Multi-scale Feature Fusion Module Research Articles

A Novel Strategy for Global Lane Detection Based on Key-Point Regression and Multi-Scale Feature Fusion

Lane detection is a fundamental task for autonomous driving. Most existed deep learning-based methods use a combination of semantic segmentation and post-processing for lane information extraction. Such methods not only tend to ignore global lane information but also bring the problem of low efficiency due to the complex models. To solve these problems, a novel global lane detection method based on key-point regression and multi-scale features fusion (KP-MFF) is proposed in this study. Firstly, a regression strategy is presented to generate key-point sequences in each grid of the image for locating the lane. Moreover, a multi-scale feature fusion module is proposed to merge feature maps of different scales. Additionally, a rule-based fast post-processing method is proposed to deal with the series of key-point sequences output by the CNN model, which further improves the lane detection accuracy. Experiments on CULane and TuSimple datasets demonstrate that the proposed method performs more effectively (417 FPS on NVIDIA 2080Ti and 91 FPS on NVIDIA Jetson AGX Xavier) while maintaining competitive accuracy compared with state-of-the-art methods. The road test also validates the practicability and effectiveness of the proposed method.

Real-Time Robotic Multigrasp Detection Using Anchor-Free Fully Convolutional Grasp Detector

Robotic grasping is essential for intelligent manufacturing. This article presents a novel anchor-free grasp detector based on fully convolutional network for detecting multiple valid grasps from RGB-D images in real time. Grasp detection is formulated as a closest horizontal or vertical rectangle regression task and a grasp angle classification task. By directing predicting grasps at feature points, our method eliminates the predefined anchors that commonly used in prior methods, and thus anchor-related hyperparameters and complex computations are avoided. For suppressing ambiguous and low-quality training samples, a new sample assignment strategy that combines center-sampling and regression weights is proposed. Our method achieves a state-of-the-art accuracy of 99.4% on Cornell and 96.2% on Jacquard dataset, and real-time speed of 104 frames per second, with approximately 2× fewer parameters and 8× less training time compared to previous one-stage detector. Moreover, an efficient multiscale feature fusion module is integrated to improve the performance of multigrasp detection by 25%. In real-world robotic grasping of novel objects, our method achieves a grasp success rate of 91.3% for single object and 83.3% for multiple objects with only 26 ms used for the whole planning. The results demonstrate that our method is robust for potential industrial applications.

Semi-Supervised Image Deraining Using Knowledge Distillation

Image deraining has achieved considerable progress based on supervised learning with synthetic training pairs, but is usually limited in handling real-world rainy images. Although semi-supervised methods are suggested to exploit real-world rainy images when training deep deraining models, their performances are still notably inferior. To address this crucial issue, this work proposes a semi-supervised image deraining network with knowledge distillation (SSID-KD) for better exploiting real-world rainy images. In particular, the consistency of feature distribution of rain streaks extracted from synthetic and real-world rainy images is enforced by adopting knowledge distillation. Moreover, as for the backbone in SSID-KD, we propose the multi-scale feature fusion module and the pyramid fusion module to better extract deep features of rainy images. SSID-KD can relieve the problem of over-deraining or under-deraining for real-world rainy images, while it can keep comparable performance with supervised deraining methods on several benchmark datasets. Extensive experiments on both synthetic and real-world rainy images have validated that our SSID-KD not only can achieve better deraining results than existing semi-supervised deraining methods but also are quantitatively comparable with state-of-the-art supervised deraining methods. Benefiting from the well exploration of real-world rainy images, our SSID-KD can obtain more visually plausible deraining results. The source code and trained models are publicly available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/cuiyixin555/SSID-KD</uri> .

Coupled Global–Local object detection for large VHR aerial images

Object detection in large aerial images generally requires splitting each image into local images in the preprocessing step, and even the state-of-the-art models currently use this preprocessing method. However, image splitting often leads to deficiencies in contextual information and incomplete detection of oversized objects. At present, many object detection methods are designed to deal with large images. However, they require complex additional structures or training steps, and their applicability is limited. To address these problems, we propose the Coupled Global–Local (CGL) network, which can be easily embedded in frequently used detection models, to efficiently capture more information. Specifically, we employ a multiscale feature fusion module to share information between the global and local branches. Furthermore, a new convolution method is proposed to adaptively adjust the receptive field for better feature extraction. In addition, we find that detection results from global branches in the existing global–local architecture hinder the performance improvement on details when the detection results from different-resolution branches are fused. Therefore, on the global branch, a proposal filter and a nonlocal suppression (NLS) algorithm are developed to prevent small positive proposals and remove unqualified detection boxes easily and efficiently, respectively. We conduct extensive experiments on the DOTA-1.0, DOTA-1.5, and DOTA-2.0 data sets. The results demonstrate that CGL can significantly improve the detection performance of various baseline models for large very-high-resolution (VHR) aerial images without bells and whistles.

