Lightweight Model Research Articles

Lightweight design of prosthetic knee joint based on variable density method

Abstract The variable density method has emerged as an efficient approach for tackling the topology optimization challenge in continuous structures. In response to the increasing demand for lightweight prosthetic joints, a lightweight design was conducted, focusing on the front linkage rod - a crucial component of prosthetic joints. This study employed Altair Inspire software to optimize the motion-based topology of the front linkage rod. Additionally, the optimized model underwent geometric reconstruction and strength analysis. Subsequently, both the pre- and post-lightweight models were subjected to 3D printing to determine the weight reduction ratio. After confirming that the weight reduction ratio fell within the acceptable error range, a lightweight model was successfully achieved, exhibiting a remarkable weight reduction of 68.61% when compared to the initial model. This research introduces a fresh perspective on the lightweight design of prosthetic joints in motion, offering valuable insights for future endeavors in this field.

PHNNs: Lightweight Neural Networks via Parameterized Hypercomplex Convolutions.

Hypercomplex neural networks have proven to reduce the overall number of parameters while ensuring valuable performance by leveraging the properties of Clifford algebras. Recently, hypercomplex linear layers have been further improved by involving efficient parameterized Kronecker products. In this article, we define the parameterization of hypercomplex convolutional layers and introduce the family of parameterized hypercomplex neural networks (PHNNs) that are lightweight and efficient large-scale models. Our method grasps the convolution rules and the filter organization directly from data without requiring a rigidly predefined domain structure to follow. PHNNs are flexible to operate in any user-defined or tuned domain, from 1-D to [Formula: see text] regardless of whether the algebra rules are preset. Such a malleability allows processing multidimensional inputs in their natural domain without annexing further dimensions, as done, instead, in quaternion neural networks (QNNs) for 3-D inputs like color images. As a result, the proposed family of PHNNs operates with 1/n free parameters as regards its analog in the real domain. We demonstrate the versatility of this approach to multiple domains of application by performing experiments on various image datasets and audio datasets in which our method outperforms real and quaternion-valued counterparts. Full code is available at: https://github.com/eleGAN23/HyperNets.

MiniSeg: An Extremely Minimum Network Based on Lightweight Multiscale Learning for Efficient COVID-19 Segmentation.

The rapid spread of the new pandemic, i.e., coronavirus disease 2019 (COVID-19), has severely threatened global health. Deep-learning-based computer-aided screening, e.g., COVID-19 infected area segmentation from computed tomography (CT) image, has attracted much attention by serving as an adjunct to increase the accuracy of COVID-19 screening and clinical diagnosis. Although lesion segmentation is a hot topic, traditional deep learning methods are usually data-hungry with millions of parameters, easy to overfit under limited available COVID-19 training data. On the other hand, fast training/testing and low computational cost are also necessary for quick deployment and development of COVID-19 screening systems, but traditional methods are usually computationally intensive. To address the above two problems, we propose MiniSeg, a lightweight model for efficient COVID-19 segmentation from CT images. Our efforts start with the design of an attentive hierarchical spatial pyramid (AHSP) module for lightweight, efficient, effective multiscale learning that is essential for image segmentation. Then, we build a two-path (TP) encoder for deep feature extraction, where one path uses AHSP modules for learning multiscale contextual features and the other is a shallow convolutional path for capturing fine details. The two paths interact with each other for learning effective representations. Based on the extracted features, a simple decoder is added for COVID-19 segmentation. For comparing MiniSeg to previous methods, we build a comprehensive COVID-19 segmentation benchmark. Extensive experiments demonstrate that the proposed MiniSeg achieves better accuracy because its only 83k parameters make it less prone to overfitting. Its high efficiency also makes it easy to deploy and develop. The code has been released at https://github.com/yun-liu/MiniSeg.

