Over-segmentation Problem Research Articles

Semantic Segmentation of Corn Leaf Blotch Disease Images Based on U-Net Integrated with RFB Structure and Dual Attention Mechanism

Northern corn leaf blight (NCLB) is caused by a fungus and can be susceptible to the disease throughout the growing period of corn, posing a significant impact on corn yield. Aiming at the problems of under-segmentation, over-segmentation, and low segmentation accuracy in the traditional segmentation model of northern corn leaf blight, this study proposes a segmentation method based on an improved U-Net network model. By introducing a convolutional layer and maximum pooling layer to a VGG19 network, the channel attention module and spatial attention module (CBAM) are fused, and the squeeze excitation (SE) attention mechanism is combined. This enhances image feature decoding, integrates feature maps of each layer, strengthens the feature extraction process, expands the sensory fields and aggregates context information, and reduces the loss of location and dense semantic information caused by the pooling operation. Findings from the study show that the proposed NCLB-Net has significantly improved the MIoU and PA indexes, reaching 92.43% and 94.71%, respectively. Compared with the traditional methods, U-Net, SETR, DAnet, OCnet, PSPNet, etc., the MIoU is improved by 20.81%, 16.10%, 9.79%, 5.27%, and 11.06%, and the PA is improved by 11.49%, 8.18%, 9.54%, 13.11%, and 6.26%, respectively.

Deep self-organizing map neural networks improve the segmentation for inadequate plantar pressure imaging data set

ABSTRACT This study introduces a deep self-organizing map neural network based on level-set (LS-SOM) for the customization of a shoe-last defined from plantar pressure imaging data. To alleviate the over-segmentation problem of images, which refers to segmenting images into more subcomponents, a domain-based segmentation model of plantar pressure images was constructed. The domain growth algorithm was subsequently modified by optimizing its parameters. A SOM with 10, 15, 20, and 30 hidden layers was compared and validated according to domain growth characteristics by using merging and splitting algorithms. Furthermore, we incorporated a level set segmentation method into the plantar pressure image algorithm to enhance its efficiency. Compared to the literature, this proposed method has significantly improved pixel accuracy, average cross-combination ratio, frequency-weighted cross-combination ratio, and boundary F1 index comparison. Using the proposed methods, shoe lasts can be designed optimally, and wearing comfort is enhanced, particularly for people with high blood pressure.

The Improvement of Density Peaks Clustering Algorithm and Its Application to Point Cloud Segmentation of LiDAR.

This work focuses on the improvement of the density peaks clustering (DPC) algorithm and its application to point cloud segmentation in LiDAR. The improvement of DPC focuses on avoiding the manual determination of the cut-off distance and the manual selection of cluster centers. And the clustering process of the improved DPC is automatic without manual intervention. The cut-off distance is avoided by forming a voxel structure and using the number of points in the voxel as the local density of the voxel. The automatic selection of cluster centers is realized by selecting the voxels whose gamma values are greater than the gamma value of the inflection point of the fitted γ curve as cluster centers. Finally, a new merging strategy is introduced to overcome the over-segmentation problem and obtain the final clustering result. To verify the effectiveness of the improved DPC, experiments on point cloud segmentation of LiDAR under different scenes were conducted. The basic DPC, K-means, and DBSCAN were introduced for comparison. The experimental results showed that the improved DPC is effective and its application to point cloud segmentation of LiDAR is successful. Compared with the basic DPC, K-means, the improved DPC has better clustering accuracy. And, compared with DBSCAN, the improved DPC has comparable or slightly better clustering accuracy without nontrivial parameters.

WDFF-Net: Weighted Dual-Branch Feature Fusion Network for Polyp Segmentation With Object-Aware Attention Mechanism.

