Generative Adversarial Network Framework Research Articles

Semantic-Aware Face Deblurring With Pixel-Wise Projection Discriminator

Most recent face deblurring methods have leveraged the distribution modeling ability of generative adversarial networks (GANs) to impose a constraint that the deblurred image should follow the distribution of sharp ground-truth images. However, generating sharp face images with high fidelity and realistic properties from a blurry face image remains challenging under the GAN framework. To this end, we focus on modeling the joint distribution of sharp face images and segmentation label maps for face image deblurring in a GAN framework. We propose a semantic-aware pixel-wise projection (SAPP) discriminator that models pixel-label matching with semantic label map information and generates a pixel-wise probability map of realness for the input image as well as a per-image probability. Moreover, we introduce a prediction-weighted (PW) loss to focus on erroneous pixels in the output of the decoder, using per-pixel real/fake probability map to re-weight the contribution of each pixel in the decoder. Furthermore, we present a coarse-to-fine training technique for the generator, which encourages the generator to focus on global consistency in the early training stages and local details in the later stages. Extensive experimental results show that our method outperforms existing methods both quantitatively and qualitatively in terms of perceptual image quality.

LsmGANs: Image-Domain Least-Squares Migration Using a New Framework of Generative Adversarial Networks

LsmGANs: Image-Domain Least-Squares Migration Using a New Framework of Generative Adversarial Networks

Adversarial Spectral Super-Resolution for Multispectral Imagery Using Spatial Spectral Feature Attention Module

Acquiring high-quality hyperspectral imagery with high spatial and spectral resolution plays an important role in remote sensing. Due to the limited capacity of sensors, providing high spatial and spectral resolution is still a challenging issue. Spectral super-resolution (SSR) increases the spectral dimensionality of multispectral images to achieve resolution enhancement. In this article, we propose a spectral resolution enhancement method based on the generative adversarial network framework without introducing additional spectral responses prior. In order to adaptively rescale informative features for capturing interdependencies in the spectral and spatial dimensions, a spatial spectral feature attention module is introduced. The proposed method jointly exploits spatio-spectral distribution in the hyperspectral manifold to increase spectral resolution while maintaining spatial content consistency. Experiments are conducted on both synthetic Landsat 8 and Sentinel-2 radiance data and real coregistered advanced land image and Hyperion (MS and HS) images, which indicates the superiority of the proposed method compared to other state-of-the-art methods.

Learning Stage-Wise GANs for Whistle Extraction in Time-Frequency Spectrograms

Whistle contour extraction aims to derive animal whistles from time-frequency spectrograms as polylines. For toothed whales, whistle extraction results can serve as the basis for analyzing animal abundance, species identity, and social activities. During the last few decades, as long-term recording systems have become affordable, automated whistle extraction algorithms were proposed to process large volumes of recording data. Recently, a deep learning-based method demonstrated superior performance in extracting whistles under varying noise conditions. However, training such networks requires a large amount of labor-intensive annotation, which is not available for many species. To overcome this limitation, we present a framework of stage-wise generative adversarial networks (GANs), which compile new whistle data suitable for deep model training via three stages: generation of background noise in the spectrogram, generation of whistle contours, and generation of whistle signals. By separating the generation of different components in the samples, our framework composes visually promising whistle data and labels even when few expert annotated data are available. Regardless of the amount of human-annotated data, the proposed data augmentation framework leads to a consistent improvement in performance of the whistle extraction model, with a maximum increase of 1.69 in the whistle extraction mean F1-score. Our stage-wise GAN also surpasses one single GAN in improving whistle extraction models with augmented data. The data and code will be available at https://github.com/Paul-LiPu/CompositeGAN\_WhistleAugment.

Operating Performance Assessment Based on Semi-Supervised Cluster Generative Adversarial Networks for Gold Flotation Process

Operating performance assessment of gold flotation process plays an important role in improving the metallurgical performances and pursues the best comprehensive economic benefits. The appearance of froth is a good indicator of flotation performance; however, labeling the data is costly. To assess the operational performance of flotation process with a small amount of labeled data, this study proposes a semi-supervised cluster generative adversarial network (SSClusterGAN). First, latent variables are sampled from one-hot encoded variables and continuous normal distribution variables. While projecting all data into the latent variable space, the specific types of clustering in the latent space is realized by controlling the type of distribution of the learned latent codes, so that the training objectives of labeled data and unlabeled data are consistent. Then, by adapting the generative adversarial networks (GANs) framework, a pair of stacked discriminators are used to learn the conditional distribution of attributes for labeled and unlabeled data, respectively. In addition, using semi-supervised learning (SSL) as the basic learning technology, the high-confidence pseudo-labels obtained by training with unlabeled data are combined with the labeled training data to increase the training samples and effectively improve the accuracy of the model. Finally, the method was applied to the gold flotation process. The experimental results show that the method effectively utilizes unlabeled samples to expand the labeled dataset, combines the advantages of SSL and GAN, obtains accurate assessment results, and effectively improves the comprehensive economic benefits of the gold flotation process.

