Feature Enhancement Module Research Articles

Stolon-YOLO: A detecting method for stolon of strawberry seedling in glass greenhouse

Stolons are an essential nutritional organ of strawberry seedlings, whose quantity and vigorous growth directly impact the quality of the seedling cultivation. Accurately detecting stolon is essential for increasing productivity in seedling factories. To alleviate these issues of low recognition accuracy of stolon caused by complex backgrounds, diverse growth types and low proportion of effective information in the bounding box, we propose a Stolon-YOLO for visual recognition of stolon for the first time. In our method, improvements are focused on two aspects: HorBlock-decoupled head and Stem Block feature enhancement module. Firstly, a new decoupled head is designed by introducing the HorBlock, which enables the interaction of high-order spatial information and enhances the precision for locating stolons. Sequential gnConv of the HorBlock realizes better global modeling for incomplete and curved stolon to improve detection efficiency. Then, the Stem Block is used to reinforce feature expression and reduce dimensionality. We fuse the processing results of the Stem Block with the output of HorBlock in a residual form. Experimental results demonstrate that the Stolon-YOLO achieves 92.5% in precision, 89.7% in recall rate, 91.1% in F1 score and 88.5% in average precision for stolon detection, which outperforms the standard YOLOv7 by 3% in recall rate and 3.4% in average precision. Compared with YOLOX, SSD, CenterNet and Faster R-CNN, the Stolon-YOLO increases recall rate by 8.3%, 27.4%, 17%, and 3.5%, and increases AP by 12.2%, 28.9%, 2.4%, and 7.5%, respectively. Meanwhile, the Stolon-YOLO achieves a frame rate of 50.25 FPS, meeting real-time detection demands. The above results highlight that this study provides a fast and effective method for identifying the stolon of strawberry seedlings in glass greenhouses.

HAFE: A Hierarchical Awareness and Feature Enhancement Network for Scene Text Recognition

Scene text recognition (STR) is to obtain text information from natural text images. Compared to the optical character recognition, STR has attracted much attentions due to the variability of scene images. However, most STR methods tend to loss some spatial information due to only simply adding Resnet-45 to the front end of the encoder. To solve this problem, we propose a Hierarchical Awareness and Feature Enhancement (HAFE) network. Our proposed Hierarchical Awareness (HA) module can effectively extract the depth information of scene text images from multiple dimensions and then improve the anti-interference ability of the network. Besides, the proposed Feature Enhancement (FE) module can enhance the stability of the model through strengthening the global features, which is helpful to improve the accuracy of text recognition. On the whole, our HAFE network tends to strengthen the information in encoder, better learn the feature maps from different levels, and then improve the recognition accuracy of non-horizontal texts. Extensive comparative experiments on seven benchmark datasets show that the proposed method has the most advanced performance and outperforms most state-of-the-art ones. Our code will be released at website: https://github.com/HAFE in the future.

Dual‐route synthetic‐to‐real adaption for single image dehazing

AbstractSingle image dehazing in real scenarios is still a particularly challenging task due to the large domain discrepancy between synthetic and real hazy images. To address this problem, the dual‐route domain‐aware adaptation framework with the bridging domain is proposed. First, inspired by the effective multistage strategy in low‐level tasks, the bridging domain is constructed with the color‐preserved adaptive histogram equalization to facilitate the adaptation and improve the performance of real hazy images. Second, the totally shared structure is relaxed and the residual dual‐path domain‐aware modules (RDDM) for synthetic and the bridging domains are proposed, which facilitates extracting the domain‐specific haze information with the different parameters. Third, a half‐cyclic constraint is proposed for the unsupervised hazy images to avoid structure distortion during the unsupervised adversarial training process. Finally, for convenience in the inference stage, feature enhancement modules (FEM) are proposed for the original real hazy images to learn the pre‐process operation in the training stage. Extensive qualitative and quantitative experiments demonstrate that the proposed method significantly improves the dehazing performance on synthetic and real hazy images.