Secure and Reliable Indoor Localization Based on Multitask Collaborative Learning for Large-Scale Buildings

Accurate and reliable indoor location estimate is crucial for many Internet-of-Things (IoT) applications in the era of smart buildings. However, the positioning accuracy and security of the existing positioning works cannot meet the demands in the large-scale smart buildings scenarios covering multiple multifloor buildings. Therefore, in this article, we focus on the reliable and accurate localization under multibuilding and multifloor environments. We propose two novel designs, including a two-step reliable feature selector and a multitask collaborative positioning model. First, we design a two-step reliable feature selector based on an access point (AP) confidence model and manifold learning, to help select the most representative and reliable fingerprint features. Second, we propose a multitask cooperative positioning model, which consists of a multiscale feature fusion module to adaptively fuse multiscale features and a multitask joint learning module to effectively constrain the cumulative error of multiscale position. Finally, based on the above two, we propose a reliable multibuilding and multifloor localization method (RMBMFL), which can achieve accurate and reliable location estimates with low computational complexity in a smart building complex. We did real-world experiments in a 20 000 <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${m^{2}}$ </tex-math></inline-formula> site that covers three multistory buildings to evaluate the performance of the proposed RMBMFL. The experimental results show that RMBMFL achieves a building identification accuracy and a floor identification accuracy of 99%, and a room-level indoor localization with an average positioning error within 2 m, and outperforms state-of-the-art solutions.

MFFRand: Semantic Segmentation of Point Clouds Based on Multi-Scale Feature Fusion and Multi-Loss Supervision

With the application of the random sampling method in the down-sampling of point clouds data, the processing speed of point clouds has been greatly improved. However, the utilization of semantic information is still insufficient. To address this problem, we propose a point cloud semantic segmentation network called MFFRand (Multi-Scale Feature Fusion Based on RandLA-Net). Based on RandLA-Net, a multi-scale feature fusion module is developed, which is stacked by encoder-decoders with different depths. The feature maps extracted by the multi-scale feature fusion module are continuously concatenated and fused. Furthermore, for the network to be trained better, the multi-loss supervision module is proposed, which could strengthen the control of the training process of the local structure by adding sub-losses in the end of different decoder structures. Moreover, the trained MFFRand network could be connected to the inference network by different decoder terminals separately, which could achieve the inference of different depths of the network. Compared to RandLA-Net, MFFRand has improved mIoU on both S3DIS and Semantic3D datasets, reaching 71.1% and 74.8%, respectively. Extensive experimental results on the point cloud dataset demonstrate the effectiveness of our method.

Detection of Defects in Power Grid Inspection Images Based on Multi-scale Fusion

With the increasing application of visual application scenarios in various industries, power grid inspection system based on vision is widely used. However, the power grid intelligent inspection system has the problem of unsatisfactory accuracy of small target detection. To address this problem, this paper proposes a multi-scale fusion-based defect detection model for grid inspection images. The object features are first acquired by improving the CSPDarknet53 network using the attention mechanism. Second, the multi-scale feature fusion module is constructed to obtain more feature information. Finally, the defect detection model of the grid inspection image is constructed based on the multi-scale feature fusion module. The experimental results of our proposed algorithm are compared and analyzed with the current mainstream algorithms in all aspects. The optimal performance accuracy of our model reaches 0.927, which is improved by 6.06%.

