Dense Attention Research Articles

SRARDA: A lightweight adaptive residual dense attention generative adversarial network for image super-resolution

Image super-resolution (SR) is the task of inferring a high resolution (HR) image from one/multiple single low resolution (LR) input(s). Traditional networks are evaluated by pixel-level metrics such as Peak-Signal-to-Noise Ratio (PSNR) etc., which do not always align with human perception of image quality. They often produce excessively smooth images that lack high-frequency texture and appear unnatural. Therefore, in this paper, we propose a lightweight adaptive residual dense attention generative adversarial network (SRARDA) for image SR. Firstly, our generator adopts the residual in residual (RIR) structure but redesigns the basic module. By using dynamic residual connection (ARC) to dynamically adjust the importance of residual and main paths, we design a novel adaptive residual dense attention block (ARDAB) that enhances the feature extraction capability of the generator. In addition, we build a high-frequency filtering unit (HFU) to extract more high-frequency features from the LR space. Finally, to fully utilize the discriminator, we use WGAN to compute the difference between the HR image and the reconstructed image. Experiments demonstrate that SRARDA effectively addresses the issue of excessive smoothing in reconstructed images, while also enhancing visual quality.

Dense object detection methods in RAW UAV imagery based on YOLOv8

Accurate, fast and lightweight dense target detection methods are highly important for precision agriculture. To detect dense apricot flowers using drones, we propose an improved dense target detection method based on YOLOv8, named D-YOLOv8. First, we introduce the Dense Feature Pyramid Networks (D-FPN) to enhance the model’s ability to extract dense features and Dense Attention Layer (DAL) to focus on dense target areas, which enhances the feature extraction ability of dense areas, suppresses features in irrelevant areas, and improves dense target detection accuracy. Finally, RAW data are used to enhance the dataset, which introduces additional original data into RAW images, further enriching the feature input of dense objects. We perform validation on the CARPK challenge dataset and constructed a dataset. The experimental results show that our proposed D-YOLOv8m achieved 98.37% AP, while the model parameters were only 13.2 million. The improved network can effectively support any task of dense target detection.

A Self-Supervised Few-Shot Semantic Segmentation Method Based on Multi-Task Learning and Dense Attention Computation.

Nowadays, autonomous driving technology has become widely prevalent. The intelligent vehicles have been equipped with various sensors (e.g., vision sensors, LiDAR, depth cameras etc.). Among them, the vision systems with tailored semantic segmentation and perception algorithms play critical roles in scene understanding. However, the traditional supervised semantic segmentation needs a large number of pixel-level manual annotations to complete model training. Although few-shot methods reduce the annotation work to some extent, they are still labor intensive. In this paper, a self-supervised few-shot semantic segmentation method based on Multi-task Learning and Dense Attention Computation (dubbed MLDAC) is proposed. The salient part of an image is split into two parts; one of them serves as the support mask for few-shot segmentation, while cross-entropy losses are calculated between the other part and the entire region with the predicted results separately as multi-task learning so as to improve the model's generalization ability. Swin Transformer is used as our backbone to extract feature maps at different scales. These feature maps are then input to multiple levels of dense attention computation blocks to enhance pixel-level correspondence. The final prediction results are obtained through inter-scale mixing and feature skip connection. The experimental results indicate that MLDAC obtains 55.1% and 26.8% one-shot mIoU self-supervised few-shot segmentation on the PASCAL-5i and COCO-20i datasets, respectively. In addition, it achieves 78.1% on the FSS-1000 few-shot dataset, proving its efficacy.

Edge and dense attention U-net for atrial scar segmentation in LGE-MRI

The segmentation of atrial scars in LGE-MRI images has huge potential value for clinical diagnosis and subsequent treatment. In clinical practice, atrial scars are usually manually calibrated by experienced experts, which is time-consuming and prone to errors. However, automatic segmentation also faces difficulties due to myocardial scars’ small size and variable shape. The present study introduces a dual branch network, incorporating edge attention, and deep supervision strategy. Edge attention is introduced to fully utilize the spatial relationship between the scar and the atrium. Besides, dense attention is embedded in bottom layer to solve feature disappearance. At the same time, deep supervision accelerates the convergence of the model and improves segmentation accuracy. The experiments were conducted on the 2022 atrial and scar segmentation challenge dataset. The results demonstrate that the proposed method has achieved superior performance.