Efficient filter pruning: Reducing model complexity through redundancy graph decomposition

Filter pruning has emerged as a crucial technique facilitating the efficient deployment of Convolutional Neural Networks (CNNs) on edge devices. The resultant lightweight models often present a challenging trade-off between performance and compression rate. Moreover, the excessive granularity of filter selection, coupled with intricate fine-tuning procedures, hampers the practical applicability of numerous pruning methodologies. To tackle these limitations, our paper introduces a novel pruning approach that initially constructs an undirected graph, effectively capturing convolutional layer redundancy by quantifying the similarity among output feature maps. Subsequently, our method iteratively eliminates highly redundant and low-value filters, employing both k-core decomposition and importance ranking criteria. In practical implementation, we adopt a one-shot pruning strategy, thereby minimizing the requirement for extensive fine-tuning. Comprehensive comparative experiments demonstrate the efficacy of our proposed approach in achieving remarkable improvements in model compression without compromising performance. For instance, with CIFAR-10, our method achieves state-of-the-art results on DenseNet-40 with a mere 70 fine-tuning epochs, while effectively removing 49.39% of FLOPs, resulting in a mere 0.01% accuracy loss. Furthermore, in the context of ImageNet, our approach excels by removing 48.81% of FLOPs and 23.71% of parameters from ResNet-50, exhibiting only 1.47% and 0.83% decline in Top-1 and Top-5 Accuracy respectively. The code is publicly available at https://github.com/lijiang99/redundancy-graph-decomposition.

Abstraction and decision fusion architecture for resource-aware image understanding with application on handwriting character classification

Resource-aware image understanding aims to achieve a soft equilibrium by effectively managing the constraints imposed by computational resources while improving the inferential capabilities of image processing systems. It serves a broad range of applications, spanning from embedded systems and IoT devices to budget-friendly smartphones and tablets. The proposed Abstraction and Decision Fusion Architecture (ADFA) addresses this problem through three tiers: abstraction, computation, and decision fusion. The first tier contains various views to generate different abstractions from the original data. These views are processed independently by an array of lightweight models forming the computation tier. They make independent decisions, where the final output is produced using a decision fusion tool in the last tier. To assess the capability of the proposed architecture, we have developed several ADFA-based models for the classification of handwriting data. In this regard, we first defined three data abstractions. Then, we trained support vector machines and fully connected neural networks according to the abstractions, which led to a set of independent basic models. Finally, an adaptive neuro-fuzzy inference system is employed as their hub to increase the accuracy by performing the decision fusion. Our experiments on the EMNIST dataset verify the efficiency and high accuracy of the proposed architecture in recognizing handwritten data, where the model size and the number of Multiply-Accumulate (MAC) operations are significantly smaller than state-of-the-art computational models.

Deep Learning-based Lightweight Models for Bactrocera Oleae Detection

The detection of Bactrocera oleae (BO), an invasive pest that poses a serious threat to olive tree cultivation, is crucial for ensuring the quality and quantity of olive oil production. This study explores the application of deep learning with lightweight models for BO detection in olive trees. The DIRT Dataset, consisting of 841 images obtained from traps, was used for training and evaluation. Different combinations of Faster R-CNN and SSD models with various backbone architectures, including EfficientNet B0, MobileNet v3, ResNet-18, and ResNet-50, were assessed. The combination of Faster R-CNN with ResNet-18 achieves comparable results to Faster R-CNN with ResNet-50 but with lower latency and a reduced parameter count. These findings highlight the potential of deep learning-based object detection models for BO detection in olive trees, offering efficient and accurate identification of infestations.

Scalable Parallel Algorithm for Graph Neural Network Interatomic Potentials in Molecular Dynamics Simulations.

Message-passing graph neural network interatomic potentials (GNN-IPs), particularly those with equivariant representations such as NequIP, are attracting significant attention due to their data efficiency and high accuracy. However, parallelizing GNN-IPs poses challenges because multiple message-passing layers complicate data communication within the spatial decomposition method, which is preferred by many molecular dynamics (MD) packages. In this article, we propose an efficient parallelization scheme compatible with GNN-IPs and develop a package, SevenNet (Scalable EquiVariance-Enabled Neural NETwork), based on the NequIP architecture. For MD simulations, SevenNet interfaces with the LAMMPS package. Through benchmark tests on a 32-GPU cluster with examples of SiO2, SevenNet achieves over 80% parallel efficiency in weak-scaling scenarios and exhibits nearly ideal strong-scaling performance as long as GPUs are fully utilized. However, the strong-scaling performance significantly declines with suboptimal GPU utilization, particularly affecting parallel efficiency in cases involving lightweight models or simulations with small numbers of atoms. We also pretrain SevenNet with a vast data set from the Materials Project (dubbed "SevenNet-0") and assess its performance on generating amorphous Si3N4 containing more than 100,000 atoms. By developing scalable GNN-IPs, this work aims to bridge the gap between advanced machine-learning models and large-scale MD simulations, offering researchers a powerful tool to explore complex material systems with high accuracy and efficiency.