A multi-task driven and reconfigurable network for cloud detection in cloud-snow coexistence regions from very-high-resolution remote sensing images

Cloud detection is a crucial step in remote sensing image preprocessing. Deep learning (DL) methods are currently preferred for cloud detection, and the original images need to be cropped into smaller patches for DL training and testing. However, in very-high-resolution (VHR) remote sensing cloud-snow coexistence scenes, the cropped patches have low feature richness and high inter-class feature similarity. As a result, the common DL methods are prone to misclassifying clouds in complex regions. Therefore, based on the expanded cloud detection dataset of VHR cloud-snow coexistence regions, this paper proposes a multi-task driven and reconfigurable network (MTDR-Net) to better perform the cloud detection work in the scenes. There are four modules in MTDR-Net: the high-resolution backbone with multiple projection heads module (HRB-MPH), the re-parametrizable multi-scale feature fusion module (RP-MFF), the lightweight and adaptive feature fusion module (LAFF), and the multi-task gradient flow guidance module (MTGFG). HRB-MPH facilitates global information interaction, enhances the multi-level granularity feature representation, and supports feature sharing between pixel-level and superpixel-level segmentation tasks. RP-MFF is used to capture the local multi-scale cloud features in the training stage and simplify losslessly the structure in the testing stage. LAFF is employed to reconstruct meaningful cloud features. MTGFG can provide unbiased guidance on dividing gradient flow between the multi-tasks. The experimental results show that the MTDR-Net obtains the greatest accuracy performance due to its anti-interference capacity for confusing ground objects and its excellent cloud boundary and thin cloud extraction ability. In addition, MTDR-Net has the fewest parameters, forward-reasoning floating point operations, memory footprint, and memory access volume. With the coarse-to-fine strategy, MTDR-Net can be further sped up by avoiding inference resource consumption in cloud-free/all-cloud scenes.

ELCD: Efficient Lunar Crater Detection Based on Attention Mechanisms and Multiscale Feature Fusion Networks from Digital Elevation Models

The detection and counting of lunar impact craters are crucial for the selection of detector landing sites and the estimation of the age of the Moon. However, traditional crater detection methods are based on machine learning and image processing technologies. These are inefficient for situations with different distributions, overlaps, and crater sizes, and most of them mainly focus on the accuracy of detection and ignore the efficiency. In this paper, we propose an efficient lunar crater detection (ELCD) algorithm based on a novel crater edge segmentation network (AFNet) to detect lunar craters from digital elevation model (DEM) data. First, in AFNet, a lightweight attention mechanism module is introduced to enhance the feature extract capabilities of networks, and a new multiscale feature fusion module is designed by fusing different multi-level feature maps to reduce the information loss of the output map. Then, considering the imbalance in the classification and the distributions of the crater data, an efficient crater edge segmentation loss function (CESL) is designed to improve the network optimization performance. Lastly, the crater positions are obtained from the network output map by the crater edge extraction (CEA) algorithm. The experiment was conducted on the PyTorch platform using two lunar crater catalogs to evaluate the ELCD. The experimental results show that ELCD has a superior detection accuracy and inference speed compared with other state-of-the-art crater detection algorithms. As with most crater detection models that use DEM data, some small craters may be considered to be noise that cannot be detected. The proposed algorithm can be used to improve the accuracy and speed of deep space probes in detecting candidate landing sites, and the discovery of new craters can increase the size of the original data set.

Flood Detection in Dual-Polarization SAR Images Based on Multi-Scale Deeplab Model

The proliferation of massive polarimetric Synthetic Aperture Radar (SAR) data helps promote the development of SAR image interpretation. Due to the advantages of powerful feature extraction capability and strong adaptability for different tasks, deep learning has been adopted in the work of SAR image interpretation and has achieved good results. However, most deep learning methods only employ single-polarization SAR images and ignore the water features embedded in multi-polarization SAR images. To fully exploit the dual-polarization SAR data and multi-scale features of SAR images, an effective flood detection method for SAR images is proposed in this paper. In the proposed flood detection method, a powerful Multi-Scale Deeplab (MS-Deeplab) model is constructed based on the dual-channel MobileNetV2 backbone and the classic DeeplabV3+ architecture to improve the ability of water feature extraction in SAR images. Firstly, the dual-channel feature extraction backbone based on the lightweight MobileNetV2 separately trains the dual-polarization SAR images, and the obtained training parameters are merged with the linear weighting to fuse dual-polarization water features. Given the multi-scale space information in SAR images, then, a multi-scale feature fusion module is introduced to effectively utilize multi-layer features and contextual information, which enhances the representation of water features. Finally, a joint loss function is constructed based on cross-entropy and a dice coefficient to deal with the imbalanced categorical distribution in the training dataset. The experimental results on the time series of Sentinel-1A SAR images show that the proposed method for flood detection has a strong ability to locate water boundaries and tiny water bodies in complex scenes. In terms of quantitative assessment, MS-Deeplab can achieve a better performance compared with other mainstream semantic segmentation models, including PSPNet, Unet and the original DeeplabV3+ model, with a 3.27% intersection over union (IoU) and 1.69% pixel accuracy (PA) improvement than the original DeeplabV3+ model.