Multi-dimensional dense attention network for pixel-wise segmentation of optic disc in colour fundus images.

Segmentation of retinal fragments like blood vessels, Optic Disc (OD), and Optic Cup (OC) enables the early detection of different retinal pathologies like Diabetic Retinopathy (DR), Glaucoma, etc. Accurate segmentation of OD remains challenging due to blurred boundaries, vessel occlusion, and other distractions and limitations. These days, deep learning is rapidly progressing in the segmentation of image pixels, and a number of network models have been proposed for end-to-end image segmentation. However, there are still certain limitations, such as limited ability to represent context, inadequate feature processing, limited receptive field, etc., which lead to the loss of local details and blurred boundaries. A multi-dimensional dense attention network, or MDDA-Net, is proposed for pixel-wise segmentation of OD in retinal images in order to address the aforementioned issues and produce more thorough and accurate segmentation results. In order to acquire powerful contexts when faced with limited context representation capabilities, a dense attention block is recommended. A triple-attention (TA) block is introduced in order to better extract the relationship between pixels and obtain more comprehensive information, with the goal of addressing the insufficient feature processing. In the meantime, a multi-scale context fusion (MCF) is suggested for acquiring the multi-scale contexts through context improvement. Specifically, we provide a thorough assessment of the suggested approach on three difficult datasets. In the MESSIDOR and ORIGA data sets, the suggested MDDA-NET approach obtains accuracy levels of 99.28% and 98.95%, respectively. The experimental results show that the MDDA-Net can obtain better performance than state-of-the-art deep learning models under the same environmental conditions.

Reconstructing Cancellous Bone From Down-Sampled Optical-Resolution Photoacoustic Microscopy Images With Deep Learning

ObjectiveBone diseases deteriorate the microstructure of bone tissue. Optical-resolution photoacoustic microscopy (OR-PAM) enables high spatial resolution of imaging bone tissues. However, the spatiotemporal trade-off limits the application of OR-PAM. The purpose of this study was to improve the quality of OR-PAM images without sacrificing temporal resolution. MethodsIn this study, we proposed the Photoacoustic Dense Attention U-Net (PADA U-Net) model, which was used for reconstructing full-scanning images from under-sampled images. Thereby, this approach breaks the trade-off between imaging speed and spatial resolution. ResultsThe proposed method was validated on resolution test targets and bovine cancellous bone samples to demonstrate the capability of PADA U-Net in recovering full-scanning images from under-sampled OR-PAM images. With a down-sampling ratio of [4, 1], compared to bilinear interpolation, the Peak Signal-to-Noise Ratio and Structural Similarity Index Measure values (averaged over the test set of bovine cancellous bone) of the PADA U-Net were improved by 2.325 dB and 0.117, respectively. ConclusionThe results demonstrate that the PADA U-Net model reconstructed the OR-PAM images well with different levels of sparsity. Our proposed method can further facilitate early diagnosis and treatment of bone diseases using OR-PAM.

A Residual Dense Attention Generative Adversarial Network for Microscopic Image Super-Resolution.

With the development of deep learning, the Super-Resolution (SR) reconstruction of microscopic images has improved significantly. However, the scarcity of microscopic images for training, the underutilization of hierarchical features in original Low-Resolution (LR) images, and the high-frequency noise unrelated with the image structure generated during the reconstruction process are still challenges in the Single Image Super-Resolution (SISR) field. Faced with these issues, we first collected sufficient microscopic images through Motic, a company engaged in the design and production of optical and digital microscopes, to establish a dataset. Secondly, we proposed a Residual Dense Attention Generative Adversarial Network (RDAGAN). The network comprises a generator, an image discriminator, and a feature discriminator. The generator includes a Residual Dense Block (RDB) and a Convolutional Block Attention Module (CBAM), focusing on extracting the hierarchical features of the original LR image. Simultaneously, the added feature discriminator enables the network to generate high-frequency features pertinent to the image's structure. Finally, we conducted experimental analysis and compared our model with six classic models. Compared with the best model, our model improved PSNR and SSIM by about 1.5 dB and 0.2, respectively.

