average coefficient of variation (CV), and 156.12% in average feature point density (AFPD). Moreover, the mean matching accuracy (MMA) of ALIKE-APPLE improved by 125.97%. Thus, ALIKE-APPLE boasts a more consistent allocation of feature points and greater precision in matching.

Contactless weighing method based on deep learning and acoustic levitation

Abstract Acoustic weighing is a promising contactless method for screening the mass of micro-nano objects as it avoids contact contamination and losses. Existing acoustic weighing methods determine the mass of an object by detecting its oscillation trajectory with a laser sensor. However, this method suffers from several limitations, such as short measurement distance, poor accuracy in measuring transparent objects, and inducing damage to photosensitive samples. To solve these issues, this work proposes a contactless weighing method based on location-aware neural network (LANet) and acoustic levitation. The proposed LANet is a deep learning-based image processing method that detects object bit oscillation trajectories completely contactless, regardless of the color, shape, and oscillation distance of the levitated object. We employ a cross-stage aggregation module and cross-mixed feature pyramid strategy to build LANet network depth for enhanced feature extraction. In addition, to create a contactless environment, we built an acoustic levitation system, which drives the oscillation of objects. Finally, we verified the accuracy and effectiveness of the method. The results show that the proposed network can accurately detect the oscillation trajectories of various objects with high detection performance, even for small objects in low-contrast backgrounds. Meanwhile, the proposed method can accurately measure the mass of objects with a percentage error of no more than 7.83%.

EMF-Net: An edge-guided multi-feature fusion network for text manipulation detection

Text images have a more uniform background than natural images, making it difficult to capture tampering traces. Thus, it poses significant challenges to the detection and localization of tampered text images. To effectively detect tampered text images, we propose the edge-guided multi-feature fusion network (EMF-Net) including four different modules, the difference semantic discriminator (DSD), the edge-guidance feature aggregation module (EFAM), the edge supervised module (ESM), and the multi-branch attention-induced fusion module (MAFM). The proposed network combines RGB images and learned noise-sensitive features to deeply excavate the hidden tampered region features to improve the detection accuracy due to the insufficient feature information. Meanwhile, the multi-branch attention-induced fusion and semantic discriminator methods are integrated to reduce false alarms by shielding the feature interference in non-tampered regions and diminish the external interference. Furthermore, we craft a dataset of 12,000 text images and their tempered versions with three tempering operations including copy-move, splicing and inpainting. Extensive experiments have shown that the proposed network can improve the generalization performance and achieve the higher detection accuracy compared to current other state-of-the-art methods.

FCPFNet: Feature Complementation Network with Pyramid Fusion for Semantic Segmentation

Traditional pyramid pooling modules have shown effective improvements in semantic segmentation tasks by capturing multi-scale feature information. However, their limitations arise from the shallow structure, which fails to fully extract contextual information, and the fused multi-scale feature information lacks distinctiveness, resulting in issues with the final segmentation discriminability. To address these issues, we proposes an effective solution called FCPFNet, which is based on global contextual prior for deep feature extraction of detailed information. Specifically, we introduce a novel deep feature aggregation module to extract semantic information from the output feature map of each layer through a deep aggregation of context information module, and expands the effective perception range. Additionally, we propose an Efficient Pyramid Pooling Module (EPPM) to capture distinctive features through communicating information between different sub-features and performs multi-scale fusion, which is integrated as a branch within the network to complement the information loss resulting from downsampling operations. Furthermore, in order to ensure the richness of image detail feature information and maintain a large receptive field to obtain more contextual information, EPPM concatenates the input feature map and the output feature map of the pyramid pooling module to acquire more comprehensive global contextual information. It has been demonstrated by experiment that the method described in this article achieves competitive performance on the challenging scene segmentation datasets Pascal VOC 2012, Cityscapes and Coco-Stuff, with MIOU of 81.0%, 78.8% and 40.1%, respectively.

Multi-Dimensional Information Fusion You Only Look Once Network for Suspicious Object Detection in Millimeter Wave Images

Millimeter wave (MMW) imaging systems have been widely used for security screening in public places due to their advantages of being able to detect a variety of suspicious objects, non-contact operation, and harmlessness to the human body. In this study, we propose an innovative, multi-dimensional information fusion YOLO network that can aggregate and capture multimodal information to cope with the challenges of low resolution and susceptibility to noise in MMW images. In particular, an MMW data information aggregation module is developed to adaptively synthesize a novel type of MMW image, which simultaneously contains pixel, depth, phase, and diverse signal-to-noise information to overcome the limitations of current MMW images containing consistent pixel information in all three channels. Furthermore, this module is capable of differentiable data enhancements to take into account adverse noise conditions in real application scenarios. In order to fully acquire the augmented contextual information mentioned above, we propose an asymptotic path aggregation network and combine it with YOLOv8. The proposed method is able to adaptively and bidirectionally fuse deep and shallow features while avoiding semantic gaps. In addition, a multi-view, multi-parameter mapping technique is designed to enhance the detection ability. The experiments on the measured MMW datasets validate the improvement in object detection using the proposed model.

A Weakly Supervised Hybrid Lightweight Network for Efficient Crowd Counting

Crowd-counting networks have become the mainstream method to deploy crowd-counting techniques on resource-constrained devices. Significant progress has been made in this field, with many outstanding lightweight models being proposed successively. However, challenges like scare-variation, global feature extraction, and fine-grained head annotation requirements still exist in relevant tasks, necessitating further improvement. In this article, we propose a weakly supervised hybrid lightweight crowd-counting network that integrates the initial layers of GhostNet as the backbone to efficiently extract local features and enrich intermediate features. The incorporation of a modified Swin-Transformer block addresses the need for effective global context information. A Pyramid Pooling Aggregation Module handles the inherent scale variation problem in crowd-counting tasks in a more computation-efficient way. This module, along with the cross-attention module, serves as bridges to promote the feature information flow between local features and global context information. Finally, a simplified regressor module is designed to allow the proposed model with weakly supervised guidance for training to avoid precise location-level annotations, noting that the omission of density map generation makes the proposed network more lightweight. Our results on the UCF-QNRF dataset indicate our model is 8.73% and 12.17% more accurate on MAE and MSE metrics, respectively, than the second-best ARNet, in which the parameters decrease by 4.52%. On the ShanghaiTech A dataset, MAE and MSE drop 1.5% and 3.2%, respectively, compared to the second-best PDDNet. The experimental results for accuracy and inference speed evaluation on some mainstream datasets validate the effective design principle of our model.

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.