Feature Fusion Module Research Articles

Disentangled feature fusion network for lightweight image super-resolution

Recently, the quality of generated images in image super-resolution (SR) has significantly improved due to the widespread application of convolutional neural networks. Existing super-resolution methods often overlook the importance of network width and instead prioritize designing deeper network structures to improve performance. Deeper networks, however, demand greater computational resources and are more challenging to train, making them less suitable for devices with limited performance. To address this issue, we propose a novel method called the disentangled feature fusion network (DFFN). Specifically, we construct dual feature extraction streams (DFES) using the strategies of feature disentanglement and parameter sharing, which increase the network's width without increasing the number of parameters. To facilitate the interaction of information between the dual feature extraction streams, we design a feature interaction module (FIM) that separately enhances high and low-frequency features. Furthermore, a feature fusion module (FFM) is presented to efficiently fuse different levels of high and low-frequency feature information. The proposed DFFN integrates DFES, FIM, and FFM, which not only widens the network without increasing parameters but also minimizes the loss of crucial feature information. Extensive experiments on five benchmark datasets demonstrate that our method significantly outperforms other state-of-the-art lightweight super-resolution methods, achieving a superior balance between parameter quantity, reconstruction performance, and model complexity. The code is available at https://github.com/zhoujy-aust/DFFN.

UAV Geo-Localization Dataset and Method Based on Cross-View Matching

The stable flight of drones relies on Global Navigation Satellite Systems (GNSS). However, in complex environments, GNSS signals are prone to interference, leading to flight instability. Inspired by cross-view machine learning, this paper introduces the VDUAV dataset and designs the VRLM network architecture, opening new avenues for cross-view geolocation. First, to address the limitations of traditional datasets with limited scenarios, we propose the VDUAV dataset. By leveraging the virtual–real mapping of latitude and longitude coordinates, we establish a digital twin platform that incorporates 3D models of real-world environments. This platform facilitates the creation of the VDUAV dataset for cross-view drone localization, significantly reducing the cost of dataset production. Second, we introduce a new baseline model for cross-view matching, the Virtual Reality Localization Method (VRLM). The model uses FocalNet as its backbone and extracts multi-scale features from both drone and satellite images through two separate branches. These features are then fused using a Similarity Computation and Feature Fusion (SCFF) module. By applying a weighted fusion of multi-scale features, the model preserves critical distinguishing features in the images, leading to substantial improvements in both processing speed and localization accuracy. Experimental results demonstrate that the VRLM model outperforms FPI on the VDUAV dataset, achieving an accuracy increase to 83.35% on the MA@20 metric and a precision of 74.13% on the RDS metric.

DecoupleCLIP: A Novel Cross-Modality Decouple Model for Painting Captioning

Image captioning aims to describe the content in an image, which plays a critical role in image understanding. Existing methods tend to generate the text for more distinct natural images. These models can not be well for paintings containing more abstract meaning due to the limitation of objective parsing without related knowledge. To alleviate, we propose a novel cross-modality decouple model to generate the objective and subjective parsing separately. Concretely, we propose to encode both subjective semantic and implied knowledge contained in the paintings. The key point of our framework is decoupled CLIP-based branches (DecoupleCLIP). For the objective caption branch, we utilize the CLIP model as the global feature extractor and construct a feature fusion module for global clues. Based on the objective caption branch structure, we add a multimodal fusion module called the artistic conception branch. In this way, the objective captions can constrain artistic conception content. We conduct extensive experiments to demonstrate our DecoupleCLIP’s superior ability over our new dataset. Our model achieves nearly 2% improvement over other comparison models on CIDEr.

Enhancing Multimodal Emotional Information Extraction in Film and Television through Adaptive Feature Fusion with DenseNe, Transformer, and 3D CNN Models

ABSTRACT The extraction of multimodal emotional information enables a more nuanced representation of the emotional subtleties embedded in film and television works. However, conventional approaches that independently extract features from images and text fail to capture the rich semantic interplay between these modalities, impeding the feature learning process within each modality. To address this, this paper presents an innovative model for extracting emotional information from multimodal film and television content. The model utilizes DenseNe for image feature extraction, enhancing network depth via the MSC module. Text feature extraction is achieved through a Transformer encoder, while video feature extraction employs a 3D CNN model. Refinements are made to the number and placement of convolutional layers, planar convolution size, and 3D convolution depth. Moreover, a multi-head scaled dot-product attention mechanism is incorporated into the interaction module to compute the similarity between each image block within the sequence and every word in the text sequence. Experimental evaluations on the CMU-MOSEI and CMU-MOSI datasets showcase superior performance compared to the baseline model, achieving accuracy and F1 scores of 0.708 and 0.698, respectively. Noteworthy is the proposed adaptive feature fusion module, which enriches the expressiveness of pivotal emotional features while eliminating redundant data.