Enhancing performance of vision transformers on small datasets through local inductive bias incorporation

Vision transformers (ViTs) achieve remarkable performance on large datasets, but tend to perform worse than convolutional neural networks (CNNs) when trained from scratch on smaller datasets, possibly due to a lack of local inductive bias in the architecture. Recent studies have therefore added locality to the architecture and demonstrated that it can help ViTs achieve performance comparable to CNNs in the small-size dataset regime. Existing methods, however, are architecture-specific or have higher computational and memory costs. Thus, we propose a module called Local InFormation Enhancer (LIFE) that extracts patch-level local information and incorporates it into the embeddings used in the self-attention block of ViTs. Our proposed module is memory and computation efficient, as well as flexible enough to process auxiliary tokens such as the classification and distillation tokens. Empirical results show that the addition of the LIFE module improves the performance of ViTs on small image classification datasets. We further demonstrate how the effect can be extended to downstream tasks, such as object detection and semantic segmentation. In addition, we introduce a new visualization method, Dense Attention Roll-Out, specifically designed for dense prediction tasks, allowing the generation of class-specific attention maps utilizing the attention maps of all tokens. The code for this project is available on Github (https://github.com/NeurAI-Lab/LIFEhttps://github.com/NeurAI-Lab/LIFE).

MRFormer: Multiscale retractable transformer for medical image progressive denoising via noise level estimation

Clear medical images are important for auxiliary diagnoses, but the images generated by various medical devices inevitably contain considerable noise. Although various models have been proposed for denoising, these methods ignore the fact that different types of medical images have different noise levels, which leads to unsatisfactory test results. In addition, collecting many medical images for training denoising models consumes many material resources. To address these issues, we formulate a progressive denoising architecture that contains preliminary and profound denoising. First, we construct a noise level estimation network to estimate the noise level via self-supervised learning and perform preliminary denoising with a dilated blind-spot network. Second, with the learned noise distribution, we synthesize noisy natural images to construct clean-noisy natural image pairs. Finally, we design a novel medical image denoising model for profound denoising by training these pairs. The proposed three-stage learning scheme and progressive denoising architecture not only solve the problem that the denoising model only adapts to a single noise level but also alleviate the lack of medical image pairs. Moreover, we integrate dense attention and sparse attention to constitute the retractable transformer module in the profound denoising model, which reconciles a wider receptive field and enhances the representation ability of the transformer, s allowing the denoising model to obtain retractable attention on the input feature and capture more local and global receptive fields simultaneously. The results of qualitative and quantitative experiments demonstrate the effectiveness of our method in removing noise at various levels.

Infrared–Visible Image Fusion through Feature-Based Decomposition and Domain Normalization

Infrared–visible image fusion is valuable across various applications due to the complementary information that it provides. However, the current fusion methods face challenges in achieving high-quality fused images. This paper identifies a limitation in the existing fusion framework that affects the fusion quality: modal differences between infrared and visible images are often overlooked, resulting in the poor fusion of the two modalities. This limitation implies that features from different sources may not be consistently fused, which can impact the quality of the fusion results. Therefore, we propose a framework that utilizes feature-based decomposition and domain normalization. This decomposition method separates infrared and visible images into common and unique regions. To reduce modal differences while retaining unique information from the source images, we apply domain normalization to the common regions within the unified feature space. This space can transform infrared features into a pseudo-visible domain, ensuring that all features are fused within the same domain and minimizing the impact of modal differences during the fusion process. Noise in the source images adversely affects the fused images, compromising the overall fusion performance. Thus, we propose the non-local Gaussian filter. This filter can learn the shape and parameters of its filtering kernel based on the image features, effectively removing noise while preserving details. Additionally, we propose a novel dense attention in the feature extraction module, enabling the network to understand and leverage inter-layer information. Our experiments demonstrate a marked improvement in fusion quality with our proposed method.