Colon polyps in colonoscopy images exhibit significant differences in color, size, shape, appearance, and location, posing significant challenges to accurate polyp segmentation. In this paper, a Weighted Dual-branch Feature Fusion Network is proposed for Polyp Segmentation, named WDFF-Net, which adopts HarDNet68 as the backbone network. First, a dual-branch feature fusion network architecture is constructed, which includes a shared feature extractor and two feature fusion branches, i.e. Progressive Feature Fusion (PFF) branch and Scale-aware Feature Fusion (SFF) branch. The branches fuse the deep features of multiple layers for different purposes and with different fusion ways. The PFF branch is to address the under-segmentation or over-segmentation problems of flat polyps with low-edge contrast by iteratively fusing the features from low, medium, and high layers. The SFF branch is to tackle the the problem of drastic variations in polyp size and shape, especially the missed segmentation problem for small polyps. These two branches are complementary and play different roles, in improving segmentation accuracy. Second, an Object-aware Attention Mechanism (OAM) is proposed to enhance the features of the target regions and suppress those of the background regions, to interfere with the segmentation performance. Third, a weighted dual-branch the segmentation loss function is specifically designed, which dynamically assigns the weight factors of the loss functions for two branches to optimize their collaborative training. Experimental results on five public colon polyp datasets demonstrate that, the proposed WDFF-Net can achieve a superior segmentation performance with lower model complexity and faster inference speed, while maintaining good generalization ability.

A rapid evaluation method of blasting effect based on optimized image segmentation algorithm and application in engineering

To quickly determine the blasting block degree and conduct an accurate and objective analysis of the tunnel blasting effect, this study has enhanced and improved upon the traditional genetic algorithm and Otsu algorithm. It has combined it with the marking watershed method and utilized ground digital acquisition to capture images of blasting debris. These images are then used in our custom-developed blasting analysis software to calculate the blasting block degree distribution and provide a quantitative analysis of blasting block degree. The research results show that the optimized image segmentation algorithm effectively improves the traditional threshold segmentation method on the poor effect of segmentation of the edge of the adherent block or the direct application of the watershed segmentation of the over-segmentation problem, to improve the segmentation accuracy based on the new segmentation technology is close to the traditional technology in terms of time. Through the self-developed software, the construction personnel in the project site to quickly obtain the blasting block degree histogram, block degree cumulative curve and other important indicators of the evaluation of the effect of blasting block degree, to provide data support for on-site construction, to assist in the modification of the blasting program, and to improve the efficiency of construction. This study realizes the rapid detection and block identification of blasting blocks, provides data support for the optimization of blasting parameters, and has good application and promotion value.

Tree Crown Segmentation and Diameter at Breast Height Prediction Based on BlendMask in Unmanned Aerial Vehicle Imagery

The surveying of forestry resources has recently shifted toward precision and real-time monitoring. This study utilized the BlendMask algorithm for accurately outlining tree crowns and introduced a Bayesian neural network to create a model linking individual tree crown size with diameter at breast height (DBH). BlendMask accurately outlines tree crown shapes and contours, outperforming traditional watershed algorithms in segmentation accuracy while preserving edge details across different scales. Subsequently, the Bayesian neural network constructs a model predicting DBH from the measured crown area, providing essential data for managing forest resources and conducting biodiversity research. Evaluation metrics like precision rate, recall rate, F1-score, and mAP index comprehensively assess the method’s performance regarding tree density. BlendMask demonstrated higher accuracy at 0.893 compared to the traditional watershed algorithm’s 0.721 accuracy based on experimental results. Importantly, BlendMask effectively handles over-segmentation problems while preserving edge details across different scales. Moreover, adjusting parameters during execution allows for flexibility in achieving diverse image segmentation effects. This study addresses image segmentation challenges and builds a model linking crown area to DBH using the BlendMask algorithm and a Bayesian neural network. The average discrepancies between calculated and measured DBH for Ginkgo biloba, Pinus tabuliformis, and Populus nigra varitalica were 0.15 cm, 0.29 cm, and 0.49cm, respectively, all within the acceptable forestry error margin of 1 cm. BlendMask, besides its effectiveness in crown segmentation, proves useful for various vegetation classification tasks like broad-leaved forests, coniferous forests, and grasslands. With abundant training data and ongoing parameter adjustments, BlendMask attains improved classification accuracy. This new approach shows great potential for real-world use, offering crucial data for managing forest resources, biodiversity research, and related fields, aiding decision-making processes.