The Unmanned Aerial Vehicle (UAV)-Based Hyperspectral Classification of Desert Grassland Plants in Inner Mongolia, China

In recent years, grassland ecosystems have faced increasingly severe desertification, which has caused continuous changes in the vegetation composition in grassland ecosystems. Therefore, effective research on grassland plant taxa is crucial to exploring the process of grassland desertification. This study proposed a solution by constructing a UAV hyperspectral remote sensing system to collect the hyperspectral data of various species in desert grasslands. This approach overcomes the limitations of traditional grassland survey methods such as a low efficiency and insufficient spatial resolution. A streamlined 2D-CNN model with different feature enhancement modules was constructed, and an improved depth-separable convolution approach was used to classify the desert grassland plants. The model was compared with existing hyperspectral classification models, such as ResNet34 and DenseNet121, under the preprocessing condition of data downscaling by combining the variance and F-norm2. The results showed that the model outperformed the other models in terms of the overall classification accuracy, kappa coefficient, and memory occupied, achieving 99.216%, 98.735%, and 16.3 MB, respectively. This model could effectively classify desert grassland species. This method provides a new approach for monitoring grassland ecosystem degradation.

PointNAC: Copula-Based Point Cloud Semantic Segmentation Network

Three-dimensional point cloud data generally contain complex scene information and diversified category structures. Existing point cloud semantic segmentation networks tend to learn feature information between sampled center points and their neighboring points, while ignoring the scale and structural information of the spatial context of the sampled center points. To address these issues, this paper introduces PointNAC (PointNet based on normal vector and attention copula feature enhancement), a network designed for point cloud semantic segmentation in large-scale complex scenes, which consists of the following two main modules: (1) The local stereoscopic feature-encoding module: this feature-encoding process incorporates distance, normal vectors, and angles calculated based on the cosine theorem, enabling the network to learn not only the spatial positional information of the point cloud but also the spatial scale and geometric structure; and (2) the copula-based similarity feature enhancement module. Based on the stereoscopic feature information, this module analyzes the correlation among points in the local neighborhood. It enhances the features of positively correlated points while leaving the features of negatively correlated points unchanged. By combining these enhancements, it effectively enhances the feature saliency within the same class and the feature distinctiveness between different classes. The experimental results show that PointNAC achieved an overall accuracy (OA) of 90.9% and a mean intersection over union (MIoU) of 67.4% on the S3DIS dataset. And on the Vaihingen dataset, PointNAC achieved an overall accuracy (OA) of 85.9% and an average F1 score of 70.6%. Compared to the segmentation results of other network models on public datasets, our algorithm demonstrates good generalization and segmentation capabilities.

A Novel Adaptive Edge Aggregation and Multiscale Feature Interaction Detector for Object Detection in Remote Sensing Images

Object detection (OD) in remote sensing (RS) images is an important task in the field of computer vision. OD techniques have achieved impressive advances in recent years. However, complex background interference, large-scale variations, and dense instances pose significant challenges for OD. These challenges may lead to misalignment between features extracted by OD models and the features of real objects. To address these challenges, we explore a novel single-stage detection framework for the adaptive fusion of multiscale features and propose a novel adaptive edge aggregation and multiscale feature interaction detector (AEAMFI-Det) for OD in RS images. AEAMFI-Det consists of an adaptive edge aggregation (AEA) module, a feature enhancement module (FEM) embedded in a context-aware cross-attention feature pyramid network (2CA-FPN), and a pyramid squeeze attention (PSA) module. The AEA module employs an edge enhancement mechanism to guide the network to learn spatial multiscale nonlocal dependencies and solve the problem of feature misalignment between the network’s focus and the real object. The 2CA-FPN employs level-by-level feature fusion to enhance multiscale feature interactions and effectively mitigate the misalignment between the scales of the extracted features and the scales of real objects. The FEM is designed to capture the local and nonlocal contexts as auxiliary information to enhance the feature representation of information interaction between multiscale features in a cross-attention manner. We introduce the PSA module to establish long-term dependencies between multiscale spaces and channels for better interdependency refinement. Experimental results obtained using the NWPU VHR-10 and DIOR datasets demonstrate the superior performance of AEAMFI-Det in object classification and localization.

Evaluation of Breast Cancer Tumor-Infiltrating Lymphocytes on Ultrasound Images Based on a Novel Multi-Cascade Residual U-Shaped Network.