MPEDA-Net: A lightweight brain tumor segmentation network using multi-perspective extraction and dense attention

Malignant brain tumors are highly deadly, necessitating the quickly precise segmentation of tumor regions. Previously, clinicians manually classified brain tumor regions utilizing magnetic resonance imaging (MRI). A new trend is the use of computer vision processing to assist clinicians in clinical analysis. Even though numerous recent methodologies based on CNN have been presented, there remains a lack of high performance when evaluating regions in MRI images. Furthermore, there is still a possibility for development in terms of parameter number and computational complexity. To collect feature information and improve the relevance of contextual feature extraction, a multi-perspective extraction (MPE) module is proposed. MPE consists of three different convolutional kernels and special operations. In addition, a dense attention (DA) module is used to provide each point with an appropriate level of attention while fusing features. The effectiveness of these two modules has been proved through ablation experiments. The proposed MPEDA-Net achieves dice of 82.52%, 93.07%, and 87.67% on the R-BraTS2021 (reconstructed-BraTS2021) in the ET, WT, and TC respectively. In addition, the BraTS2018 and BraTS2019 experiments illustrate that the dice of ET, WT, and TC reach 82.44%, 91.38%, 88.27%, and 83.73%, 91.87%, 88.71%, respectively. The more effective segmentation performance shows that MPEDA-Net can significantly enhance brain tumor segmentation accuracy, exceeding several existing methods. The MPEDA-Net code is already available on GitHub: https://github.com/luohaohaoluo/MPEDANet-pytorch.

A Dense Attention Railway Foreign Object Detection Algorithm Based on Mask R-CNN

A Dense Attention Railway Foreign Object Detection Algorithm Based on Mask R-CNN

RMLANet: Random Multi-Level Attention Network for Shadow Detection and Removal

This paper addresses the problem of shadow detection and shadow removal from a single image. Despite awareness of utilizing both local and global contexts, previous works only aggregate features level by level in a coarse-to-fine manner. To overcome this problem, we present RMLANet, a novel Random Multi-Level Attention Network. To be specific, we first design a shuffled multi-level feature aggregation module to fuse the multi-level features and the guiding features using the self-attention mechanism. Nevertheless, the computational complexity of dense self-attention is unaffordable when processing high-resolution inputs. We argue that dense attention between any pixel pair is unnecessary due to the local consistency in images. Then we further propose a sparse attention mechanism to reduce the number of attention pairs, which greatly reduces the computational complexity. Through extensive experiments on four shadow detection and three shadow removal benchmark datasets, our proposed RMLANet achieves superior performance over current state-of-the-art approaches for both shadow detection and shadow removal. Codes are publicly available at https://github.com/LeipingJie/RMLANet.

Automatic detection and localisation of myocardial infarction using multi-channel dense attention neural network

Myocardial infarction (MI) is a life-threatening cardiac disease. Electrocardiogram (ECG) is crucial in the clinical diagnosis of MI, facilitating early localisation and diagnosis, guiding timely intervention, and improving patient outcomes. However, most previous studies required additional denoising operations and did not propose an effective method to reduce differences between patients. Therefore, this paper presents a multi-channel dense attention neural network (MCDANN) for MI detection and localisation using 12-lead ECG signals without denoising preprocessing. MCDANN is composed of 12 parallel channels dedicated to the automatic extraction of heartbeat features. The extracted features from all channels are then aggregated for MI detection and localisation. Each channel consists primarily of the dense layers and the squeeze-and-excitation (SE) module. The dense layers enhance the reuse of feature information within the leads, whereas the SE module weakens less critical information in the ECG classification task. We conduct inter-patient and intra-patient experiments to evaluate the MCDANN model based on the PTB diagnostic database. For MI detection, MCDANN achieved 99.94% accuracy and 99.92% F1 score in intra-patient experiments and 98.27% accuracy and 96.84% F1 score in inter-patient experiments. For MI localisation, MCDANN achieved 99.85% accuracy and 99.84% F1 score in intra-patient experiments, and 81.70% accuracy and 80.73% F1 score in inter-patient experiments. Based on the experimental findings, it is clear that the MCDANN model exhibits remarkable performance in MI detection and localisation, both for intra-patient and inter-patient cases. Notably, it substantially enhances the performance of inter-patient MI localisation compared to other existing methods.

