Lightweight Backbone Research Articles

DecoupleNet: A Lightweight Backbone Network With Efficient Feature Decoupling for Remote Sensing Visual Tasks

DecoupleNet: A Lightweight Backbone Network With Efficient Feature Decoupling for Remote Sensing Visual Tasks

Semantic segmentation of substation tools using an improved ICNet network

<p>In the field of substation operation and maintenance, real-time detection and precise segmentation of tools play an important role in maintaining the safe operation of the power grid and guiding operators to work safely. To improve the accuracy and real-time performance of semantic segmentation of substation operation and maintenance tools, we have proposed an improved, light-weight, real-time, semantic segmentation network based on an efficient image cascade network architecture (ICNet). The network uses multiscale branches and cascaded feature fusion units to extract rich multilevel features. We designed a semantic segmentation and purification module to deal with redundant and conflicting information in multiscale feature fusion. A lightweight backbone network was used in the feature extraction stage at different resolutions, and a recursive gated convolution was used in the upsampling stage to achieve high-order spatial interactions, thereby improving segmentation accuracy. Due to the lack of a substation tool semantic segmentation data set, we constructed one. Training and testing on the data set showed that the proposed model improved the accuracy of tool detection while ensuring real-time performance. Compared with the currently popular semantic segmentation network, it had better performance in real-time and accuracy, and provided a new semantic segmentation method for embedded platforms.</p>

Deep learning-based models for environmental management: Recognizing construction, renovation, and demolition waste in-the-wild

The construction industry generates a substantial volume of solid waste, often destinated for landfills, causing significant environmental pollution. Waste recycling is decisive in managing waste yet challenging due to labor-intensive sorting processes and the diverse forms of waste. Deep learning (DL) models have made remarkable strides in automating domestic waste recognition and sorting. However, the application of DL models to recognize the waste derived from construction, renovation, and demolition (CRD) activities remains limited due to the context-specific studies conducted in previous research. This paper aims to realistically capture the complexity of waste streams in the CRD context. The study encompasses collecting and annotating CRD waste images in real-world, uncontrolled environments. It then evaluates the performance of state-of-the-art DL models for automatically recognizing CRD waste in-the-wild. Several pre-trained networks are utilized to perform effectual feature extraction and transfer learning during DL model training. The results demonstrated that DL models, whether integrated with larger or lightweight backbone networks can recognize the composition of CRD waste streams in-the-wild which is useful for automated waste sorting. The outcome of the study emphasized the applicability of DL models in recognizing and sorting solid waste across various industrial domains, thereby contributing to resource recovery and encouraging environmental management efforts.

FOA-BDNet: A behavior detection algorithm for elevator maintenance personnel based on first-order deep network architecture

<p>The operation space of the vertical lift shaft is small, the components are complex, the occluding and different behavior space characteristics are similar, and the unsafe behavior is not easy to detect, which makes the operation safety of maintenance personnel in the elevator greatly threatened. This paper proposes an elevator maintenance personnel behavior detection algorithm based on the first-order deep network architecture (FOA-BDNet). First, a lightweight backbone feature extraction network is designed to meet the online real-time requirements of elevator maintenance environment monitoring video stream detection. Then, the feature fusion network structure of "far intersection and close connection" is proposed to fuse the fine-grained information with the coarse-grained information and to enhance the expression ability of deep semantic features. Finally, a first-order deep target detection algorithm adapted to the elevator scene is designed to identify and locate the behavior of maintenance personnel and to correctly detect unsafe behaviors. Experiments show that the detection accuracy rate on the self-built data set in this paper is 98.68%, which is 4.41% higher than that of the latest target detection model YOLOv8-s, and the reasoning speed reaches 69.51fps/s, which can be easily deployed in common edge devices and meet the real-time detection requirements for the unsafe behaviors of elevator scene maintenance personnel.</p>

Real-Time Semantic Segmentation via a Densely Aggregated Bilateral Network.