Breast cancer has become the leading cancer of the 21st century. Tumor-infiltrating lymphocytes (TILs) have emerged as effective biomarkers for predicting treatment response and prognosis in breast cancer. The work described here was aimed at designing a novel deep learning network to assess the levels of TILs in breast ultrasound images. We propose the Multi-Cascade Residual U-Shaped Network (MCRUNet), which incorporates a gray feature enhancement (GFE) module for image reconstruction and normalization to achieve data synergy. Additionally, multiple residual U-shaped (RSU) modules are cascaded as the backbone network to maximize the fusion of global and local features, with a focus on the tumor's location and surrounding regions. The development of MCRUNet is based on data from two hospitals and uses a publicly available ultrasound data set for transfer learning. MCRUNet exhibits excellent performance in assessing TILs levels, achieving an area under the receiver operating characteristic curve of 0.8931, an accuracy of 85.71%, a sensitivity of 83.33%, a specificity of 88.64% and an F1 score of 86.54% in the test group. It outperforms six state-of-the-art networks in terms of performance. The MCRUNet network based on breast ultrasound images of breast cancer patients holds promise for non-invasively predicting TILs levels and aiding personalized treatment decisions.

An Efficient Detector with Auxiliary Network for Remote Sensing Object Detection

Over the past few decades, advances in satellite and aerial imaging technology have made it possible to acquire high-quality remote sensing images. As one of the most popular research directions of computer vision, remote sensing object detection is widely researched due to the wide application in military and civil fields. The algorithms based on convolutional neural network have made great achievements in the field of object detection. However, plenty of small and densely distributed remote sensing objects against complex background pose some challenges to object detection. In this work, an efficient anchor-free based remote sensing object detector based on YOLO (You Only Look Once) is constructed. Firstly, the backbone network is simplified for the high efficiency of detection. In order to extract the features of densely distributed objects effectively, the detection scales are adjusted based on the backbone network. Secondly, aiming at the shortcomings of CBAM, the improved CJAM (Coordinate Joint Attention Mechanism) is proposed to deal with object detection under complex background. In addition, feature enhancement modules DPFE (Dual Path Feature Enhancement) and IRFE (Inception-ResNet-Feature Enhancement) as well as PRes2Net (Parallel Res2Net) are proposed. We combine CJAM with the above modules to create DC-CSP_n, CSP-CJAM-IRFE, and CJAM-PRes2Net for better feature extraction. Thirdly, a lightweight auxiliary network is constructed to integrate the low-level and intermediate information extracted from remote sensing images into the high-level semantic information of the backbone network. The auxiliary network allows the detector to locate the target efficiently. Fourthly, Swin Transformer is introduced into the ‘Neck’ part of the network so that the network can effectively grasp the global information. The mAP on DOTA1.5 and VEDAI datasets, which both contain a large number of small objects, reached 77.07% and 63.83%, respectively. Compared with advanced algorithms such as YOLO V4, YOLO V5s, YOLO V5l, and YOLO V7, our approach achieves the highest mAP.

AGLC-GAN: Attention-based global-local cycle-consistent generative adversarial networks for unpaired single image dehazing

Image dehazing is a critical image pre-processing task to estimate the haze-free images corresponding to the input hazy images. Despite the recent advances, the task of image dehazing remains challenging, especially in the unsupervised scenario. Several efforts can be found in the literature to dehaze images in a supervised set-up, where a huge number of paired (clear and hazy images) images are required for training. The supervised approaches often become biased towards the nature of haze present in the training hazy images, and produce less realistic images for query hazy images. We propose an Attention-based Global–Local Cycle-consistent Generative Adversarial Network (AGLC-GAN) for Unpaired Single Image Dehazing. The proposed CycleGAN-based AGLC-GAN model contains a dehazing generator encapsulating an autoencoder-like network with an attention mechanism comprising channel attention and pixel attention to deal with uneven haze intensity across the image. We use a global–local consistent discriminator to identify spatially varying haze and improve the stability of the discriminator. We adopt cyclic perceptual consistency loss to maintain consistency in the feature space. A dynamic feature enhancement module and an adaptive mix-up module are included in the proposed generator to dynamically obtain more spatially structured features and hence, adaptively preserve the flow of shallow features. Furthermore, we extensively experiment with the proposed model on multiple benchmark datasets for evaluating the efficacy of removing haze. The results of the experiments conducted in the study, demonstrate a significant quantitative and qualitative improvement over the existing methods for unpaired image dehazing. 11The code is available at the following link: https://github.com/jr925/AGLC-GAN

SiamHSFT: A Siamese Network-Based Tracker with Hierarchical Sparse Fusion and Transformer for UAV Tracking.