A multi-scale global attention network for blood vessel segmentation from fundus images

Accurate segmentation of retinal fundus vessel images is vital to clinical diagnosis. Due to the intricate vascular morphology, high noise and low contrast of fundus vessel images, retinal fundus vessel segmentation is still a challenging task, especially for thin vessel segmentation. In recent years, on account of strong context feature extraction ability of deep learning, it has shown a remarkable performance in the automatic segmentation of retinal fundus vessels. However, it still exhibits certain limitations, such as information loss on micro objects or details, inadequate treatment of local features, etc. Faced with these challenging factors, we present a new multi-scale global attention network (MGA-Net). To realize effective feature representation, a dense attention U-Net network is proposed. Meanwhile, we design a global context attention (GCA) block to realize multi-scale feature fusion, allowing the global features from the deep network layers to flow to the shallow network layers. Further, aimed at retinal fundus vessel segmentation task again the class imbalance issue, the AG block is also introduced. Related experiments are conducted on CHASE_DB1, DRIVE and STARE datasets to show the effectiveness of proposed segmentation model. The experimental results demonstrate the robustness of the proposed method with F1 exceeding 82% on all three datasets and effectively improve the segmentation performance of thin vessels. The source code of proposed MGA-Net is available at https://github.com/gegao310/workspace.git.

Dense attention network identifies EEG abnormalities during working memory performance of patients with schizophrenia.

Patients with schizophrenia typically exhibit deficits in working memory (WM) associated with abnormalities in brain activity. Alterations in the encoding, maintenance and retrieval phases of sequential WM tasks are well established. However, due to the heterogeneity of symptoms and complexity of its neurophysiological underpinnings, differential diagnosis remains a challenge. We conducted an electroencephalographic (EEG) study during a visual WM task in fifteen schizophrenia patients and fifteen healthy controls. We hypothesized that EEG abnormalities during the task could be identified, and patients successfully classified by an interpretable machine learning algorithm. We tested a custom dense attention network (DAN) machine learning model to discriminate patients from control subjects and compared its performance with simpler and more commonly used machine learning models. Additionally, we analyzed behavioral performance, event-related EEG potentials, and time-frequency representations of the evoked responses to further characterize abnormalities in patients during WM. The DAN model was significantly accurate in discriminating patients from healthy controls, ACC = 0.69, SD = 0.05. There were no significant differences between groups, conditions, or their interaction in behavioral performance or event-related potentials. However, patients showed significantly lower alpha suppression in the task preparation, memory encoding, maintenance, and retrieval phases F(1,28) = 5.93, p = 0.022, η2 = 0.149. Further analysis revealed that the two highest peaks in the attention value vector of the DAN model overlapped in time with the preparation and memory retrieval phases, as well as with two of the four significant time-frequency ROIs. These results highlight the potential utility of interpretable machine learning algorithms as an aid in diagnosis of schizophrenia and other psychiatric disorders presenting oscillatory abnormalities.

Super-Resolution Reconstruction of Particleboard Images Based on Improved SRGAN

As an important forest product, particleboard can greatly save forestry resources and promote low-carbon development by reusing wood processing residues. The size of the entire particleboard is large, and there are problems with less image feature information and blurred defect outlines when obtaining the particleboard images. The super-resolution reconstruction technology can improve the quality of the particleboard surface images, making the defects clearer. In this study, the super-resolution dense attention generative adversarial network (SRDAGAN) model was improved to solve the problem that the super-resolution generative adversarial network (SRGAN) reconstructed image would produce artifacts and its performance needed to be improved. The Batch Normalization (BN) layer was removed, the convolutional block attention module (CBAM) was optimized to construct the dense block, and the dense blocks were constructed via a densely skip connection. Then, the corresponding 52,400 image blocks with high resolution and low resolution were trained, verified, and tested according to the ratio of 3:1:1. The model was comprehensively evaluated from the effect of image reconstruction and the three indexes of PSNR, SSIM, and LPIPS. It was found that compared with BICUBIC, SRGAN, and SWINIR, the PSNR index of SRDAGAN increased by 4.88 dB, 3.25 dB, and 2.68 dB, respectively; SSIM increased by 0.0507, 0.1122, and 0.0648, respectively; and LPIPS improved by 0.1948, 0.1065, and 0.0639, respectively. The reconstructed images not only had a clearer texture, but also had a more realistic expression of various features, and the performance of the model had been greatly improved. At the same time, this study also emphatically discussed the image reconstruction effect with defects. The result showed that the SRDAGAN proposed in this study can complete the super-resolution reconstruction of the particleboard images with high quality. In the future, it can also be further combined with defect detection for the actual production to improve the quality of forestry products and increase economic benefits.