Research on defect detection of bottle cap interior based on low-angle and large divergence angle vision system.

During the machine vision inspection of the inner section of bottle caps within pharmaceutical packaging, the unique conca bottom and convex side walls often create obstructions to the illumination. Consequently, this results in challenges such as irregular background and diminished feature contrast in the image, ultimately leading to the misidentification of defects. As a solution, a vision system characterized by a Low-Angle and Large Divergence Angle (LALDA) is presented in this paper. Using the large divergence angle of LED, combined with low-angle illumination, a uniform image of the side wall region with bright-field characteristics and a uniform image of inner circle region at the bottom with dark-field characteristics are obtained, thus solving the problems of light being obscured and brightness overexposure of the background. Based on the imaging characteristics of LALDA, a multi-channel segmentation (MCS) algorithm is designed. The HSV color space has been transformed, and the image is automatically segmented into multiple sub-regions by mutual calculation of different channels. Further, image homogenization and enhancement are used to eliminate fluctuations in the background and to enhance the contrast of defects. In addition, a variety of defect extraction methods are designed based on the imaging characteristics of different sub-regions, which can avoid the problem of over-segmentation in detection. In this paper, the LALDA is applied to the defect detection inside the cap of capsule medicine bottle, the detection speed is better than 400 pcs/min and the detection accuracy is better than 95%, which can meet the actual production line capacity and detection requirements.

Filament-necking localization method via combining improved PSO with rotated rectangle algorithm for safflower-picking robots

Safflower is one of the most important specialty economic crops in the world. Safflower blossoms are harvested continuously for 3–5 times. Untimely picking can affect the opening of filaments in the next crop, resulting in reduced filament production and serious economic losses. However, the safflower images collected by picking robots were affected by lighting conditions or complex backgrounds. The existing segmentation algorithms have the problem of insufficient segmentation or over-segmentation, because of the small safflower size and localization area. Therefore, combining improved particle swarm optimization (PSO) with a rotated rectangle algorithm based on the filament-necking localization method for safflower-picking robots was proposed. The R-component image in RGB color space was extracted as a preprocessing sample by combining the color features of safflower. The inertia weights in the PSO algorithm are improved to enhance the performance of the algorithm. An adaptive nonlinear function is introduced to search for the optimal threshold and initially segmented to obtain the binary image. Then, the barycenter and minimum outer rectangle of the contour were set up using the rotated rectangle algorithm based on the geometric features of the filaments. The circular region of interest (Cir-ROI) of filament-necking is determined. The Zhang-Suen refinement algorithm was used for skeleton extraction to design an algorithm for localizing the picking point of safflower filaments. To test safflower images collected in complex environments, the results of average processing time were 0.14 s, and the average relative target area error rate was 19.33 %. Moreover, the localization accuracy of the picking point was 89.75 %. The filament-necking localization method provides a theoretical basis and experimental data support for damage reduction and efficient harvesting of safflower filaments.

EfficientUNet: An efficient solution for breast tumour segmentation in ultrasound images

AbstractThe accurate segmentation of breast tumours is important for the diagnosis and treatment of breast cancer. When using the classic U‐Net, Attention‐UNet, and UNet++ to segment the tumour, there are problems of oversegmentation, incorrect segmentation, and poor edge continuity. In this paper, an effective tumour segmentation method, EfficientUNet, is proposed. EfficientUNet adopts a step‐by‐step enhancement method, combining ResNet18, a channel attention mechanism and deep supervision. ResNet18, as the encoder of the whole network, solves the problem of gradient disappearanceand improves the feature extraction ability of the model. The channel attention module makes the model more accurate in tumour edge processing. The deep supervision technology accelerates the model training and provides the convergence direction for the model. In addition, it is found that when adjusting the size of the image, the method of image filling before clipping (or zooming) is more conducive to model learning than the direct interpolation method. And a comparative experiment wasperformed on dataset B. Compared to U‐Net, Attention‐UNet and UNet++, EfficientUNet has the highest performance. Finally, the ablation experiment also indicated the effectiveness of each module in EfficientUNet.