Establishing effective learning bridge cross multi-scale feature maps for object detection and semantic segmentation

ABSTRACT In the field of image object detection and semantic segmentation, improving the accuracy of object identification and segmentation is a primary goal. To achieve this, leveraging the potential of multi-scale information through feature map refinement and fusion has been widely recognized. However, existing feature fusion methods either design more complex feature pyramid networks, replace existing detectors, or incrementally introduce feature fusion modules, overlooking the effective approach of enhancing spatial information in deep feature maps. We propose a novel pluggable feature fusion paradigm termed ‘Effective Learning Bridge’. Our research introduces an efficient and adaptive learning mechanism that builds learning bridges between feature maps at different scales within the feature pyramid, thereby enhancing the spatial information of objects in deep feature maps. This mechanism is specifically designed for multi-scale feature maps and can be seamlessly integrated into any network incorporating feature maps. By altering the model’s backpropagation path, we successfully improve learning efficiency, which in turn enhances the accuracy of object detection and segmentation. Our proposed paradigm and method were extensively evaluated through experiments on SIMD, HRSID, and WHDLD datasets and benchmark models. The results unequivocally demonstrate the effectiveness of our approach in significantly improving the accuracy of object detection and semantic segmentation, as well as the overall learning efficiency of the model.

DMDLK-Net: A dynamic multi-scale feature fusion network with deformable large kernel for medical segmentation

Abstract. Skin disease image segmentation is a crucial component of computer-aided diagnosis, providing precise localization and delineation of lesions that enhance diagnostic accuracy and efficiency. Despite significant advancements in convolutional neural networks (CNNs), there remains substantial room for improvement in segmentation performance due to the diverse and complex nature of skin lesions. In this study, we propose DMDLK-Net, a dynamic multi-scale feature fusion network with deformable large kernels, specifically designed to address the challenges in skin disease segmentation. Our network incorporates a Dynamic Deformable Large Kernel (DDLK) module and a Dynamic Multi-Scale Feature Fusion (DMFF) module, enhancing the model's ability to capture intricate lesion features. We present the performance of DMDLK-Net on the ISIC-2018 dataset, highlighting its promising results. Key contributions of this work include the innovative use of deformable large kernels for adaptive feature extraction and the introduction of dynamic multi-scale fusion to balance local and global information. Our experimental results confirm the effectiveness of DMDLK-Net in delivering high-precision segmentation, thus providing a reliable tool for clinical applications.

A Multi-Hierarchical Complementary Feature Interaction Network for Accelerated Multi-Modal MR Imaging

Magnetic resonance (MR) imaging is widely used in the clinical field due to its non-invasiveness, but the long scanning time is still a bottleneck for its popularization. Using the complementary information between multi-modal imaging to accelerate imaging provides a novel and effective MR fast imaging solution. However, previous technologies mostly use simple fusion methods and fail to fully utilize their potential sharable knowledge. In this study, we introduced a novel multi-hierarchical complementary feature interaction network (MHCFIN) to realize joint reconstruction of multi-modal MR images with undersampled data and thus accelerate multi-modal imaging. Firstly, multiple attention mechanisms are integrated with a dual-branch encoder–decoder network to represent shared features and complementary features of different modalities. In the decoding stage, the multi-modal feature interaction module (MMFIM) acts as a bridge between the two branches, realizing complementary knowledge transfer between different modalities through cross-level fusion. The single-modal feature fusion module (SMFFM) carries out multi-scale feature representation and optimization of the single modality, preserving better anatomical details. Extensive experiments are conducted under different sampling patterns and acceleration factors. The results show that this proposed method achieves obvious improvement compared with existing state-of-the-art reconstruction methods in both visual quality and quantity.