A Few-Shot Learning Method for SAR Images Based on Weighted Distance and Feature Fusion

Convolutional Neural Network (CNN) has been widely applied in the field of synthetic aperture radar (SAR) image recognition. Nevertheless, CNN-based recognition methods usually encounter the problem of poor feature representation ability due to insufficient labeled SAR images. In addition, the large inner-class variety and high cross-class similarity of SAR images pose a challenge for classification. To alleviate the problems mentioned above, we propose a novel few-shot learning (FSL) method for SAR image recognition, which is composed of the multi-feature fusion network (MFFN) and the weighted distance classifier (WDC). The MFFN is utilized to extract input images’ features, and the WDC outputs the classification results based on these features. The MFFN is constructed by adding a multi-scale feature fusion module (MsFFM) and a hand-crafted feature insertion module (HcFIM) to a standard CNN. The feature extraction and representation capability can be enhanced by inserting the traditional hand-crafted features as auxiliary features. With the aid of information from different scales of features, targets of the same class can be more easily aggregated. The weight generation module in WDC is designed to generate category-specific weights for query images. The WDC distributes these weights along the corresponding Euclidean distance to tackle the high cross-class similarity problem. In addition, weight generation loss is proposed to improve recognition performance by guiding the weight generation module. Experimental results on the Moving and Stationary Target Acquisition and Recognition (MSTAR) dataset and the Vehicle and Aircraft (VA) dataset demonstrate that our proposed method surpasses several typical FSL methods.

MS-IAF: Multi-Scale Information Augmentation Framework for Aircraft Detection

Aircrafts have been an important object of study in the field of multi-scale image object detection due to their important strategic role. However, the multi-scale detection of aircrafts and their key parts from remote sensing images can be a challenge, as images often present complex backgrounds and obscured conditions. Most of today’s multi-scale datasets consist of independent objects and lack mixed annotations of aircrafts and their key parts. In this paper, we contribute a multi-scale aircraft dataset (AP-DATA) consisting of 7000 aircraft images that were taken in complex environments and obscured conditions. Our dataset includes mixed annotations of aircrafts and their key parts. We also present a multi-scale information augmentation framework (MS-IAF) to recognize multi-scale aircrafts and their key parts accurately. First, we propose a new deep convolutional module ResNeSt-D as the backbone, which stacks scattered attention in a multi-path manner and makes the receptive field more suitable for the object. Then, based on the combination of Faster R-CNN with ResNeSt-D, we propose a multi-scale feature fusion module called BFPCAR. BFPCAR overcomes the attention imbalance problem of the non-adjacent layers of the FPN module by reducing the loss of information between different layers and including more semantic features during information fusion. Based on AP-DATA, a dataset with three types of features, the average precision (AP) of MS-IAF reached 0.884, i.e., 2.67% higher than that of the original Faster R-CNN. The APs of these two modules were improved by 2.32% and 1.39%, respectively. The robustness of our proposed model was validated using the open sourced RSOD remote sensing image dataset, and the best accuracy was achieved.

Water Body Extraction in Remote Sensing Imagery Using Domain Adaptation-Based Network Embedding Selective Self-Attention and Multi-Scale Feature Fusion

A water body is a common object in remote sensing images and high-quality water body extraction is important for some further applications. With the development of deep learning (DL) in recent years, semantic segmentation technology based on deep convolution neural network (DCNN) brings a new way for automatic and high-quality body extraction from remote sensing images. Although several methods have been proposed, there exist two major problems in water body extraction, especially for high resolution remote sensing images. One is that it is difficult to effectively detect both large and small water bodies simultaneously and accurately predict the edge position of water bodies with DCNN-based methods, and the other is that DL methods need a large number of labeled samples which are often insufficient in practical application. In this paper, a novel SFnet-DA network based on the domain adaptation (DA) embedding selective self-attention (SSA) mechanism and multi-scale feature fusion (MFF) module is proposed to deal with these problems. Specially, the SSA mechanism is used to increase or decrease the space detail and semantic information, respectively, in the bottom-up branches of the network by selective feature enhancement, thus it can improve the detection capability of water bodies with drastic scale change and can prevent the prediction from being affected by other factors, such as roads and green algae. Furthermore, the MFF module is used to accurately acquire edge information by changing the number of the channel of advanced feature branches with a unique fusion method. To skip the labeling work, SFnet-DA reduces the difference in feature distribution between labeled and unlabeled datasets by building an adversarial relationship between the feature extractor and the domain classifier, so that the trained parameters of the labeled datasets can be directly used to predict the unlabeled images. Experimental results demonstrate that the proposed SFnet-DA has better performance on water body segmentation than state-of-the-art methods.