Advanced EGAN-IDS Framework for Resilience against Adversarial Attacks using Multi-headed Attention Module

A network's security depends heavily on its network intrusion detection system (NIDS). Monitoring network traffic continuously and spotting any unusual activity is its main duty. Even though several methods, such as machine learning-based network intrusion detection systems (ML-NIDSs), have been put forth and put into practice to improve the detection of malicious network traffic, many of the studies that have already been done do not adequately replicate real-world situations, where new attack types are continually emerging. Thus, more research is needed to determine how resistant intrusion detection systems are too adversarial and zero-day attacks. We provide and expand the Enhanced Generative Adversarial Network (EGAN-IDS) framework in this study. This framework creates hostile attack routes that are intended to avoid being discovered by five distinct Black Box ML-based intrusion detection systems. Artificial neural networks and self-awareness processes are used by EGAN-IDS to generate artificial adversarial attacks that are able to evade detection. According to our evaluation results, EGAN-IDS has produced adversarial patterns for a variety of attacks, which has reduced detection rates by an average of 15.93% for each of the five IDSs. These conclusions highlight the model's resilience and broad applicability.

Ultrafast Ultrasound Localization Microscopy by Conditional Generative Adversarial Network.

Ultrasound localization microscopy (ULM) overcomes the acoustic diffraction limit and enables the visualization of microvasculature at subwavelength resolution. However, challenges remain in ultrafast ULM implementation, where short data acquisition time, efficient data processing speed, and high imaging resolution need to be considered simultaneously. Recently, deep learning (DL)-based methods have exhibited potential in speeding up ULM imaging. Nevertheless, a certain number of ultrasound (US) data ( L frames) are still required to accumulate enough localized microbubble (MB) events, leading to an acquisition time within a time span of tens of seconds. To further speed up ULM imaging, in this article, we present a new DL-based method, termed as ULM-GAN. By using a modified conditional generative adversarial network (cGAN) framework, ULM-GAN is able to reconstruct a superresolution image directly from a temporal mean low-resolution (LR) image generated by averaging l -frame raw US images with l being significantly smaller than L . To evaluate the performance of ULM-GAN, a series of numerical simulations and phantom experiments are both implemented. The results of the numerical simulations demonstrate that when performing ULM imaging, ULM-GAN allows ∼40 -fold reduction in data acquisition time and ∼61 -fold reduction in computational time compared with the conventional Gaussian fitting method, without compromising spatial resolution according to the resolution scaled error (RSE). For the phantom experiments, ULM-GAN offers an implementation of ULM with ultrafast data acquisition time ( ∼0.33 s) and ultrafast data processing speed ( ∼0.60 s) that makes it promising to observe rapid biological activities in vivo.

FingerGAN: A Constrained Fingerprint Generation Scheme for Latent Fingerprint Enhancement.

Latent fingerprint enhancement is an essential preprocessing step for latent fingerprint identification. Most latent fingerprint enhancement methods try to restore corrupted gray ridges/valleys. In this paper, we propose a new method that formulates latent fingerprint enhancement as a constrained fingerprint generation problem within a generative adversarial network (GAN) framework. We name the proposed network FingerGAN. It can enforce its generated fingerprint (i.e, enhanced latent fingerprint) indistinguishable from the corresponding ground truth instance in terms of the fingerprint skeleton map weighted by minutia locations and the orientation field regularized by the FOMFE model. Because minutia is the primary feature for fingerprint recognition and minutia can be retrieved directly from the fingerprint skeleton map, we offer a holistic framework that can perform latent fingerprint enhancement in the context of directly optimizing minutia information. This will help improve latent fingerprint identification performance significantly. Experimental results on two public latent fingerprint databases demonstrate that our method outperforms the state of the arts significantly. The codes will be available for non-commercial purposes from https://github.com/HubYZ/LatentEnhancement.