GLE-net: global-local information enhancement for semantic segmentation of remote sensing images

Remote sensing (RS) images contain a wealth of information with expansive potential for applications in image segmentation. However, Convolutional Neural Networks (CNN) face challenges in fully harnessing the global contextual information. Leveraging the formidable capabilities of global information modeling with Swin-Transformer, a novel RS images segmentation model with CNN (GLE-Net) was introduced. This integration gives rise to a revamped encoder structure. The subbranch initiates the process by extracting features at varying scales within the RS images using the Multiscale Feature Fusion Module (MFM), acquiring rich semantic information, discerning localized finer features, and adeptly handling occlusions. Subsequently, Feature Compression Module (FCM) is introduced in main branch to downsize the feature map, effectively reducing information loss while preserving finer details, enhancing segmentation accuracy for smaller targets. Finally, we integrate local features and global features through Spatial Information Enhancement Module (SIEM) for comprehensive feature modeling, augmenting the segmentation capabilities of model. We performed experiments on public datasets provided by ISPRS, yielding notably remarkable experimental outcomes. This underscores the substantial potential of our model in the realm of RS image segmentation within the context of scientific research.

Adversarial Unsupervised Domain Adaptation for 3D Semantic Segmentation with 2D Image Fusion of Dense Depth

AbstractUnsupervised domain adaptation (UDA) is increasingly used for 3D point cloud semantic segmentation tasks due to its ability to address the issue of missing labels for new domains. However, most existing unsupervised domain adaptation methods focus only on uni‐modal data and are rarely applied to multi‐modal data. Therefore, we propose a cross‐modal UDA on multi‐modal datasets that contain 3D point clouds and 2D images for 3D Semantic Segmentation. Specifically, we first propose a Dual discriminator‐based Domain Adaptation (Dd‐bDA) module to enhance the adaptability of different domains. Second, given that the robustness of depth information to domain shifts can provide more details for semantic segmentation, we further employ a Dense depth Feature Fusion (DdFF) module to extract image features with rich depth cues. We evaluate our model in four unsupervised domain adaptation scenarios, i.e., dataset‐to‐dataset (A2D2 → SemanticKITTI), Day‐to‐Night, country‐to‐country (USA → Singapore), and synthetic‐to‐real (VirtualKITTI → SemanticKITTI). In all settings, the experimental results achieve significant improvements and surpass state‐of‐the‐art models.

Global induced local network for infrared: dim small target detection

Abstract It is challenging to detect infrared dim targets submerged in complicated backgrounds due to their small size and faint intensity. The previous attention-based detection networks frequently require global long-range dependence. Significant calculations are required to determine the target's sparse but meaningful position. To prevent wasting calculations on the background, this paper offers a detection network guided by global context for local feature learning, named GILNet (Global Induced Local Network). It designs a global location module (GLM) and a local feature interaction module (LFIM) to capture the global position and features of targets, respectively. More specifically, using global context interaction, the GLM finds the region that might contain dim small targets, that is, the coarse location. In the coarsely located regions, the LFIM further acquires feature information about targets. Next, we also design an eight-directional attention operation to obtain the contour information of targets in the low feature map. It is fused with the high feature map in the multi-directional feature fusion module (MFFM), which retains more semantic and spatial information about targets. Finally, quantitative and qualitative analysis show that the GILNet performs better than eight comparison methods on two public datasets.

MFE-SSNet: Multi-Modal Fusion-Based End-to-End Steering Angle and Vehicle Speed Prediction Network

AbstractIn the field of autonomous vehicles, accurately predicting steering angle and speed is a pivotal task. This task affects the accuracy of the final decision of the autonomous vehicle and is the basis for ensuring the safe and efficient operation of the autonomous vehicle. Previous studies have often relied on data from only one or two modalities to make predictions for steering angle and vehicle speed, which were often inadequate. In this paper, the authors propose a Multi-Modal Fusion-Based End-to-End Steering Angle and Vehicle Speed Prediction Network (MFE-SSNet). The network innovatively extends the one-stream and two-stream structure to a three-stream structure and cleverly extracts features of images, steering angles, and vehicle speeds using HRNet and LSTM layers. In addition, in order to fully fuse the feature information of different modal data, this paper also proposes a local attention-based feature fusion module. This module improves the fusion of different modal feature vectors by capturing the interdependencies in the local channels. Experimental results demonstrate that MFE-SSNet outperforms the current state-of-the-art model on the publicly available Udacity dataset.