APHS-YOLO: A Lightweight Model for Real-Time Detection and Classification of Stropharia Rugoso-Annulata.

The classification of Stropharia rugoso-annulata is currently reliant on manual sorting, which may be subject to bias. To improve the sorting efficiency, automated sorting equipment could be used instead. However, sorting naked mushrooms in real time remains a challenging task due to the difficulty of accurately identifying, locating and sorting large quantities of them simultaneously. Models must be deployable on resource-limited devices, making it challenging to achieve both a high accuracy and speed. This paper proposes the APHS-YOLO (YOLOv8n integrated with AKConv, CSPPC and HSFPN modules) model, which is lightweight and efficient, for identifying Stropharia rugoso-annulata of different grades and seasons. This study includes a complete dataset of runners of different grades in spring and autumn. To enhance feature extraction and maintain the recognition accuracy, the new multi-module APHS-YOLO uses HSFPNs (High-Level Screening Feature Pyramid Networks) as a thin-neck structure. It combines an improved lightweight PConv (Partial Convolution)-based convolutional module, CSPPC (Integration of Cross-Stage Partial Networks and Partial Convolution), with the Arbitrary Kernel Convolution (AKConv) module. Additionally, to compensate for the accuracy loss due to lightweighting, APHS-YOLO employs a knowledge refinement technique during training. Compared to the original model, the optimized APHS-YOLO model uses 57.8% less memory and 62.5% fewer computational resources. It has an FPS (frames per second) of over 100 and even achieves 0.1% better accuracy metrics than the original model. These research results provide a valuable reference for the development of automatic sorting equipment for forest farmers.

Real-time lightweight drone detection model: Fine-grained Identification of four types of drones based on an improved Yolov7 model

Recently, the rapid progress of artificial intelligence has enhanced the human-robot relationship through the development of several autonomous robots; such as drones. The overwhelming rise of drones has brought both relevant advantages and threatening risks into our daily lives. They are strongly deployed to achieve hard tasks and reach critical areas within a short period of time. In the past few years, drones' anarchic and malicious use has generated many incidents and accidents, causing far-reaching impacts. Due to their dual-edged nature, they are regarded as potential threats. Recognizing the type of flying drones represents a challenging task for anti-drone systems to reinforce airspace safety and security. To this end, we propose a novel lightweight model able to differentiate between the main types of drones in real-time using appropriate modules. To find the optimal compromise between size, performance and speed, a lightweight model is proposed based on Yolov7 with the incorporation of CNeB and C3C2 modules as well as the Re- Parameterization Decoupled (RePD) head structure. Further, a CNeB module is used to reduce the model’s size and improve the model’s performance, a C3C2 module to enhance the feature extraction and fusion and a Re- Parameterization Decoupled RePD to improve both the efficacy and accuracy of the prediction part while reducing the inference time. The experimental results show that the Yolov7-CNeB-C3C2-RePD meets the anti-drone requirements by improving the detection performance, minimizing the inference speed and reducing the model size. In comparison to other models, the developed has provided high results reaching 91.9% precision, 90.5% recall, 95% mAP@0.5, and 20.3 ms during the inference stage, which outperforms the other algorithms in the benchmark. Furthermore, the model has an average detection speed of 0.003 ms per image and a capacity to process 27.88 frames per second, thus satisfying the real-time requirements.

Early detection of nicosulfuron toxicity and physiological prediction in maize using multi-branch deep learning models and hyperspectral imaging

The misuse of herbicides in fields can cause severe toxicity in maize, resulting in significant reductions in both yield and quality. Therefore, it is crucial to develop early and efficient methods for assessing herbicide toxicity, protecting maize production, and maintaining the field environment. In this study, we utilized maize crops treated with the widely used nicosulfuron herbicide and their hyperspectral images to develop the HerbiNet model. After 4 d of nicosulfuron treatment, the model achieved an accuracy of 91.37 % in predicting toxicity levels, with correlation coefficient R² values of 0.82 and 0.73 for soil plant analysis development (SPAD) and water content, respectively. Additionally, the model exhibited higher generalizability across datasets from different years and seasons, which significantly surpassed support vector machines, AlexNet, and partial least squares regression models. A lightweight model, HerbiNet-Lite, exhibited significantly low complexity using 18 spectral wavelengths. After 4 d of nicosulfuron treatment, the HerbiNet-Lite model achieved an accuracy of 87.93 % for toxicity prediction and R² values of 0.80 and 0.71 for SPAD and water content, respectively, while significantly reducing overfitting. Overall, this study provides an innovative approach for the early and accurate detection of nicosulfuron toxicity within maize fields.