Motion artifact correction in fetal MRI based on a Generative Adversarial network method

Fetal MR imaging is subject to artifacts, where the most common type is caused by motion. These artifacts can appear as blurring and/or ghosting in the affected sequences. Currently if the motion artifact is severe or covers essential fetal tissue, the sequence acquisition must be repeated for diagnostic decision-making. We propose a novel deep learning network to reduce and remove motion artifacts in fetal MRIs. It follows a Generative Adversarial Network (GAN) framework where the Generator consists of an Autoencoder structure containing Residual blocks with Squeeze and Excitation (SE), and the Discriminator uses a sequential Convolutional Neural Network (CNN) design. The loss function is composed of weighted subcomponents involving WGAN, L1, and perceptual losses. The proposed network was trained on a synthetically created motion artifact dataset, and further validated on real motion-degraded images. The creation of the synthetic dataset consisted of randomly modifying the k-space of each scan. On the synthetic dataset, the proposed network achieved an average SSIM and PSNR of 93.7 % and 33.5 dB respectively. For the real motion affected dataset, the proposed network attained an average BRISQUE score of 21.1. These results outperformed current state-of-the-art techniques including BM3D, RED-Net, NLM filtering, and WGAN-VGG. The presented network facilitates rapid and accurate post-processing for fetal MRI. It can also improve diagnostic accuracy and can save time and money by reducing the number of rescans caused by severe motion artifacts.

BPDGAN: A GAN-Based Unsupervised Back Project Dense Network for Multi-Modal Medical Image Fusion.

Single-modality medical images often cannot contain sufficient valid information to meet the information requirements of clinical diagnosis. The diagnostic efficiency is always limited by observing multiple images at the same time. Image fusion is a technique that combines functional modalities such as positron emission computed tomography (PET) and single-photon emission computed tomography (SPECT) with anatomical modalities such as computed tomography (CT) and magnetic resonance imaging (MRI) to supplement the complementary information. Meanwhile, fusing two anatomical images (like CT-MRI) is often required to replace single MRI, and the fused images can improve the efficiency and accuracy of clinical diagnosis. To this end, in order to achieve high-quality, high-resolution and rich-detail fusion without artificial prior, an unsupervised deep learning image fusion framework is proposed in this paper. It is named the back project dense generative adversarial network (BPDGAN) framework. In particular, we construct a novel network based on the back project dense block (BPDB) and convolutional block attention module (CBAM). The BPDB can effectively mitigate the impact of black backgrounds on image content. Conversely, the CBAM improves the performance of BPDGAN on the texture and edge information. To conclude, qualitative and quantitative experiments are tested to demonstrate the superiority of BPDGAN. In terms of quantitative metrics, BPDGAN outperforms the state-of-the-art comparisons by approximately 19.58%, 14.84%, 10.40% and 86.78% on AG, EI, Qabf and Qcv metrics, respectively.

A bearing fault diagnosis method for high-noise and unbalanced dataset

Purpose The purpose is using generative adversarial network (GAN) to solve the problem of sample augmentation in the case of imbalanced bearing fault data sets and improving residual network is used to improve the diagnostic accuracy of the bearing fault intelligent diagnosis model in the environment of high signal noise. Design/methodology/approach A bearing vibration data generation model based on conditional GAN (CGAN) framework is proposed. The method generates data based on the adversarial mechanism of GANs and uses a small number of real samples to generate data, thereby effectively expanding imbalanced data sets. Combined with the data augmentation method based on CGAN, a fault diagnosis model of rolling bearing under the condition of data imbalance based on CGAN and improved residual network with attention mechanism is proposed. Findings The method proposed in this paper is verified by the western reserve data set and the truck bearing test bench data set, proving that the CGAN-based data generation method can form a high-quality augmented data set, while the CGAN-based and improved residual with attention mechanism. The diagnostic model of the network has better diagnostic accuracy under low signal-to-noise ratio samples. Originality/value A bearing vibration data generation model based on CGAN framework is proposed. The method generates data based on the adversarial mechanism of GAN and uses a small number of real samples to generate data, thereby effectively expanding imbalanced data sets. Combined with the data augmentation method based on CGAN, a fault diagnosis model of rolling bearing under the condition of data imbalance based on CGAN and improved residual network with attention mechanism is proposed.

Deep and Domain Transfer Learning Aided Photoacoustic Microscopy: Acoustic Resolution to Optical Resolution.