Due to high maneuverability as well as hardware limitations of Unmanned Aerial Vehicle (UAV) platforms, tracking targets in UAV views often encounter challenges such as low resolution, fast motion, and background interference, which make it difficult to strike a compatibility between performance and efficiency. Based on the Siamese network framework, this paper proposes a novel UAV tracking algorithm, SiamHSFT, aiming to achieve a balance between tracking robustness and real-time computation. Firstly, by combining CBAM attention and downward information interaction in the feature enhancement module, the provided method merges high-level and low-level feature maps to prevent the loss of information when dealing with small targets. Secondly, it focuses on both long and short spatial intervals within the affinity in the interlaced sparse attention module, thereby enhancing the utilization of global context and prioritizing crucial information in feature extraction. Lastly, the Transformer's encoder is optimized with a modulation enhancement layer, which integrates triplet attention to enhance inter-layer dependencies and improve target discrimination. Experimental results demonstrate SiamHSFT's excellent performance across diverse datasets, including UAV123, UAV20L, UAV123@10fps, and DTB70. Notably, it performs better in fast motion and dynamic blurring scenarios. Meanwhile, it maintains an average tracking speed of 126.7 fps across all datasets, meeting real-time tracking requirements.

Perceptual contrast and residual self-attention generative adversarial network-based for highly under-sampled MRI reconstruction

ObjectiveThe study of deep learning-based fast magnetic resonance imaging (MRI) reconstruction methods has become popular in recent years. However, it is still challenging to obtain high-fidelity MRI images from undersampled data with large acceleration factors. The objective of this study was to improve the reconstruction quality of under-sampled MR images by exploring the correlation between pixels in the image and the perceptual similarities between the initial reconstructed images and the fully sampled ones. MethodsExisting reconstruction methods often suffer from aliasing artefacts and blurry texture information. We propose a novel perceptual residual self-attention generative adversarial model, SRSAGAN, to tackle this issue for fast and accurate MRI reconstruction. The proposed method involves two operation steps. In step 1, leveraging the average information and similarity weights of high-level features, we designed a novel feature enhancement module based on residual self-attention to restore fine texture details in the reconstructed image. In step 2, to improve the accuracy of network inference, we compare the perceptual similarities between the initial reconstructed images and the fully sampled ones. ResultThe proposed algorithm was evaluated using the fastMRI dataset. It demonstrates that SRSAGAN outperforms state-of-art methods and reconstructs complex image texture with only 10 % retained the k-space raw data (10-fold acceleration). The reconstruction time is only about 7.78 ms.

Lightweight semantic segmentation network with configurable context and small object attention.

The current semantic segmentation algorithms suffer from encoding feature distortion and small object feature loss. Context information exchange can effectively address the feature distortion problem, but it has the issue of fixed spatial range. Maintaining the input feature resolution can reduce the loss of small object information but would slow down the network's operation speed. To tackle these problems, we propose a lightweight semantic segmentation network with configurable context and small object attention (CCSONet). CCSONet includes a long-short distance configurable context feature enhancement module (LSCFEM) and a small object attention decoding module (SOADM). The LSCFEM differs from the regular context exchange module by configuring long and short-range relevant features for the current feature, providing a broader and more flexible spatial range. The SOADM enhances the features of small objects by establishing correlations among objects of the same category, avoiding the introduction of redundancy issues caused by high-resolution features. On the Cityscapes and Camvid datasets, our network achieves the accuracy of 76.9 mIoU and 73.1 mIoU, respectively, while maintaining speeds of 87 FPS and 138 FPS. It outperforms other lightweight semantic segmentation algorithms in terms of accuracy.

RGB depth salient object detection via cross‐modal attention and boundary feature guidance

AbstractRGB depth (RGB‐D) salient object detection (SOD) is a meaningful and challenging task, which has achieved good detection performance in dealing with simple scenes using convolutional neural networks, however, it cannot effectively handle scenes with complex contours of salient objects or similarly coloured salient objects and background. A novel end‐to‐end framework is proposed for RGB‐D SOD, which comprises of four main components: the cross‐modal attention feature enhancement (CMAFE) module, the multi‐level contextual feature interaction (MLCFI) module, the boundary feature extraction (BFE) module, and the multi‐level boundary attention guidance (MLBAG) module. The CMAFE module retains the more effective salient features by employing a dual‐attention mechanism to filter noise from two modalities. In the MLCFI module, a shuffle operation is used for high‐level and low‐level channels to promote cross‐channel information communication, and rich semantic information is extracted. The BFE module converts salient features into boundary features to generate boundary maps. The MLBAG module produces saliency maps by aggregating multi‐level boundary saliency maps to guide cross‐modal features in the decode stage. Extensive experiments are conducted on six public benchmark datasets, with the results demonstrating that the proposed model significantly outperforms 23 state‐of‐the‐art RGB‐D SOD models with regards to multiple evaluation metrics.