Global superpixel-merging via set maximum coverage

Due to better boundary adherence and low computational cost, the superpixel segmentation algorithm SLIC (simple linear iterative clustering) has been widely applied in vision-based applications. However, limit to unavoidable over-segmentation problem, one has to consider region-merging to reconstruct entire objects from the segmented superpixels (or called regions). The existing region-merging methods are generated from data clustering, and avoidably suffer from error-merging, slow convergence speed, or easily dropping in LOCAL optimal problems, especially for high-resolution RS (remote sensing) images. In this paper, instead of data clustering, we propose a fast GLOBAL method based on Set Maximum Coverage, termed as MaxCov-merging. Theoretically, the existence of the maximum coverage is proved by using Bayes optimal decision principle. To speed up MaxCov-merging, some heuristic strategies are also provided. Finally, extensive verification and comparison are carried on the public and our collected high-resolution images. Compared with the state-of-the-art methods, the comparison shows the superiority of our MaxCov in terms of the performance of globality, ease of use and fast region-merging speed.

Bone tumor examination based on FCNN-4s and CRF fine segmentation fusion algorithm

Background and objectiveBone tumor is a kind of harmful orthopedic disease, there are benign and malignant points. Aiming at the problem that the accuracy of the existing machine learning algorithm for bone tumor image segmentation is not high, a bone tumor image segmentation algorithm based on improved full convolutional neural network which consists fully convolutional neural network (FCNN-4s) and conditional random field (CRF). MethodologyThe improved fully convolutional neural network (FCNN-4s) was used to perform coarse segmentation on preprocessed images. Batch normalization layers were added after each convolutional layer to accelerate the convergence speed of network training and improve the accuracy of the trained model. Then, a fully connected conditional random field (CRF) was fused to refine the bone tumor boundary in the coarse segmentation results, achieving the fine segmentation effect. ResultsThe experimental results show that compared with the traditional convolutional neural network bone tumor image segmentation algorithm, the algorithm has a great improvement in segmentation accuracy and stability, the average Dice can reach 91.56%, the real-time performance is better. ConclusionCompared with the traditional convolutional neural network segmentation algorithm, the algorithm in this paper has a more refined structure, which can effectively solve the problem of over-segmentation and under-segmentation of bone tumors. The segmentation prediction has better real-time performance, strong stability, and can achieve higher segmentation accuracy.

A Semantic Information-Based Optimized vSLAM in Indoor Dynamic Environments

In unknown environments, mobile robots can use visual-based Simultaneous Localization and Mapping (vSLAM) to complete positioning tasks while building sparse feature maps and dense maps. However, the traditional vSLAM works in the hypothetical environment of static scenes and rarely considers the dynamic objects existing in the actual scenes. In addition, it is difficult for the robot to perform high-level semantic tasks due to its inability to obtain semantic information from sparse feature maps and dense maps. In order to improve the ability of environment perception and accuracy of mapping for mobile robots in dynamic indoor environments, we propose a semantic information-based optimized vSLAM algorithm. The optimized vSLAM algorithm adds the modules of dynamic region detection and semantic segmentation to ORB-SLAM2. First, a dynamic region detection module is added to the vision odometry. The dynamic region of the image is detected by combining single response matrix and dense optical flow method to improve the accuracy of pose estimation in dynamic environment. Secondly, the semantic segmentation of images is implemented based on BiSeNet V2 network. For the over-segmentation problem in semantic segmentation, a region growth algorithm combining depth information is proposed to optimize the 3D segmentation. In the process of map building, semantic information and dynamic regions are used to remove dynamic objects and build an indoor map containing semantic information. The system not only can effectively remove the effect of dynamic objects on the pose estimation, but also use the semantic information of images to build indoor maps containing semantic information. The proposed algorithm is evaluated and analyzed in TUM RGB-D dataset and real dynamic scenes. The results show that the accuracy of our algorithm outperforms that of ORB-SLAM2 and DS-SLAM in dynamic scenarios.