TMSDNet: Transformer with multi‐scale dense network for single and multi‐view 3D reconstruction

Abstract3D reconstruction is a long‐standing problem. Recently, a number of studies have emerged that utilize transformers for 3D reconstruction, and these approaches have demonstrated strong performance. However, transformer‐based 3D reconstruction methods tend to establish the transformation relationship between the 2D image and the 3D voxel space directly using transformers or rely solely on the powerful feature extraction capabilities of transformers. They ignore the crucial role played by deep multi‐scale representation of the object in the voxel feature domain, which can provide extensive global shape and local detail information about the object in a multi‐scale manner. In this article, we propose a novel framework TMSDNet (transformer with multi‐scale dense network) for single‐view and multi‐view 3D reconstruction with transformer to solve this problem. Based on our well‐designed combined‐transformer Block, which is canonical encoder–decoder architecture, voxel features with spatial order can be extracted from the input image, which are used to further extract multi‐scale global features in parallel using a multi‐scale residual attention module. Furthermore, a residual dense attention block is introduced for deep local features extraction and adaptive fusion. Finally, the reconstructed objects are produced with the voxel reconstruction block. Experiment results on the benchmarks such as ShapeNet and Pix3D datasets demonstrate that TMSDNet outperforms the existing state‐of‐the‐art reconstruction methods substantially.

SAR Ship Detection Algorithm Based on Deep Dense Sim Attention Mechanism Network

Ship detection is of great significance in the interpretation of synthetic aperture radar (SAR) images. However, SAR generates inherent speckle noise when producing images, which poses many challenges for ship detection tasks. One major issue during detection is low accuracy caused by noise interference near the ship. To address this issue, this study aims to design a deep dense attention detection network for improving the accuracy of ship target detection in SAR. The proposed algorithm primarily uses a multilayer deep dense network to preliminarily extract ship image features, and subsequently introduces an attention network to further enhance these features. Finally, an anchor point mechanism is utilized to perform ship positioning regression estimation. Experimental results on public SAR ship datasets, including SAR ship detection dataset (SSDD) and SAR-Ship-Dataset, demonstrate that the proposed algorithm performs well in terms of speed and accuracy, and has better robustness and real-time performance compared to similar detection algorithms.

Dynamic scene deblurring with continuous cross-layer attention transmission

The deep convolutional neural networks (CNNs) using attention mechanism have achieved great success for dynamic scene deblurring. In most of these networks, only the features refined by the attention maps can be passed to the next layer and the attention maps of different layers are separated from each other, which does not make full use of the attention information from different layers in the CNN. To address this problem, we introduce a new continuous cross-layer attention transmission (CCLAT) mechanism that can exploit hierarchical attention information from all the convolutional layers. Based on the CCLAT mechanism, we use a very simple attention module to construct a novel residual dense attention fusion block (RDAFB). In RDAFB, the attention maps inferred from the outputs of the preceding RDAFB and each layer are directly connected to the subsequent ones, leading to a CCLAT mechanism. Taking RDAFB as the building block, we design an effective architecture for dynamic scene deblurring named RDAFNet. The experiments on benchmark datasets show that the proposed model outperforms the state-of-the-art deblurring approaches, and demonstrate the effectiveness of CCLAT mechanism. The source code is available on: https://github.com/xjmz6/RDAFNet.