A digital twin-driven human-machine interactive assembly method based on lightweight multi-target detection and assembly feature generation

ABSTRACT In the field of industrial assembly, human-machine interactive assembly methods are frequently used. Lack of virtual and physical mapping, a convoluted guiding system, and low effect precision in the interactive process, a digital twin-driven human-machine interactive assembly method system is proposed as a solution to the mentioned issues. The YOLOv7-tiny lightweight model is used to perform accurate detection of parts. By incorporating attention modules into the backbone network, the feature extraction capability of the model in complicated assembly environments is enhanced. The assembly method proposed is validated using the assembly process of the reducer as an instance. The OpenCV method is employed to produce geometric reference features for parts. The experimental results show that the proposed assembly method can provide visual guidance for the assembly process, improve the traditional list-type assembly component retrieval method, solve the drawbacks of the pre-set assembly guidance in the guidance system that may not be able to adapt to the changes of the assembly results in the actual operation, and can accurately instruct novices how to assemble, which is characterised by easy implementation, low cost and high accuracy, and is of great significance for improving the success rate and assembly efficiency of human-machine interactive assembly.

Taoism-Net: A Fruit Tree Segmentation Model Based on Minimalism Design for UAV Camera

The development of precision agriculture requires unmanned aerial vehicles (UAVs) to collect diverse data, such as RGB images, 3D point clouds, and hyperspectral images. Recently, convolutional networks have made remarkable progress in downstream visual tasks, while often disregarding the trade-off between accuracy and speed in UAV-based segmentation tasks. The study aims to provide further valuable insights using an efficient model named Taoism-Net. The findings include the following: (1) Prescription maps in agricultural UAVs requires pixel-level precise segmentation, with many focusing solely on accuracy at the expense of real-time processing capabilities, being incapable of satisfying the expectations of practical tasks. (2) Taoism-Net is a refreshingly segmented model, overcoming the challenges of complexity in deep learning, based on minimalist design, which is used to generate prescription maps through pixel level classification mapping of geodetic coordinates (the lychee tree aerial dataset in Guangdong is used for experiments). (3) Compared with mainstream lightweight models or mature segmentation algorithms, Taoism-Net achieves significant improvements, including an improvement of at least 4.8% in mIoU, and manifested a superior performance in the accuracy–latency curve. (4) “The greatest truths are concise” is a saying widely spread by ancient Taoism, indicating that the most fundamental approach is reflected through the utmost minimalism; moreover, Taoism-Net expects to a build bridge between academic research and industrial deployment, for example, UAVs in precision agriculture.

A light-weight rectangular decomposition large kernel convolution network for deformable medical image registration

The performance and speed of medical image registration have been greatly boosted by advanced deep-learning based methods. However, most current methods are challenged by large deformations between input images, which necessitate a compromise in computational cost to enhance the model’s receptive field and its ability to model long-range spatial relationships for improving registration performance. In order to enhance the performance of registration for images with large deformations at a lower computational cost, in this paper, we propose a light-weight registration model with the ability to model large receptive fields and long-range spatial relationships, named LL-Net. The core components of LL-Net consist of a Rectangular Decomposition Large Kernel Attention (RD-LKA) layer and a Spatial and Channel Fusion Attention (SC-Fusion) layer. The RD-LKA layer utilizes anisotropic depth-wise large kernel convolutions to capture large receptive fields with an extremely low parameter count while modeling long-range spatial relationships. Moreover, the SC-Fusion layer enhances the model’s feature fusion capability and strengthens feature representations at critical locations. Our LL-Net exhibits state-of-the-art performance across multiple datasets. Specifically, it achieves a Dice score of 76.7% and an HD95 of 2.983 mm on the IXI dataset, and a Dice score of 87.8% and an HD95 of 1.042 mm on the OASIS dataset. Experimental results substantiate the efficacy of LL-Net in capturing large receptive fields and modeling long-range spatial relationships. The code for LL-Net is available at https://github.com/BoyOfChu/LL_Net.