Self-knowledge distillation for surgical phase recognition.

Advances in surgical phase recognition are generally led by training deeper networks. Rather than going further with a more complex solution, we believe that current models can be exploited better. We propose a self-knowledge distillation framework that can be integrated into current state-of-the-art (SOTA) models without requiring any extra complexity to the models or annotations. Knowledge distillation is a framework for network regularization where knowledge is distilled from a teacher network to a student network. In self-knowledge distillation, the student model becomes the teacher such that the network learns from itself. Most phase recognition models follow an encoder-decoder framework. Our framework utilizes self-knowledge distillation in both stages. The teacher model guides the training process of the student model to extract enhanced feature representations from the encoder and build a more robust temporal decoder to tackle the over-segmentation problem. We validate our proposed framework on the public dataset Cholec80. Our framework is embedded on top of four popular SOTA approaches and consistently improves their performance. Specifically, our best GRU model boosts performance by [Formula: see text] accuracy and [Formula: see text] F1-score over the same baseline model. We embed a self-knowledge distillation framework for the first time in the surgical phase recognition training pipeline. Experimental results demonstrate that our simple yet powerful framework can improve performance of existing phase recognition models. Moreover, our extensive experiments show that even with 75% of the training set we still achieve performance on par with the same baseline model trained on the full set.

Brain image segmentation of the corpus callosum by combining Bi-Directional Convolutional LSTM and U-Net using multi-slice CT and MRI

Background and objectiveTraditional disease diagnosis is usually performed by experienced physicians, but misdiagnosis or missed diagnosis still exists. Exploring the relationship between changes in the corpus callosum and multiple brain infarcts requires extracting corpus callosum features from brain image data, which requires addressing three key issues. (1) automation, (2) completeness, and (3) accuracy. Residual learning can facilitate network training, Bi-Directional Convolutional LSTM (BDC-LSTM) can exploit interlayer spatial dependencies, and HDC can expand the receptive domain without losing resolution. MethodsIn this paper, we propose a segmentation method by combining BDC-LSTM and U-Net to segment the corpus callosum from multiple angles of brain images based on computed tomography (CT) and magnetic resonance imaging (MRI) in which two types of sequence, namely T2-weighted imaging as well as the Fluid Attenuated Inversion Recovery (Flair), were utilized. The two-dimensional slice sequences are segmented in the cross-sectional plane, and the segmentation results are combined to obtain the final results. Encoding, BDC- LSTM, and decoding include convolutional neural networks. The coding part uses asymmetric convolutional layers of different sizes and dilated convolutions to get multi-slice information and extend the convolutional layers' perceptual field. ResultsThis paper uses BDC-LSTM between the encoding and decoding parts of the algorithm. On the image segmentation of the brain in multiple cerebral infarcts dataset, accuracy rates of 0.876, 0.881, 0.887, and 0.912 were attained for the intersection of union (IOU), dice similarity coefficient (DS), sensitivity (SE), and predictive positivity value (PPV). The experimental findings demonstrate that the algorithm outperforms its rivals in accuracy. ConclusionThis paper obtained segmentation results for three images using three models, ConvLSTM, Pyramid-LSTM, and BDC-LSTM, and compared them to verify that BDC-LSTM is the best method to perform the segmentation task for faster and more accurate detection of 3D medical images. We improve the convolutional neural network segmentation method to obtain medical images with high segmentation accuracy by solving the over-segmentation problem.