Acoustic resolution photoacoustic micros- copy (AR-PAM) can achieve deeper imaging depth in biological tissue, with the sacrifice of imaging resolution compared with optical resolution photoacoustic microscopy (OR-PAM). Here we aim to enhance the AR-PAM image quality towards OR-PAM image, which specifically includes the enhancement of imaging resolution, restoration of micro-vasculatures, and reduction of artifacts. To address this issue, a network (MultiResU-Net) is first trained as generative model with simulated AR-OR image pairs, which are synthesized with physical transducer model. Moderate enhancement results can already be obtained when applying this model to in vivo AR imaging data. Nevertheless, the perceptual quality is unsatisfactory due to domain shift. Further, domain transfer learning technique under generative adversarial network (GAN) framework is proposed to drive the enhanced image's manifold towards that of real OR image. In this way, perceptually convincing AR to OR enhancement result is obtained, which can also be supported by quantitative analysis. Peak Signal to Noise Ratio (PSNR) and Structural Similarity Index (SSIM) values are significantly increased from 14.74 dB to 19.01 dB and from 0.1974 to 0.2937, respectively, validating the improvement of reconstruction correctness and overall perceptual quality. The proposed algorithm has also been validated across different imaging depths with experiments conducted in both shallow and deep tissue. The above AR to OR domain transfer learning with GAN (AODTL-GAN) framework has enabled the enhancement target with limited amount of matched in vivo AR-OR imaging data.

Geometrically-driven generation of mechanical designs through deep convolutional GANs

Despite the freedom Additive Manufacturing (AM) offers when manufacturing organic shapes, it still requires some geometrical criteria to avoid a part's collapse during printing. The most synergetic design approach to AM is Topology Optimization (TO), which finds an optimal free-form given mechanical constraints. However, it is hard for TO to integrate these layout geometry-related constraints and it seldom proposes printable shapes. Therefore, this work leverages the Deep Learning (DL) capability to handle spatial correlations within the mechanical design process by integrating the layout and mechanical constraints at the conceptual level. It proposes a DL-layout-driven solution (DL-TO) trained via a triple-discriminator Generative Adversarial Network (GAN) framework. The DL-TO's performance is demonstrated by generating mechanically valid 2D designs conforming with layout constraints in a fraction of a second. DL-TO's creativity is illustrated by its capability to generate designs with unseen input constraints (passive/active elements) and to propose several shapes for the same input mechanical constraints, a task that is hard for a traditional TO to achieve.

Multi-Attribute Discriminative Representation Learning for Prediction of Adverse Drug-Drug Interaction.

Adverse drug-drug interaction (ADDI) is a significant life-threatening issue, posing a leading cause of hospitalizations and deaths in healthcare systems. This paper proposes a unified Multi-Attribute Discriminative Representation Learning (MADRL) model for ADDI prediction. Unlike the existing works that equally treat features of each attribute without discrimination and do not consider the underlying relationship among drugs, we first develop a regularized optimization problem based on CUR matrix decomposition for joint representative drug and discriminative feature selection such that the selected drugs and features can well approximate the original feature spaces and the critical factors discriminative to ADDIs can be properly explored. Different from the existing models that ignore the consistent and unique properties among attributes, a Generative Adversarial Network (GAN) framework is then designed to capture the inter-attribute shared and intra-attribute specific representations of adverse drug pairs for exploiting their consensus and complementary information in ADDI prediction. Meanwhile, MADRL is compatible with any kind of attributes and capable of exploring their respective effects on ADDI prediction. An iterative algorithm based on the alternating direction method of multipliers is developed for optimization. Experiments on publicly available dataset demonstrate the effectiveness of MADRL when compared with eleven baselines and its six variants.

Robust Medical Image Colorization with Spatial Mask-Guided Generative Adversarial Network.

Color medical images provide better visualization and diagnostic information for doctors during clinical procedures than grayscale medical images. Although generative adversarial network-based image colorization approaches have shown promising results, in these methods, adversarial training is applied to the whole image without considering the appearance conflicts between the foreground objects and the background contents, resulting in generating various artifacts. To remedy this issue, we propose a fully automatic spatial mask-guided colorization with generative adversarial network (SMCGAN) framework for medical image colorization. It generates colorized images with fewer artifacts by introducing spatial masks, which encourage the network to focus on the colorization of the foreground regions instead of the whole image. Specifically, we propose a novel spatial mask-guided method by introducing an auxiliary foreground segmentation branch combined with the main colorization branch to obtain the spatial masks. The spatial masks are then used to generate masked colorized images where most background contents are filtered out. Moreover, two discriminators are utilized for the generated colorized images and masked generated colorized images, respectively, to assist the model in focusing on the colorization of foreground regions. We validate our proposed framework on two publicly available datasets, including the Visible Human Project (VHP) dataset and the prostate dataset from NCI-ISBI 2013 challenge. The experimental results demonstrate that SMCGAN outperforms the state-of-the-art GAN-based image colorization approaches with an average improvement of 8.48% in the PSNR metric. The proposed SMCGAN can also generate colorized medical images with fewer artifacts.