FCP-Net: A Feature-Compression-Pyramid Network Guided by Game-Theoretic Interactions for Medical Image Segmentation.

Medical image segmentation is a crucial step in diagnosis and analysis of diseases for clinical applications. Deep convolutional neural network methods such as DeepLabv3+ have successfully been applied for medical image segmentation, but multi-level features are seldom integrated seamlessly into different attention mechanisms, and few studies have fully explored the interactions between medical image segmentation and classification tasks. Herein, we propose a feature-compression-pyramid network (FCP-Net) guided by game-theoretic interactions with a hybrid loss function (HLF) for the medical image segmentation. The proposed approach consists of segmentation branch, classification branch and interaction branch. In the encoding stage, a new strategy is developed for the segmentation branch by applying three modules, e.g., embedded feature ensemble, dilated spatial mapping and channel attention (DSMCA), and branch layer fusion. These modules allow effective extraction of spatial information, efficient identification of spatial correlation among various features, and fully integration of multi-receptive field features from different branches. In the decoding stage, a DSMCA module and a multi-scale feature fusion module are used to establish multiple skip connections for enhancing fusion features. Classification and interaction branches are introduced to explore the potential benefits of the classification information task to the segmentation task. We further explore the interactions of segmentation and classification branches from a game theoretic view, and design an HLF. Based on this HLF, the segmentation, classification and interaction branches can collaboratively learn and teach each other throughout the training process, thus applying the conjoint information between the segmentation and classification tasks and improving the generalization performance. The proposed model has been evaluated using several datasets, including ISIC2017, ISIC2018, REFUGE, Kvasir-SEG, BUSI, and PH2, and the results prove its competitiveness compared with other state-of-the-art techniques.

Method of feature pyramid and attention enhancement network for pavement crack detection

The detection of pavement crack information plays an important role in the service evaluation of infrastructure engineering performance. The randomness of road noise, pollution, and spots make it difficult to ensure the accuracy of traditional image processing algorithms to detect cracks. Therefore, the crack identification technology under complex background conditions is particularly important. We propose a network architecture, named feature pyramid and attention enhancement network (FPAEN) for pavement crack detection. Through a new multiscale feature fusion module, the deep semantic information is integrated into the low-level convolution stage layer by layer, so the context information circulates throughout the whole network. In addition, the attention enhancement mechanism and the multiscale dilated convolution module are used to strengthen the network model’s ability to locate the crack pixels. To evaluate the performance of the proposed approach, we train FPAEN on the DeepCrack dataset and test it on four datasets. Experimental results show that our method outperforms other advanced crack detection approaches in F-score and mean intersection over union.

Video Super-Resolution Using Multi-Scale and Non-Local Feature Fusion

Video super-resolution can generate corresponding to high-resolution video frames from a plurality of low-resolution video frames which have rich details and temporally consistency. Most current methods use two-level structure to reconstruct video frames by combining optical flow network and super-resolution network, but this process does not deeply mine the effective information contained in video frames. Therefore, we propose a video super-resolution method that combines non-local features and multi-scale features to extract more in-depth effective information contained in video frames. Our method obtains long-distance effective information by calculating the similarity between any two pixels in the video frame through the non-local module, extracts the local information covered by different scale convolution cores through the multi-scale feature fusion module, and fully fuses feature information using different connection modes of convolution cores. Experiments on different data sets show that the proposed method is superior to the existing methods in quality and quantity.

Dynamic graph convolutional network for assembly behavior recognition based on attention mechanism and multi-scale feature fusion.