DSGA-Net: Deeply separable gated transformer and attention strategy for medical image segmentation network

To address the problems of under-segmentation and over-segmentation of small organs in medical image segmentation. We present a novel medical image segmentation network model with Depth Separable Gating Transformer and a Three-branch Attention module (DSGA-Net). Firstly, the model adds a Depth Separable Gated Visual Transformer (DSG-ViT) module into its Encoder to enhance (i) the contextual links among global, local, and channels and (ii) the sensitivity to location information. Secondly, a Mixed Three-branch Attention (MTA) module is proposed to increase the number of features in the up-sampling process. Meanwhile, the loss of feature information is reduced when restoring the feature image to the original image size. By validating Synapse, BraTs2020, and ACDC public datasets, the Dice Similarity Coefficient (DSC) of the results of DSGA-Net reached 81.24%,85.82%, and 91.34%, respectively. Moreover, the Hausdorff Score (HD) decreased to 20.91% and 5.27% on the Synapse and BraTs2020. There are 10.78% and 0.69% decreases compared to the Baseline TransUNet. The experimental results indicate that DSGA-Net achieves better segmentation than most advanced methods.

Hepatocellular carcinoma segmentation and pathological differentiation degree prediction method based on multi-task learning

Hepatocellular carcinoma (HCC) is the most common liver malignancy, where HCC segmentation and prediction of the degree of pathological differentiation are two important tasks in surgical treatment and prognosis evaluation. Existing methods usually solve these two problems independently without considering the correlation of the two tasks. In this paper, we propose a multi-task learning model that aims to accomplish the segmentation task and classification task simultaneously. The model consists of a segmentation subnet and a classification subnet. A multi-scale feature fusion method is proposed in the classification subnet to improve the classification accuracy, and a boundary-aware attention is designed in the segmentation subnet to solve the problem of tumor over-segmentation. A dynamic weighted average multi-task loss is used to make the model achieve optimal performance in both tasks simultaneously. The experimental results of this method on 295 HCC patients are superior to other multi-task learning methods, with a Dice similarity coefficient (Dice) of (83.9 ± 0.88)% on the segmentation task, while the average recall is (86.08 ± 0.83)% and an F1 score is (80.05 ± 1.7)% on the classification task. The results show that the multi-task learning method proposed in this paper can perform the classification task and segmentation task well at the same time, which can provide theoretical reference for clinical diagnosis and treatment of HCC patients.

Point Cloud Data Processing Optimization in Spectral and Spatial Dimensions Based on Multispectral Lidar for Urban Single-Wood Extraction

Lidar can effectively obtain three-dimensional information on ground objects. In recent years, lidar has developed rapidly from single-wavelength to multispectral hyperspectral imaging. The multispectral airborne lidar Optech Titan is the first commercial system that can collect point cloud data on 1550, 1064, and 532 nm channels. This study proposes a method of point cloud segmentation in the preprocessed intensity interpolation process to solve the problem of inaccurate intensity at the boundary during point cloud interpolation. The entire experiment consists of three steps. First, a multispectral lidar point cloud is obtained using point cloud segmentation and intensity interpolation; the spatial dimension advantage of the multispectral point cloud is used to improve the accuracy of spectral information interpolation. Second, point clouds are divided into eight categories by constructing geometric information, spectral reflectance information, and spectral characteristics. Accuracy evaluation and contribution analysis are also conducted through point cloud truth value and classification results. Lastly, the spatial dimension information is enhanced by point cloud drop sampling, the method is used to solve the error caused by airborne scanning and single-tree extraction of urban trees. Classification results showed that point cloud segmentation before intensity interpolation can effectively improve the interpolation and classification accuracies. The total classification accuracy of the data is improved by 3.7%. Compared with the extraction result (377) of single wood without subsampling treatment, the result of the urban tree extraction proved the effectiveness of the proposed method with a subsampling algorithm in improving the accuracy. Accordingly, the problem of over-segmentation is solved, and the final single-wood extraction result (329) is markedly consistent with the real situation of the region.