With the growing demands of applications on online devices, the speed-accuracy trade-off is critical in the semantic segmentation system. Recently, the bilateral segmentation network has shown promising capacity to achieve the balance between favorable accuracy and fast speed, and has become the mainstream backbone in real-time semantic segmentation. Segmentation of target objects relies on high-level semantics, whereas it requires detailed low-level features to model specific local patterns for accurate location. However, the lightweight backbone of bilateral architecture limits the extraction of semantic context and spatial details. And the late fusion of the bilateral streams incurs the insufficient aggregation of semantic context and spatial details. In this article, we propose a densely aggregated bilateral network (DAB-Net) for real-time semantic segmentation. In the context path, a patchwise context enhancement (PCE) module is proposed to efficiently capture the local semantic contextual information from spatialwise and channelwise, respectively. Meanwhile, a context-guided spatial path (CGSP) is designed to exploit more spatial information by encoding finer details from the raw image and the transition from the context path. Finally, with multiple interactions between bilateral branches, the intertwined outputs from bilateral streams are combined in a unified decoder for a final interaction to further enhance the feature representation, which generates the final segmentation prediction. Experimental results on three public benchmarks demonstrate that our proposed method achieves higher accuracy with a limited decay in speed, which performs favorably against state-of-the-art real-time approaches and runs at 31.1 frames/s (FPS) on the high resolution of [Formula: see text] . The source code is released at https://github.com/isyangshu/DABNet.

Transformer Fusion and Pixel-Level Contrastive Learning for RGB-D Salient Object Detection

Current RGB-D salient object detection (RGB-D SOD) methods mainly develop a generalizable model trained by binary cross-entropy (BCE) loss based on convolutional or Transformer backbones. However, they usually exploit convolutional modules to fuse multi-modality features, with little attention paid to capturing the long-range multi-modality interactions for feature fusion. Furthermore, BCE loss does not explicitly explore intra- and inter-pixel relationships in a joint embedding space. To address these issues, we propose a cross-modality interaction parallel-transformer (CIPT) module, which better captures the long-range multi-modality interactions, generating more comprehensive fusion features. Besides, we propose a pixel-level contrastive learning (PCL) method that improves inter-pixel discrimination and intra-pixel compactness, resulting in a well-structured embedding space and a better saliency detector. Specifically, we propose an asymmetric network (TPCL) for RGB-D SOD, which consists of a Swin V2 Transformer-based backbone and a designed lightweight backbone (LDNet). Moreover, an edge-guided module and a feature enhancement (FE) module are proposed to refine the learned fusion features. Extensive experiments demonstrate that our method achieves excellent performance against 15 state-of-the-art methods on seven public datasets. We expect our work to facilitate the exploration of applying Transformer and contrastive learning for RGB-D SOD tasks. Our codes and predicted saliency maps will be released on GitHub <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/TomorrowJW/TPCL_RGBDSOD</uri> .

UCM-Net: A U-Net-Like Tampered-Region-Related Framework for Copy-Move Forgery Detection

Copy-move forgery causes a big challenge to copy-move forgery detection (CMFD) due to that the photometrical characteristics of genuine and tampered regions in the same image remain highly consistent. A novel U-Net-like architecture with multiple asymmetric cross-layer connections associated with self-correlation and atrous spatial pyramid pooling (ASPP) between feature extraction module (FEM) and tampered region localization module (TRLM), called UCM-Net, is proposed in this paper. Different from existing deep learning based CMFD networks which indiscriminately process large or small tampered regions without considering the statistical characteristics of regions, FEM differentially treats large or small tampered regions by exploiting deep backbone networks to extract high-level features with rich semantic information for large tampered regions while utilizing lightweight backbone networks to extract low-level features for small tampered regions. Multiple cross-layer connections between two modules utilize the self-correlation calculation and ASPP to remove as much irrelevant semantic information as possible while retaining multi-scale tampered features from shallow to deep convolutional layers of FEM. Unlike the previous CMFD networks, which cannot capture multi-scale features because of simply stacking convolution blocks in the upsampling step, TRLM exploits multiple U-shaped residual U-block modules with different depths to change the receptive field of each point in the tampered feature maps so as to capture global and local information, greatly improving the localization accuracy of tampered regions. Experimental results on three publicly available databases demonstrate that UCM-Net outperforms several state-of-the-art algorithms in terms of various evaluation metrics.