A Lightweight Model to Skin Disease Recognition

A Lightweight Model to Skin Disease Recognition

LitefusionNet: Boosting the performance for medical image classification with an intelligent and lightweight feature fusion network

Medical image analysis plays a crucial role in modern healthcare for accurate diagnosis and treatment. However, the inherent challenges and limitations posed by the complexity and variability of medical images, coupled with the shortcomings of existing methods, necessitate the development of novel approaches. In this study, we propose LiteFusionNet (Lightweight Fusion Network), a lightweight model that effectively addresses these challenges, offering the advantage of accurate and efficient medical image classification while mitigating the computational demands. The LitefusionNet leverages the power of deep convolutional neural networks (DCNNs) and feature fusion techniques to achieve improved performance in medical image classification. LitefusionNet combines the strengths of MobileNet and MobileNetV2 architectures to extract robust features from medical images. These features capture discriminative information from different levels of abstraction, enhancing the model's ability to capture fine-grained patterns. The fusion process employs a concatenation method to combine the extracted features, resulting in a more comprehensive representation that improves the model's classification accuracy. To evaluate the effectiveness of LitefusionNet, extensive experiments are conducted on a diverse set of publicly available medical image datasets, including brain MRI, skin, CT, X-ray, and histology. The results demonstrate that LitefusionNet outperforms several existing models in terms of classification accuracy, showcasing its efficacy in different medical imaging modalities. Furthermore, we provide interpretability to the model's predictions by performing Grad-CAM analysis, enabling insights into the important regions in the medical images that contribute to the classification decision. In addition, we compare LitefusionNet with five pre-trained models. LiteFusionNet excels in medical image classification, boasting impressive accuracies across diverse datasets: 97.33% for brain MRI, 91.11% for skin, 99.00% for CT, 98.15% for X-ray, and 92.11% for histology. These results underscore LiteFusionNet's robust and versatile performance, making it a compelling solution for accurate and efficient medical image analysis. Overall, LitefusionNet strikes a balance between accuracy, efficiency, and real-time performance. Our findings demonstrate its potential as a promising solution for accurate and efficient medical image analysis, with applications in diagnostic support systems and clinical decision-making.

Classification and identification of tea diseases based on improved YOLOv7 model of MobileNeXt

To address the issues of low accuracy and slow response speed in tea disease classification and identification, an improved YOLOv7 lightweight model was proposed in this study. The lightweight MobileNeXt was used as the backbone network to reduce computational load and enhance efficiency. Additionally, a dual-layer routing attention mechanism was introduced to enhance the model’s ability to capture crucial details and textures in disease images, thereby improving accuracy. The SIoU loss function was employed to mitigate missed and erroneous judgments, resulting in improved recognition amidst complex image backgrounds.The revised model achieved precision, recall, and average precision of 93.5%, 89.9%, and 92.1%, respectively, representing increases of 4.5%, 1.9%, and 2.6% over the original model. Furthermore, the model’s volum was reduced by 24.69M, the total param was reduced by 12.88M, while detection speed was increased by 24.41 frames per second. This enhanced model efficiently and accurately identifies tea disease types, offering the benefits of lower parameter count and faster detection, thereby establishing a robust foundation for tea disease monitoring and prevention efforts.

Surface defect detection and semantic segmentation with a novel lightweight deep neural network

Current approaches to defect detection and segmentation make essential use of machine learning methods. To develop lightweight models is one of key tasks for many defect detection and segmentation applications. In this work, we present a lightweight trilateral parallel feature extraction with multi-feature aggregation network (TriMFANet) for surface defect detection and segmentation. In TriMFANet, the top lateral is the feature-rich extraction used to capture detailed information. The other two laterals, efficient semantic feature extraction (ESFE) and reverse ESFE, leverage Hadamard product attention to jointly extract deep-level global feature information. Additionally, the MFA module employs origin-symmetric sigmoid attention to enhance deep feature information and integrates the triple features. We conducted binary defect segmentation tasks on the SD-saliency-900 and RSDDs datasets, achieving outstanding performance in both S α and E ξ . For multi-class defect detection tasks on the NEU-Seg and MSD datasets, we rank first with mIoU scores of 79.0% and 81.2% respectively. Experimental results demonstrate that our lightweight model with only 90 K parameters exhibits excellent performance.