Characterization of the in-focus droplets in shadowgraphy systems via deep learning-based image processing method

It is important to accurately identify and measure in-focus droplets from shadowgraph droplet images that typically contain a large number of defocused droplets for the research of multiphase flow. However, conventional in-focus droplet identification methods are time-consuming and laborious due to the noise and background illumination in experimental data. In this paper, a deep learning-based method called focus-droplet generative adversarial network (FocGAN) is developed to automatically detect and characterize the focused droplets in shadow images. A generative adversarial network framework is adopted by our model to output binarized images containing only in-focus droplets, and inception blocks are used in the generator to enhance the extraction of multi-scale features. To emulate the real shadow images, an algorithm based on the Gauss blur method is developed to generate paired datasets to train the networks. The detailed architecture and performance of the model were investigated and evaluated by both the synthetic data and spray experimental data. The results show that the present learning-based method is far superior to the traditional adaptive threshold method in terms of effective extraction rate and accuracy. The comprehensive performance of FocGAN, including detection accuracy and robustness to noise, is higher than that of the model based on a convolutional neural network. Moreover, the identification results of spray images with different droplet number densities clearly exhibit the feasibility of FocGAN in real experiments. This work indicates that the proposed learning-based approach is promising to be widely applied as an efficient and universal tool for processing particle shadowgraph images.

Zero-Shot Learning via Structure-Aligned Generative Adversarial Network.

In this article, we propose a structure-aligned generative adversarial network framework to improve zero-shot learning (ZSL) by mitigating the semantic gap, domain shift, and hubness problem. The proposed framework contains two parts, i.e., a generative adversarial network with a softmax classifier part, and a structure-aligned part. In the first part, the generative adversarial network aims at generating pseudovisual features through the guiding generator and discriminator play the minimax two-player game together. At the same time, the softmax classifier is committed to increasing the interclass distance and reducing intraclass distance. Then, the harmful effect of domain shift and hubness problems can be mitigated. In another part, we introduce a structure-aligned module where the structural consistency between visual space and semantic space is learned. By aligning the structure between visual space and semantic space, the semantic gap between them can be bridged. The performance of classification is improved when the structure-aligned visual-semantic embedding space is transferred to the unseen classes. Our framework reformulates the ZSL as a standard fully supervised classification task using the pseudovisual features of unseen classes. Extensive experiments conducted on five benchmark data sets demonstrate that the proposed framework significantly outperforms state-of-the-art methods in both conventional and generalized settings.

Accuracy of Prostate Segmentation on Synthetic MR Images Generated from CT Scans Using a Novel Generative Adversarial Network for High-Dose-Rate Prostate Brachytherapy