Fruit tree canopy segmentation from UAV orthophoto maps based on a lightweight improved U-Net

Segmenting fruit tree canopies from drone remote sensing images is a prerequisite for achieving accurate agricultural monitoring and precision aerial spraying at the individual tree (instance) level. However, current instance segmentation algorithms for canopy segmentation require repeated sampling of digital orthophoto maps (DOMs), resulting in high computational complexity and poor performance in densely planted orchards. Therefore, we propose a canopy labeling method suitable for U-Net and a lightweight segmentation network, where a lightweight backbone network, an attention mechanism, and a focal loss function are introduced to improve the U-Net decoder, greatly reducing the computational complexity required for large-scale canopy segmentation. The feasibility and effectiveness of the proposed method were validated using datasets from two seasons of two different lychee orchards. The average recognition rate of the improved U-Net was 90.98 %, 28.59 % higher than the basic model U-Net, and the floating-point operations (FLOPs) amounted to 50.86 GFLOPS, 27.67 % lower than the basic model U-Net. Under the same experimental conditions, the proposed model outperformed mainstream semantic segmentation models such as Deeplabv3 + and ResNet50-U-Net, and it was more efficient than previous instance segmentation methods based on YOLACT, as it does not require repeated sampling of the same region. For the same area, the number of sampled small images decreased from 194 to 78, resulting in a 148 % overall efficiency improvement while achieving better segmentation results. Thus, the proposed model can be used to extract and locate the crown of a lychee orchard, aiding in accurate management of a lychee orchard and in a differentiated agricultural analysis and decision-making for individual lychee trees.

An Enhanced Detector for Vulnerable Road Users Using Infrastructure-Sensors-Enabled Device.

The precise and real-time detection of vulnerable road users (VRUs) using infrastructure-sensors-enabled devices is crucial for the advancement of intelligent traffic monitoring systems. To overcome the prevalent inefficiencies in VRU detection, this paper introduces an enhanced detector that utilizes a lightweight backbone network integrated with a parameterless attention mechanism. This integration significantly enhances the feature extraction capability for small targets within high-resolution images. Additionally, the design features a streamlined 'neck' and a dynamic detection head, both augmented with a pruning algorithm to reduce the model's parameter count and ensure a compact architecture. In collaboration with the specialized engineering dataset De_VRU, the model was deployed on the Hisilicon_Hi3516DV300 platform, specifically designed for infrastructure units. Rigorous ablation studies, employing YOLOv7-tiny as the baseline, confirm the detector's efficacy on the BDD100K and LLVIP datasets. The model not only achieved an improvement of over 12% in the mAP@50 metric but also realized a reduction in parameter count by more than 40%, and a 50% decrease in inference time. Visualization outcomes and a case study illustrate the detector's proficiency in conducting real-time detection with high-resolution imagery, underscoring its practical applicability.

MAUNext: a lightweight segmentation network for medical images

Objective. The primary objective of this study is to enhance medical image segmentation techniques for clinical research by prioritizing accuracy and the number of parameters. Approach. To achieve this objective, a novel codec-based MAUNext approach is devised, focusing on lightweight backbone design and the integration of skip connections utilizing multiscale, attention mechanism, and other strategic components. The approach is composed of three core modules: a multi-scale attentional convolution module for improved accuracy and parameter reduction, a collaborative neighbourhood-attention MLP encoding module to enhance segmentation performance, and a tiny skip-connected cross-layer semantic fusion module to bridge the semantic gap between encoder and decoder. Main results. The study extensively evaluates the MAUNext approach alongside eight state-of-the-art methods on three renowned datasets: Kagglelung, ISIC, and Brain. The experimental outcomes robustly demonstrate that the proposed approach surpasses other methods in terms of both parameter numbers and accuracy. This achievement holds promise for effectively addressing medical image segmentation tasks. Significance. Automated medical image segmentation, particularly in organ and lesion identification, plays a pivotal role in clinical diagnosis and treatment. Manual segmentation is resource-intensive, thus automated methods are highly valuable. The study underscores the clinical significance of automated segmentation by providing an advanced solution through the innovative MAUNext approach. This approach offers substantial improvements in accuracy and efficiency, which can significantly aid clinical decision-making and patient treatment.

Toward a Vision-Based Intelligent System: A Stacked Encoded Deep Learning Framework for Sign Language Recognition.