Hot rolled steel surface defect detection and classification using an automatic ensemble approach

This study introduces an ensemble-based Deep Neural Network (DNN) model for detecting defects on steel surfaces. The method suggested in this study classifies steel surface conditions into six possible fault categories, namely, crazing, inclusion, rolled in, pitted surface, scratches, and patches. The images undergo preprocessing and extraction of features in spatial and frequency domains using image segmentation techniques such as grey level difference method (GLDM), fast Fourier Transform (FFT), grey level co-occurrence matrix (GLCM), texture analysis and discrete wavelet transform (DWT). The ensembling of image features into a fused feature pool is carried out after the preprocessing of input images that are provided as input to a light-weight neural network model for training and testing. The performance of the model is comprehensively evaluated via an ablation study both before and after ensembling. In addition, the model capability is effectively analyzed using receiver operating characteristics (ROC) curve, confusion matrix from which classification accuracy of the model could be obtained and other parameters including precision and f1-score. It was observed that the proposed deep learning network presents phenomenally high accuracy of 99.72% for detection and classification of steel surface faults. This result was found to be superior when compared with the performance of the same neural network over each feature type individually. This study also compares the classification results of the model built based on the ensembled feature set with the results of various other classification approaches available in literature. The ensemble-based model could potentially be integrated into existing inspection systems for real-time, efficient and robust condition monitoring of steel surfaces.

Intelligent Detection of Muskmelon Ripeness in Greenhouse Environment Based on YOLO-RFEW

Accurate detection of muskmelon fruit ripeness is crucial to ensure fruit quality, optimize picking time, and enhance economic benefits. This study proposes an improved lightweight YOLO-RFEW model based on YOLOv8n, aiming to address the challenges of low efficiency in muskmelon fruit ripeness detection and the complexity of deploying a target detection model to a muskmelon picking robot. Firstly, the RFAConv replaces the Conv in the backbone part of YOLOv8n, allowing the network to focus more on regions with significant contributions in feature extraction. Secondly, the feature extraction and fusion capability are enhanced by improving the C2f module into a C2f-FE module based on FasterNet and an Efficient Multi-Scale attention (EMA) mechanism within the lightweight model. Finally, Weighted Intersection over Union (WIoU) is optimized as the loss function to improve target frame prediction capability and enhance target detection accuracy. The experimental results demonstrate that the YOLO-RFEW model achieves high accuracy, with precision, recall, F1 score, and mean Average Precision (mAP) values of 93.16%, 83.22%, 87.91%, and 90.82%, respectively. Moreover, it maintains a lightweight design and high efficiency with a model size of 4.75 MB and an inference time of 1.5 ms. Additionally, in the two types of maturity tests (M-u and M-r), APs of 87.70% and 93.94% are obtained, respectively, by the YOLO-RFEW model. Compared to YOLOv8n, significant improvements in detection accuracy have been achieved while reducing both model size and computational complexity using the proposed approach for muskmelon picking robots’ real-time detection requirements. Furthermore, when compared to lightweight models such as YOLOv3-Tiny, YOLOv4-Tiny, YOLOv5s, YOLOv7-Tiny, YOLOv8s, and YOLOv8n, the YOLO-RFEW model demonstrates superior performance with only 28.55%, 22.42%, 24.50%, 40.56%, 22.12%, and 79.83% of their respective model sizes, respectively, while achieving the highest F1 score and mAP values among these seven models. The feasibility and effectiveness of our improved scheme are verified through comparisons between thermograms generated by YOLOv8n and YOLO-RFEW as well as detection images. In summary, the YOLO-RFEW model not only improves the accuracy rate of muskmelon ripeness detection but also successfully realizes the lightweight and efficient performance, which has important theoretical support and application value in the field of muskmelon picking robot development.

Unstructured Road Extraction in UAV Images based on Lightweight Model

There is no unified planning standard for unstructured roads, and the morphological structures of these roads are complex and varied. It is important to maintain a balance between accuracy and speed for unstructured road extraction models. Unstructured road extraction algorithms based on deep learning have problems such as high model complexity, high computational cost, and the inability to adapt to current edge computing devices. Therefore, it is best to use lightweight network models. Considering the need for lightweight models and the characteristics of unstructured roads with different pattern shapes, such as blocks and strips, a TMB (Triple Multi-Block) feature extraction module is proposed, and the overall structure of the TMBNet network is described. The TMB module was compared with SS-nbt, Non-bottleneck-1D, and other modules via experiments. The feasibility and effectiveness of the TMB module design were proven through experiments and visualizations. The comparison experiment, using multiple convolution kernel categories, proved that the TMB module can improve the segmentation accuracy of the network. The comparison with different semantic segmentation networks demonstrates that the TMBNet network has advantages in terms of unstructured road extraction.