AEFF-SSC: an attention-enhanced feature fusion for 3D semantic scene completion

Abstract Three-dimensional (3D) occupancy perception technology aims to enable autonomous vehicles to observe and understand dense 3D environments. Estimating the complete geometry and semantics of a scene solely from visual images is challenging. However, humans can easily conceive the complete form of objects based on partial key information and their own experience. This ability is crucial for recognizing and interpreting the surrounding environment. To equip 3D occupancy perception systems with a similar capability, a 3D semantic scene completion method called AEFF-SSC is proposed. This method deeply explores boundary and multi-scale information in voxels, aiming to reconstruct 3D geometry more accurately. We have specifically designed an attention-enhanced feature fusion module that effectively fuses image feature information from different scales and focuses on feature boundary information, thereby more efficiently extracting voxel features. Additionally, we introduce a semantic segmentation module driven by a 3D attention-UNet network. This module combines a 3D U-Net network with a 3D attention mechanism. Through feature fusion and feature weighting, it aids in restoring 3D spatial information and significantly improves the accuracy of segmentation results. Experimental verification on the SemanticKITTI dataset demonstrates that AEFF-SSC significantly outperforms other existing methods in terms of both geometry and semantics. Specifically, within a 12.8 m × 12.8 m area ahead, our geometric occupancy accuracy has achieved a significant improvement of 71.58%, and at the same time, the semantic segmentation accuracy has also increased remarkably by 54.20%.

TransImg: A Translation Algorithm of Visible-to-Infrared Image Based on Generative Adversarial Network

Infrared images of sensitive targets are difficult to obtain and cannot meet the design and training needs of target detection and tracking algorithms for mobile platforms such as aircraft. This paper proposes an image translation algorithm TransImg, which can achieve visible light image translation to the infrared domain to enrich the dataset. First, the algorithm designed a generator structure consisting of a deep residual connected encoder and a region perception feature fusion module to enhance feature learning, thereby avoiding issues such as generating infrared images with insufficient details in the transfer task. Afterward, a multi-scale discriminator and a composite loss function were designed to further improve the transfer effect. Finally, an automatic mixed-precision training strategy was designed for the overall migration algorithm architecture to accelerate the training and generation of infrared images. Experiments have shown that the image translation algorithm TransImg has good algorithm accuracy, and the infrared image generated by visible light image translation has richer texture details, faster generation speed, and lower video memory consumption, and the performance exceeds the mainstream traditional algorithm, and the generated images can meet the requirements of target detection and tracking algorithms design and training for mobile platforms such as aircraft.

GCS-YOLOv8: A Lightweight Face Extractor to Assist Deepfake Detection

To address the issues of target feature blurring and increased false detections caused by high compression rates in deepfake videos, as well as the high computational resource requirements of existing face extractors, we propose a lightweight face extractor to assist deepfake detection, GCS-YOLOv8. Firstly, we employ the HGStem module for initial downsampling to address the issue of false detections of small non-face objects in deepfake videos, thereby improving detection accuracy. Secondly, we introduce the C2f-GDConv module to mitigate the low-FLOPs pitfall while reducing the model’s parameters, thereby lightening the network. Additionally, we add a new P6 large target detection layer to expand the receptive field and capture multi-scale features, solving the problem of detecting large-scale faces in low-compression deepfake videos. We also design a cross-scale feature fusion module called CCFG (CNN-based Cross-Scale Feature Fusion with GDConv), which integrates features from different scales to enhance the model’s adaptability to scale variations while reducing network parameters, addressing the high computational resource requirements of traditional face extractors. Furthermore, we improve the detection head by utilizing group normalization and shared convolution, simplifying the process of face detection while maintaining detection performance. The training dataset was also refined by removing low-accuracy and low-resolution labels, which reduced the false detection rate. Experimental results demonstrate that, compared to YOLOv8, this face extractor achieves the AP of 0.942, 0.927, and 0.812 on the WiderFace dataset’s Easy, Medium, and Hard subsets, representing improvements of 1.1%, 1.3%, and 3.7% respectively. The model’s parameters and FLOPs are only 1.68 MB and 3.5 G, reflecting reductions of 44.2% and 56.8%, making it more effective and lightweight in extracting faces from deepfake videos.