<h3>Purpose/Objective(s)</h3> For patients undergoing prostate HDR brachytherapy at our institution, both CT and MRI scans are acquired to identify catheters and delineate the prostate, respectively. However, acquiring both scans increases the time and hospital resources needed for treatment. We hypothesized that a synthetic MRI (sMRI) could be generated with sufficiently high soft-tissue contrast from the planning CT using a novel Generative Adversarial Network (GAN) framework with the goal to use sMRI to delineate the prostate without requiring the MRI. <h3>Materials/Methods</h3> Our newly developed PixCycleGAN, a hybrid of Pix2Pix and CycleGAN to maximize the advantage of Pix2Pix and to add the strength of CycleGAN training with paired CT-MRI datasets acquired on the same day for previously treated HDR prostate patients. PixCycleGAN is composed of two Residual Network generators and four discriminators with two transformation cycles. To train PixCycleGAN, 58 paired CT and T2-weighted MRI datasets without metal or motion artifacts were selected. Using 10 independent datasets, the image quality of sMRI was evaluated using metrics including the mean absolute error (MAE), mean squared error (MSE), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index (SSIM). The accuracy of prostate segmentation was evaluated by having radiation oncologists manually delineate the prostate on real MRI (rMRI) and sMRI images. All rMRI and sMRI images were presented in a random order to reduce potential bias in segmentation. The dice similarity coefficient (DSC) and Hausdorff distance (HD) were evaluated for the prostate contours delineated on sMRI vs. rMRI. To estimate interobserver variability (IOV), DSC and HD of prostate contours independently delineated on rMRI by two radiation oncologists were calculated. <h3>Results</h3> Qualitatively, sMRI images showed enhanced soft-tissue contrast on the boundary of the prostate and rectum compared to the CT scans. In addition, sMRI contained features from both MRI and CT showing clear visualization of the boundary of the prostate and the catheters. Image quality metrics were: MAE=0.13±0.03, MSE=0.03±0.01, PSNR=69.2±1.42 (dB), SSIM=0.78±0.07. The DSC of the prostate contour was 0.85±0.05 for sMRI vs. rMRI and 0.84±0.07 for IOV (p=0.13). However, the HD for sMRI vs. rMRI (7.77±1.83 mm) was smaller than the HD for IOV (9.53±2.74 mm). <h3>Conclusion</h3> We developed a novel PixCycleGAN framework to generate a sMRI from the planning CT which depicted enhanced soft-tissue contrast at the boundaries of the prostate. The agreement between DSC of the prostate on sMRI vs. rMRI and DSC of IOV shows that the accuracy of prostate segmentation on sMRI is within the segmentation variation on rMRI between two radiation oncologists. In the future, we will evaluate dosimetric impact of using sMRI in prostate segmentation for HDR treatment planning.

Multiple auxiliary classifiers GAN for controllable image generation: Application to license plate recognition

One of the main challenges in developing machine learning (ML) applications is the lack of labeled and balanced datasets. In the literature, different techniques tackle this problem via augmentation, rendering, and over-sampling. Still, these methods produce datasets that appear less natural, exhibit poor balance, and have less variation. One potential solution is to leverage the Generative Adversarial Network (GAN) which achieves remarkable results in the generation of high-fidelity natural images. However, expanding the ability of GANs' to control generated image attributes with supervisory information remains a challenge. This research aims to propose an efficient method to generate high-fidelity natural images with total control of its main attributes. Therefore, this paper proposes a novel Multiple Auxiliary Classifiers GAN (MAC-GAN) framework based on Auxiliary Classifier GAN (AC-GAN), multi-conditioning, Wasserstein distance, gradient penalty, and dynamic loss. It is therefore presented as an efficient solution for highly controllable image synthesis red that allows to enrich and re-balance datasets beyond data augmentation. Furthermore, the effectiveness of MAC-GAN images on a target ML application called Automatic License Plate Recognition (ALPR) under limited resource constraints is probed. The improvement achieved is over 5% accuracy, which is mainly due to the ability of the MAC-GAN to create a balanced dataset with controllable synthesis and produce multiple (different) images with the same attributes, thus increasing the variation of the dataset in a more elaborate way than data augmentation techniques.

Attention-based Feature Fusion Generative Adversarial Network for yarn-dyed fabric defect detection

Defects on the surface of yarn-dyed fabrics are one of the important factors affecting the quality of fabrics. Defect detection is the core link of quality control. Due to the diversity of yarn-dyed fabric patterns and the scarcity of defect samples, reconstruction-based unsupervised deep learning algorithms have received extensive attention in the field of fabric defect detection. However, most existing deep learning algorithms cannot fully extract shallow, high-frequency and high-level information, which limits their ability to reconstruct yarn-dyed fabric images. In this article, we propose an Attention-based Feature Fusion Generative Adversarial Network framework for unsupervised defect detection of yarn-dyed fabrics. The framework utilizes a modified Feature Pyramid Network to fuse multi-level information and utilizes an attention mechanism to enhance the model's feature representation capabilities. The Attention-based Feature Fusion Generative Adversarial Network consists of an attention fusion generator and a patch-level discriminator. In the attention fusion generator, the Feature Pyramid Network with EfficientNetV2 as the backbone is used as the core building block, and different feature fusion methods are used to avoid the loss of information in the process of network deepening. The attention mechanism is used to enhance the channel and spatial-wise correlation of features, which helps the model to focus on more meaningful information by recalibrating the feature maps. In the discriminator, the patch-level discriminator is used to calculate the similarity between the reconstructed image and the original image from a local perspective, thereby improving the model's attention to texture details. Experimental results on public datasets demonstrate the effectiveness of the proposed method compared to other methods.