Sign language recognition, an essential interface between the hearing and deaf-mute communities, faces challenges with high false positive rates and computational costs, even with the use of advanced deep learning techniques. Our proposed solution is a stacked encoded model, combining artificial intelligence (AI) with the Internet of Things (IoT), which refines feature extraction and classification to overcome these challenges. We leverage a lightweight backbone model for preliminary feature extraction and use stacked autoencoders to further refine these features. Our approach harnesses the scalability of big data, showing notable improvement in accuracy, precision, recall, F1-score, and complexity analysis. Our model's effectiveness is demonstrated through testing on the ArSL2018 benchmark dataset, showcasing superior performance compared to state-of-the-art approaches. Additional validation through an ablation study with pre-trained convolutional neural network (CNN) models affirms our model's efficacy across all evaluation metrics. Our work paves the way for the sustainable development of high-performing, IoT-based sign-language-recognition applications.

LRTransDet: A Real-Time SAR Ship-Detection Network with Lightweight ViT and Multi-Scale Feature Fusion

In recent years, significant strides have been made in the field of synthetic aperture radar (SAR) ship detection through the application of deep learning techniques. These advanced methods have substantially improved the accuracy of ship detection. Nonetheless, SAR images present distinct challenges, including complex backgrounds, small ship targets, and noise interference, thereby rendering the detectors particularly demanding. In this paper, we introduce LRTransDet, a real-time SAR ship detector. LRTransDet leverages a lightweight vision transformer (ViT) and a multi-scale feature fusion neck to address these challenges effectively. First, our model implements a lightweight backbone that combines convolutional neural networks (CNNs) and transformers, thus enabling it to simultaneously capture both local and global features from input SAR images. Moreover, we boost the model’s efficiency by incorporating the faster weighted feature fusion (Faster-WF2) module and coordinate attention (CA) mechanism within the feature fusion neck. These components optimize computational resources while maintaining the model’s performance. To overcome the challenge of detecting small ship targets in SAR images, we refine the original loss function and use the normalized Wasserstein distance (NWD) metric and the intersection over union (IoU) scheme. This combination improves the detector’s ability to efficiently detect small targets. To prove the performance of our proposed model, we conducted experiments on four challenging datasets (the SSDD, the SAR-Ship Dataset, the HRSID, and the LS-SSDD-v1.0). The results demonstrate that our model surpasses both general object detectors and state-of-the-art SAR ship detectors in terms of detection accuracy (97.8% on the SSDD and 93.9% on the HRSID) and speed (74.6 FPS on the SSDD and 75.8 FPS on the HRSID), all while demanding 3.07 M parameters. Additionally, we conducted a series of ablation experiments to illustrate the impact of the EfficientViT, the Faster-WF2 module, the CA mechanism, and the NWD metric on multi-scale feature fusion and detection performance.

ElegantYOLO: An Agile and Real-time Vehicle Detection System for Live Environments

Vehicle detection plays a crucial role in traffic monitoring systems, but it often faces challenges like occlusion, vehicle size, and shape variations. Existing systems struggle with misinterpreting parts of other vehicles as actual vehicles, leading to losses and distorted shapes. To address these issues, a new approach called ElegantYOLO is introduced in this study. ElegantYOLO combines elements from LittleYOLO-SPP and ShortYOLO-CSP while modifying baseline network layer depths using the improved Ghost Module Extended connection method to reduce computational costs. The model's learning capabilities are improved by incorporating spatial attributes through the concatenation of Spatial Pooling blocks. The study employs the Alpha-IoU as the bounding box loss function, minimizing the disparity between predicted and ground truth boxes. This enhances vehicle detection accuracy and robustness. Additionally, the study uses the slicing-aided hyper inference (SAHI) technique, which allows the lightweight backbone network to capture more detailed vehicle information by processing higher-resolution images. Through extensive testing on various datasets such as PASCAL VOC 2007, 2012, and MS COCO 2014, the proposed model not only excels in detecting small vehicles but also demonstrates improved detection accuracy across different environmental conditions. The performance of ElegantYOLO surpasses both LittleYOLO and ShortYOLO by achieving an almost 10% higher mean average precision (mAP). Specifically, the model achieves outstanding results on PASCAL VOC and COCO datasets, with mAPs of 96.45% and 79.28%, respectively. Moreover, the proposed technique significantly enhances accuracy while reducing detection time.