A multi-scale feature extraction and fusion-based model for retinal vessel segmentation in fundus images.

In response to the challenge of low accuracy in retinal vessel segmentation attributed to the minute nature of the vessels, this paper proposes a retinal vessel segmentation model based on an improved U-Net, which combines multi-scale feature extraction and fusion techniques. An improved dilated residual module was first used to replace the original convolutional layer of U-Net, and this module, coupled with a dual attention mechanism and diverse expansion rates, facilitates the extraction of multi-scale vascular features. Moreover, an adaptive feature fusion module was added at the skip connections of the model to improve vessel connectivity. To further optimize network training, a hybrid loss function is employed to mitigate the class imbalance between vessels and the background. Experimental results on the DRIVE dataset and CHASE_DB1 dataset show that the proposed model has an accuracy of 96.27% and 96.96%, sensitivity of 81.32% and 82.59%, and AUC of 98.34% and 98.70%, respectively, demonstrating superior segmentation performance.

A Novel Detection Transformer Framework for Ship Detection in Synthetic Aperture Radar Imagery Using Advanced Feature Fusion and Polarimetric Techniques

Ship detection in synthetic aperture radar (SAR) imagery faces significant challenges due to the limitations of traditional methods, such as convolutional neural network (CNN) and anchor-based matching approaches, which struggle with accurately detecting smaller targets as well as adapting to varying environmental conditions. These methods, relying on either intensity values or single-target characteristics, often fail to enhance the signal-to-clutter ratio (SCR) and are prone to false detections due to environmental factors. To address these issues, a novel framework is introduced that leverages the detection transformer (DETR) model along with advanced feature fusion techniques to enhance ship detection. This feature enhancement DETR (FEDETR) module manages clutter and improves feature extraction through preprocessing techniques such as filtering, denoising, and applying maximum and median pooling with various kernel sizes. Furthermore, it combines metrics like the line spread function (LSF), peak signal-to-noise ratio (PSNR), and F1 score to predict optimal pooling configurations and thus enhance edge sharpness, image fidelity, and detection accuracy. Complementing this, the weighted feature fusion (WFF) module integrates polarimetric SAR (PolSAR) methods such as Pauli decomposition, coherence matrix analysis, and feature volume and helix scattering (Fvh) components decomposition, along with FEDETR attention maps, to provide detailed radar scattering insights that enhance ship response characterization. Finally, by integrating wave polarization properties, the ability to distinguish and characterize targets is augmented, thereby improving SCR and facilitating the detection of weakly scattered targets in SAR imagery. Overall, this new framework significantly boosts DETR’s performance, offering a robust solution for maritime surveillance and security.

Using lightweight denoising network to suppress multiple barrage jamming in range‐Doppler domain

AbstractBarrage jamming technology poses a significant threat to radar detection functions, yet there is limited research on anti‐multiple barrage jamming using single‐channel radar systems. Addressing this gap, this paper proposes a framework for anti‐multiple barrage jamming based on range‐Doppler domain denoising. This framework first performs pulse compression and Doppler processing on the echo signal to enhance the signal‐to‐jamming ratio. The range‐Doppler spectrum data is then input into a deep denoising network to suppress multiple types of jamming. Additionally, the paper proposes a lightweight deep denoising network comprising a feature extraction module, a jamming suppression module, and a feature fusion module. The feature extraction module preliminarily extracts jamming features using multiple layers of depth‐wise and point‐wise convolution. The jamming suppression module removes noise at various resolutions through a lightweight encoding and decoding structure, effectively suppressing the barrage jamming signal. The feature fusion module uses convolution kernels with different sizes to merge the multiple features output by the jamming suppression module. Simulation results demonstrate that the proposed method effectively suppresses ten types of barrage jamming. Furthermore, a measured dataset is constructed to verify the method’s effectiveness.