A New Smoke Segmentation Method Based on Improved Adaptive Density Peak Clustering

Smoke image segmentation plays a vital role in the accuracy of target extraction. In order to improve the performance of the traditional fire image segmentation algorithm, a new smoke segmentation method based on improved double truncation distance self-adaptive density peak clustering(TSDPC) is proposed. Firstly, the smoke image is over-segmented into multiple superpixels to reduce the time cost, and the local density of sample points corresponding to each superpixel is redefined by location information and color space information. Secondly, TSDPC combines the information entropy theory to find the optimal double truncation distance. Finally, TSDPC uses trigonometric functions to determine clustering centers in the decision diagram, which can solve the problem of over-segmentation. Then, it assigns labels to the remain sample points for obtaining the clustering result. Compared with other algorithms, the accuracy of TSDPC is increased by 5.68% on average, and the F1 value is increased by 6.69% on average, which shows its high accuracy and effectiveness. In public dataset, TSDPC has also demonstrated its effectiveness.

A local tangent plane distance-based approach to 3D point cloud segmentation via clustering

This paper proposes an effective measure for the planar segmentation problem based on the clustering method. It uses the distance from a point to the local plane as a metric to characterize the relationship between data. As a result, the data points of the coplanar have a high similarity to distinguish each plane. A dissimilarity matrix of the input point cloud can be evaluated, and multidimensional scaling analysis is performed to reconstruct the correlation information between data points in the 3D Euclidean space. The obtained reconstructed point cloud shows the separation between different planes. An adaptive DBSCAN clustering method based on density stratification is developed to perform cluster segmentation on the reconstructed point cloud. Experimental results show that the proposed method can effectively solve the over-segmentation problem, and at the same time provide high segmentation accuracy.

LUN-BiSeNetV2: A lightweight unstructured network based on BiSeNetV2 for road scene segmentation

With the continuous introduction of automatic driving technology, the research of road scene segmentation algorithm in machine vision has become very important. In traditional methods, most researchers use machine learning methods to segment thresholds. However, the introduction of deep learning in recent years makes convolutional neural networks widely used in this field. Aiming at the problem that the traditional threshold segmentation method is difficult to effectively extract the threshold value of road image in multiple scenes and the serious problem of over-segmentation caused by deep neural network training data directly, this paper proposes a road scene segmentation method based on a lightweight unstructured network based on BiSeNetV2. The network contains backbone segmentation network and BiSeNetV2 network. The Mobilenetv2 network is used in the backbone network to replace the Xception feature extraction network in the decoder. In addition, grouping convolution is used to replace common convolution in Mobilenetv2 network. And it selects the batch specification layer to reduce the number of parameters, without affecting the accuracy and improving the efficiency of segmentation. At the same time, due to the relatively fixed distribution position of unstructured roads in the image, attention mechanism is introduced to process advanced semantic features, so as to improve the sensitivity and accuracy of the network. The BiSeNetV2 network enhances the dominant relationship between channel features by adding a compression excitation module based on channel attention mechanism after the detail branch, so as to perceive key areas and highlight local features. The lightweight feature pyramid attention mechanism is used to optimize semantic branches, improve the feature integration between contexts, extract high-level road semantic information more efficiently and retain spatial location information to the maximum extent. Finally, local semantic features and high-level semantic features are fused to improve the effect of unstructured road detection. The experiment is trained on the open data set. The results show that compared with other state-of-theart networks, the accuracy and real-time performance of proposed LUN-BiSeNetV2 in this paper are good, and the false segmentation and edge clarity are better. Compared with the classical algorithm, the average intersection is improved by 2.2% compared with mIoU, the average pixel accuracy is improved by 7.6%, and the frame rate is improved by 24.5%.