Feature Refinement Method Based on the Two-Stage Detection Framework for Similar Pest Detection in the Field.

Efficient pest identification and control is critical for ensuring food safety. Therefore, automatic detection of pests has high practical value for Integrated Pest Management (IPM). However, complex field environments and the similarity in appearance among pests can pose a significant challenge to the accurate identification of pests. In this paper, a feature refinement method designed for similar pest detection in the field based on the two-stage detection framework is proposed. Firstly, we designed a context feature enhancement module to enhance the feature expression ability of the network for different pests. Secondly, the adaptive feature fusion network was proposed to avoid the suboptimal problem of feature selection on a single scale. Finally, we designed a novel task separation network with different fusion features constructed for the classification task and the localization task. Our method was evaluated on the proposed dataset of similar pests named SimilarPest5 and achieved a mean average precision (mAP) of 72.7%, which was better than other advanced object detection methods.

AFRE-Net: Adaptive Feature Representation Enhancement for Arbitrary Oriented Object Detection

Arbitrary-oriented object detection (AOOD) is a crucial task in aerial image analysis but is also faced with significant challenges. In current AOOD detectors, commonly used multi-scale feature fusion modules fall short in spatial and semantic information complement between scales. Additionally, fixed feature extraction structures are usually used following a fusion model, resulting in the inability of detectors to self-adjust. At the same time, feature fusion and extraction modules are designed in isolation and the internal synergy between them is ignored. The above problems result in feature representation deficiency, thus affecting the overall detection precision. To solve these problems, we first create a fine-grained feature pyramid network (FG-FPN) that not only provides richer spatial and semantic features, but also completes neighbor scale features in a self-learning mode. Subsequently, we propose a novel feature enhancement module (FEM) to fit FG-FPN. FEM authorizes the detection unit to automatically adjust the sensing area and adaptively suppress background interference, thereby generating stronger feature representations. Our proposed solution was tested through extensive experiments on challenging datasets, including DOTA (77.44% mAP), HRSC2016 (97.82% mAP), UCAS-AOD (91.34% mAP), as well as ICDAR2015 (86.27% F-score) and its effectiveness and high applicability are verified on all the above datasets.

A lightweight cow mounting behavior recognition system based on improved YOLOv5s

To improve the detection speed of cow mounting behavior and the lightness of the model in dense scenes, this study proposes a lightweight rapid detection system for cow mounting behavior. Using the concept of EfficientNetV2, a lightweight backbone network is designed using an attention mechanism, inverted residual structure, and depth-wise separable convolution. Next, a feature enhancement module is designed using residual structure, efficient attention mechanism, and Ghost convolution. Finally, YOLOv5s, the lightweight backbone network, and the feature enhancement module are combined to construct a lightweight rapid recognition model for cow mounting behavior. Multiple cameras were installed in a barn with 200 cows to obtain 3343 images that formed the cow mounting behavior dataset. Based on the experimental results, the inference speed of the model put forward in this study is as high as 333.3 fps, the inference time per image is 4.1 ms, and the model mAP value is 87.7%. The mAP value of the proposed model is shown to be 2.1% higher than that of YOLOv5s, the inference speed is 0.47 times greater than that of YOLOv5s, and the model weight is 2.34 times less than that of YOLOv5s. According to the obtained results, the model proposed in the current work shows high accuracy and inference speed and acquires the automatic detection of cow mounting behavior in dense scenes, which would be beneficial for the all-weather real-time monitoring of multi-channel cameras in large cattle farms.