FireViT: An Adaptive Lightweight Backbone Network for Fire Detection

Fire incidents pose a significant threat to human life and property security. Accurate fire detection plays a crucial role in promptly responding to fire outbreaks and ensuring the smooth execution of subsequent firefighting efforts. Fixed-size convolutions struggle to capture the irregular variations in smoke and flames that occur during fire incidents. In this paper, we introduce FireViT, an adaptive lightweight backbone network that combines a convolutional neural network (CNN) and transformer for fire detection. The FireViT we propose is an improved backbone network based on MobileViT. We name the lightweight module that combines deformable convolution with a transformer as th DeformViT block and compare multiple builds of this module. We introduce deformable convolution in order to better adapt to the irregularly varying smoke and flame in fire scenarios. In addition, we introduce an improved adaptive GELU activation function, AdaptGELU, to further enhance the performance of the network model. FireViT is compared with mainstream lightweight backbone networks in fire detection experiments on our self-made labeled fire natural light dataset and fire infrared dataset, and the experimental results show the advantages of FireViT as a backbone network for fire detection. On the fire natural light dataset, FireViT outperforms the PP-LCNet lightweight network backbone for fire target detection, with a 1.85% increase in mean Average Precision (mAP) and a 0.9 M reduction in the number of parameters. Additionally, compared to the lightweight network backbone MobileViT-XS, which similarly combines a CNN and transformer, FireViT achieves a 1.2% higher mAP while reducing the Giga-Floating Point Operations (GFLOPs) by 1.3. FireViT additionally demonstrates strong detection performance on the fire infrared dataset.

EDKSANet: An Efficient Dual-Kernel Split Attention Neural Network for the Classification of Tibetan Medicinal Materials

Tibetan medicine has received wide acclaim for its unique diagnosis and treatment methods. The identification of Tibetan medicinal materials, which are a vital component of Tibetan medicine, is a key research area in this field. However, traditional deep learning-based visual neural networks face significant challenges in efficiently and accurately identifying Tibetan medicinal materials due to their large number, complex morphology, and the scarcity of public visual datasets. To address this issue, we constructed a computer vision dataset with 300 Tibetan medicinal materials and proposed a lightweight and efficient cross-dimensional attention mechanism, the Dual-Kernel Split Attention (DKSA) module, which can adaptively share parameters of the kernel in both spatial and channel dimensions. Based on the DKSA module, we achieve efficient unification of convolution and self-attention under the CNN architecture and develop a new lightweight backbone architecture, EDKSANet, to provide enhanced performance for various computer vision tasks. As compared to RedNet, the top-1 accuracy is improved by 1.2% on an ImageNet dataset, and a larger margin of +1.5 box AP for object detection and an improvement of +1.3 mask AP for instance segmentation on MS-COCO dataset are obtained. Moreover, EDKSANet achieved excellent classification performance on the Tibetan medicinal materials dataset, with an accuracy of up to 96.85%.

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.

Lightweight Distortion-Aware Network for Salient Object Detection in Omnidirectional Images

Compared with 2D image salient object detection (SOD), SOD in omnidirectional images (or 360° images) usually suffers from geometric distortion. Although existing omnidirectional image SOD (ODI-SOD) methods have improved the detection accuracy obviously, their application may be cumbersome in real scenes due to their high computational cost. To avoid distortion and reduce the computational cost simultaneously in ODI-SOD, we propose a novel lightweight distortion-aware network, named LDNet, in this paper. First, to extract features with less distortion from ODIs, we integrate the distortion-aware convolution and depth-wise separable convolution (DSConv) into distortion-aware DSConv (DDSConv) and replace the regular convolutions in the last two blocks of the ResNet-18 with DDSConvs to obtain our lightweight backbone network (LD-ResNet-18). To enhance spatial information in each channel of the extracted features at each level comprehensively, then, we propose a lightweight distortion-aware channel-wise enhancement (DCE) module (only 0.05M parameters) including DDSConvs with various dilation rates, channel shuffle operation and attention mechanism, and employ a high-to-low dense connection structure to modulate the enhanced multi-level features. Besides, we design a distortion-aware self-correlation (DSC) module (only 0.02M parameters) for mining the contextual dependency of the features via a coarse-fine strategy, and the correlated features are refined by DCE modules and integrated by another dense connection structure. The final saliency map is predicted from the densely integrated features. Compared with 12 state-of-the-art methods on two public datasets, our lightweight LDNet achieves competitive or even better performance with only 2.9M parameters and 3.4G FLOPs, which balances the efficiency and performance.