Intelligent recognition of assembly behaviors of workshop production personnel is crucial to improve production assembly efficiency and ensure production safety. This paper proposes a graph convolutional network model for assembly behavior recognition based on attention mechanism and multi-scale feature fusion. The proposed model learns the potential relationship between assembly actions and assembly tools for recognizing assembly behaviors. Meanwhile, the introduction of an attention mechanism helps the network to focus on the key information in assembly behavior images. Besides, the multi-scale feature fusion module is introduced to enable the network to better extract image features at different scales. This paper constructs a data set containing 15 types of workshop production behaviors, and the proposed assembly behavior recognition model is tested on this data set. The experimental results show that the proposed model achieves good recognition results, with an average assembly recognition accuracy of 93.1%.

Camouflage Target Segmentation Algorithm Using Multi-Scale Feature Extraction and Multi-Level Attention Mechanism

There are two problems when semantic segmentation algorithm is used in the battlefield environment: The camouflage target segmentation boundary is not ideal, and the small target segmentation accuracy is low. So a camouflage target semantic segmentation algorithm CSS-Net is proposed, which combines multi-scale feature extraction method and multi-level attention mechanism. The algorithm consistes of an encoder-decoder structure. In the encoder part, a lightweight and deep separable convolution with residual structure is used to construct the feature encoder to extract features from the reconnaissance images. In the decoder part, a strategy-selected multi-scale feature fusion module and a multi-level attention feature enhancement module are designed to obtain the multi-scale representation information and the channel information of the images. It furtherly enhances the semantic decoding process while gradually refining the segmentation results. The experiment results show that the CSS-Net algorithm can effectively complete the segmentation and recognition of camouflage targets in complex battlefield environment. The overall mIoU value reaches 91.98%, and the segmentation boundary is improved. Compared with DeepLab v3+ algorithm, the mIoU value of CSS-Net on camouflage small targets is increased by 3.71 percents, and the mIoU value on multi-scale targets exceeds 85%. The segmentation effects are improved significantly.

MFFE: Multi-scale Feature Fusion Enhanced Net for image dehazing

It is very challenging to perform dehazing operations on single haze images taken due to information degradation, for which we propose a novel single-image dehazing network based on U-Net. Most existing image dehazing algorithms only focus on whether they can remove the haze, but ignore the quality of the final dehazed image, which can lead to problems such as loss of image detail information and texture blurring, affecting the final recovery. We use the enhancement and error feedback mechanisms inside the super-resolution algorithm to gradually recover the haze-free images by introducing the Strengthen-Operate-subtract strategy in the decoder and demonstrate their effectiveness for the dehazing problem. Meanwhile, to solve the problem of missing information and utilization in the traditional U-Net structure, we design a multiscale feature fusion module that can effectively compensate for the missing contextual information and make full use of the disjoint features. In addition, an improved local binary pattern and SE attention mechanism are used to help the network obtain clearer details and texture images, and the accuracy of the enhanced dehazing network is improved using residual learning. We analyze the effectiveness of the proposed algorithm and show through extensive evaluation that our proposed model can be effectively used for single image dehazing and also has good performance compared to state-of-the-art methods on the test dataset.

AFENet: Attention Fusion Enhancement Network for Optic Disc Segmentation of Premature Infants.

Retinopathy of prematurity and ischemic brain injury resulting in periventricular white matter damage are the main causes of visual impairment in premature infants. Accurate optic disc (OD) segmentation has important prognostic significance for the auxiliary diagnosis of the above two diseases of premature infants. Because of the complexity and non-uniform illumination and low contrast between background and the target area of the fundus images, the segmentation of OD for infants is challenging and rarely reported in the literature. In this article, to tackle these problems, we propose a novel attention fusion enhancement network (AFENet) for the accurate segmentation of OD in the fundus images of premature infants by fusing adjacent high-level semantic information and multiscale low-level detailed information from different levels based on encoder–decoder network. Specifically, we first design a dual-scale semantic enhancement (DsSE) module between the encoder and the decoder inspired by self-attention mechanism, which can enhance the semantic contextual information for the decoder by reconstructing skip connection. Then, to reduce the semantic gaps between the high-level and low-level features, a multiscale feature fusion (MsFF) module is developed to fuse multiple features of different levels at the top of encoder by using attention mechanism. Finally, the proposed AFENet was evaluated on the fundus images of preterm infants for OD segmentation, which shows that the proposed two modules are both promising. Based on the baseline (Res34UNet), using DsSE or MsFF module alone can increase Dice similarity coefficients by 1.51 and 1.70%, respectively, whereas the integration of the two modules together can increase 2.11%. Compared with other state-of-the-art segmentation methods, the proposed AFENet achieves a high segmentation performance.