A lightweight license plate detection algorithm based on deep learning

AbstractLicense plate detection is an important task in Intelligent Transportation Systems (ITS) and has a wide range of applications in vehicle management, traffic control, and public safety. In order to improve the accuracy and speed of mobile recognition, an improved lightweight YOLOv5s model is proposed for license plate detection. First, an improved Stemblock network is used to replace the original Focus layer in the network, which ensures strong feature expression capability and reduces a large number of parameters to lower the computational complexity; then, an improved lightweight network, ShuffleNetv2, is used to replace the backbone network of the YOLOv5s, which makes the model lighter and ensures the detection accuracy at the same time. Then, a feature enhancement module is designed to reduce the information loss caused by the rearrangement of the backbone network channels, which facilitates the information interaction in the feature fusion process; finally, the low‐, medium‐ and high‐level features in the Shufflenetv2 network structure are fused to form the final high‐level output features. Experimental results on the CCPD dataset show that compared to other methods this paper obtains better performance and faster speed in the license plate detection task, in which the average precision mean value reaches 96.6%, and can achieve a detection speed of 43.86 frame/s, and the parameter volume is reduced to 5.07 M.

Learning convolutional multi-level transformers for image-based person re-identification

As a vital vision task, person re-identification (Re-ID) aims to retrieve the same person under non-overlapping cameras. It is a very challenging task due to the presence of complex backgrounds, diverse illuminations and different perspectives. In this work, we integrate the advantages of convolutional neural networks (CNNs) and transformers, and propose a novel learning framework named convolutional multi-level transformer (CMT) for image-based person Re-ID. More specifically, we first propose a scale-aware feature enhancement (SFE) module to extract multi-scale local features from a pre-trained CNN backbone. Then, we introduce a part-aware transformer encoder (PTE) to further mine discriminative local information guided by global semantics. Finally, a deeply-supervised learning (DSL) technique is adopted to optimize the proposed CMT and improve its training efficiency. Extensive experiments on four large-scale Re-ID benchmarks demonstrate that our method performs favorably against several state-of-the-art methods.

A new method for safety helmet detection based on convolutional neural network.

Considering practical issues such as cost control of hardware facilities in engineering projects, it is a challenge to design a robust safety helmet detection method, which can be implemented on mobile or embedded devices with limited computing power. This paper presents an approach to optimize the BottleneckCSP structure in the YOLOv5 backbone network, which can greatly reduce the complexity of the model without changing the size of the network input and output. To eliminate the information loss caused by upsampling and enhance the semantic information of the feature map on the reverse path, this paper designs an upsampling feature enhancement module. Besides, To avoid the negative impact of redundant information generated by feature fusion on the detection results, this paper introduces a self-attention mechanism. That is, using the designed channel attention module and location attention module, adjacent shallow feature maps and upsampled feature maps are adaptively fused to generate new feature maps with strong semantics and precise location information. Compared with the existing methods with the fastest inference speed, under the same compute capability, the proposed method not only has faster inference speed, the FPS can reach 416, but also has better performance with mAP of 94.2%.

An attention-based dual-encoding network for fire flame detection using optical remote sensing

Automatic extraction of flame area plays an important role in forest fire detection, which can accurately understand the spatial distribution and development trend of forest fire, so as to effectively realize the protection of forest resources. However, due to the instability and spread of fires, and the complexity of the background, accurate early fire detection is extremely challenging. At the same time, the image pixel proportion of the flame area in early stage is much smaller than that in the background, which causes a serious class imbalance problem. With the fast development of deep learning, some achievements have been made in flame extraction, but there are still some deficiencies in the existing networks, such as limited feature representation, poor feature capturing ability on micro objects, insufficiency processing of local features, etc. This paper proposes an attention-based dual-encoding segmentation network, abbreviated as ADE-Net, for pixelwise early fire detection. To realize strong feature representation, a dual-encoding path, consisting of semantic units and spatial units, is introduced to extract richer features, and an attention fusion module (AFM) is introduced to fully integrate spatial and semantic information and achieve effective feature aggregation. In addition, faced with the class imbalance problem, a multi-attention fusion (MAF) module is introduced to obtain more discriminating features to make the segmentation network to focus on the key pixel areas. Furthermore, a feature enhancement module, named attention-guided enhancement (AGE) module, is proposed to enrich the feature representation of local feature maps. Finally, to realize better multi-scale global feature extraction and fusion, a global context fusion (GCF) module is proposed into the bottleneck layer for multi-scale feature enhancement. Experimental results show that the proposed ADE-Net has a good early fire detection ability from remote sensing images, and it has obtained a competitive advantage compared with advanced segmentation models.