LCA-YOLOv8-Seg: An Improved Lightweight YOLOv8-Seg for Real-Time Pixel-Level Crack Detection of Dams and Bridges

Remotely operated vehicles (ROVs) and unmanned aerial vehicles (UAVs) provide a solution for dam and bridges structural health information acquisition, but problems like effective damage-related information extraction also occur. Vision-based crack detection methods can replace traditional manual inspection and achieve fast and accurate crack detection. This paper thereby proposes a lightweight, real-time, pixel-level crack detection method based on an improved instance segmentation model. A lightweight backbone and a novel efficient prototype mask branch are proposed to decrease the complexity of the model and maintain the accuracy of the model. The proposed method attains an accuracy of 0.945 at 129 frames per second (FPS). Moreover, our model has smaller volume, lower computational requirements and is suitable for low-performance devices.

Sfnet: Faster and Accurate Semantic Segmentation Via Semantic Flow

In this paper, we focus on exploring effective methods for faster and accurate semantic segmentation. A common practice to improve the performance is to attain high-resolution feature maps with strong semantic representation. Two strategies are widely used: atrous convolutions and feature pyramid fusion, while both are either computationally intensive or ineffective. Inspired by the Optical Flow for motion alignment between adjacent video frames, we propose a Flow Alignment Module (FAM) to learn Semantic Flow between feature maps of adjacent levels and broadcast high-level features to high-resolution features effectively and efficiently. Furthermore, integrating our FAM to a standard feature pyramid structure exhibits superior performance over other real-time methods, even on lightweight backbone networks, such as ResNet-18 and DFNet. Then to further speed up the inference procedure, we also present a novel Gated Dual Flow Alignment Module to directly align high-resolution feature maps and low-resolution feature maps where we term the improved version network as SFNet-Lite. Extensive experiments are conducted on several challenging datasets, where results show the effectiveness of both SFNet and SFNet-Lite. In particular, when using Cityscapes test set, the SFNet-Lite series achieve 80.1 mIoU while running at 60 FPS using ResNet-18 backbone and 78.8 mIoU while running at 120 FPS using STDC backbone on RTX-3090. Moreover, we unify four challenging driving datasets (i.e., Cityscapes, Mapillary, IDD, and BDD) into one large dataset, which we named Unified Driving Segmentation (UDS) dataset. It contains diverse domain and style information. We benchmark several representative works on UDS. Both SFNet and SFNet-Lite still achieve the best speed and accuracy trade-off on UDS, which serves as a strong baseline in such a challenging setting. The code and models are publicly available at https://github.com/lxtGH/SFSegNets.

A lightweight network based on dual-stream feature fusion and dual-domain attention for white blood cells segmentation.

Accurate white blood cells segmentation from cytopathological images is crucial for evaluating leukemia. However, segmentation is difficult in clinical practice. Given the very large numbers of cytopathological images to be processed, diagnosis becomes cumbersome and time consuming, and diagnostic accuracy is also closely related to experts' experience, fatigue and mood and so on. Besides, fully automatic white blood cells segmentation is challenging for several reasons. There exists cell deformation, blurred cell boundaries, and cell color differences, cells overlapping or adhesion. The proposed method improves the feature representation capability of the network while reducing parameters and computational redundancy by utilizing the feature reuse of Ghost module to reconstruct a lightweight backbone network. Additionally, a dual-stream feature fusion network (DFFN) based on the feature pyramid network is designed to enhance detailed information acquisition. Furthermore, a dual-domain attention module (DDAM) is developed to extract global features from both frequency and spatial domains simultaneously, resulting in better cell segmentation performance. Experimental results on ALL-IDB and BCCD datasets demonstrate that our method outperforms existing instance segmentation networks such as Mask R-CNN, PointRend, MS R-CNN, SOLOv2, and YOLACT with an average precision (AP) of 87.41%, while significantly reducing parameters and computational cost. Our method is significantly better than the current state-of-the-art single-stage methods in terms of both the number of parameters and FLOPs, and our method has the best performance among all compared methods. However, the performance of our method is still lower than the two-stage instance segmentation algorithms. in future work, how to design a more lightweight network model while ensuring a good accuracy will become an important problem.