Detection Method for Bolt Missing Pins of Transmission Line Based on Improved SSD Network

The bolt pin structure plays a role in connecting and fixing various components in transmission lines, and the pins are prone to detachment, which may lead to the disintegration of key components and cause serious consequences. The bolt pin target size in the transmission line inspection image is small, the background is complex, and the detection effect using traditional object detection algorithms is poor. This article proposes a method for detecting missing bolts in transmission lines based on an improved SSD (Single Shot MultiBox Detector) network. Using the traditional SSD network as the foundation, a feature fusion module is first added to the network to enrich the semantic information of low-level feature maps and the detail information of high-level feature maps through the mutual fusion of high-level and low-level feature maps. Secondly, embedding a channel attention mechanism module enhances the channel features corresponding to the bolt pin structure and suppresses channel features that are irrelevant to the background; At the same time, the introduction of Focal Loss loss function reduces the unevenness of positive and negative sample classification. Finally, experiments were conducted on the collected bolt missing pin detection dataset, and the results showed that the method proposed in this paper has good performance in detecting missing pins, and the detection accuracy has been significantly improved.

FG-UNet: fine-grained feature-guided UNet for segmentation of weeds and crops in UAV images.

Semantic segmentation of weed and crop images is a key component and prerequisite for automated weed management. For weeds in unmanned aerial vehicle (UAV) images, which are usually characterized by small size and easily confused with crops at early growth stages, existing semantic segmentation models have difficulties to extract sufficiently fine features. This leads to their limited performance in weed and crop segmentation of UAV images. We proposed a fine-grained feature-guided UNet, named FG-UNet, for weed and crop segmentation in UAV images. Specifically, there are two branches in FG-UNet, namely the fine-grained feature branch and the UNet branch. In the fine-grained feature branch, a fine feature-aware (FFA) module was designed to mine fine features in order to enhance the model's ability to segment small objects. In the UNet branch, we used an encoder-decoder structure to realize high-level semantic feature extraction in images. In addition, a contextual feature fusion (CFF) module was designed for the fusion of the fine features and high-level semantic features, thus enhancing the feature discrimination capability of the model. The experimental results showed that our proposed FG-UNet, achieved state-of-the-art performance compared to other semantic segmentation models, with mean intersection over union (MIOU) and mean pixel accuracy (MPA) of 88.06% and 92.37%, respectively. The proposed method in this study lays a solid foundation for accurate detection and intelligent management of weeds. It will have a positive impact on the development of smart agriculture. © 2024 Society of Chemical Industry.

A Depth Awareness and Learnable Feature Fusion Network for Enhanced Geometric Perception in Semantic Correspondence.

Deep learning is becoming the most widely used technology for multi-sensor data fusion. Semantic correspondence has recently emerged as a foundational task, enabling a range of downstream applications, such as style or appearance transfer, robot manipulation, and pose estimation, through its ability to provide robust correspondence in RGB images with semantic information. However, current representations generated by self-supervised learning and generative models are often limited in their ability to capture and understand the geometric structure of objects, which is significant for matching the correct details in applications of semantic correspondence. Furthermore, efficiently fusing these two types of features presents an interesting challenge. Achieving harmonious integration of these features is crucial for improving the expressive power of models in various tasks. To tackle these issues, our key idea is to integrate depth information from depth estimation or depth sensors into feature maps and leverage learnable weights for feature fusion. First, depth information is used to model pixel-wise depth distributions, assigning relative depth weights to feature maps for perceiving an object's structural information. Then, based on a contrastive learning optimization objective, a series of weights are optimized to leverage feature maps from self-supervised learning and generative models. Depth features are naturally embedded into feature maps, guiding the network to learn geometric structure information about objects and alleviating depth ambiguity issues. Experiments on the SPair-71K and AP-10K datasets show that the proposed method achieves scores of 81.8 and 83.3 on the percentage of correct keypoints (PCK) at the 0.1 level, respectively. Our approach not only demonstrates significant advantages in experimental results but also introduces the depth awareness module and a learnable feature fusion module, which enhances the understanding of object structures through depth information and fully utilizes features from various pre-trained models, offering new possibilities for the application of deep learning in RGB and depth data fusion technologies. We will also continue to focus on accelerating model inference and optimizing model lightweighting, enabling our model to operate at a faster speed.