An improved YOLOv8n for tunnel lining disease ground-penetrating radar detection identificationZhang, Huaizhi; Liu, Yulang; Zhang, Xiaozhi
doi: 10.1117/12.3061932pmid: N/A
In tunnel lining disease GPR (Ground-penetrating Radar) detection identification, the manually identification are often with slow processing speeds and high workload. To improve the situation, a new tunnel lining disease identification method based on an improved YOLOv8n model is proposed. First, GprMax is used to simulate three types of tunnel lining diseases and adverse geological including multi-layer media cavity, multi-layer media water cavity, multi-layer media crack, and these are the dataset. Secondly, a new lightweight GD-YOLOv8 (GPR Disease-YOLOv8) model for tunnel lining disease GPR detection identification is introduced. Numerical results demonstrate that the proposed lightweight GD-YOLOv8 model achieves a mean average precision (mAP) of 96.4% at IoU 0.5 and a recall rate of 93.1%, outperforming classical models such as YOLOv8, YOLOv5, SSD, and Faster-RCNN. Compared to YOLOv8, GD-YOLOv8 shows 0.2% improvement in mAP0.5, 0.7% increase in precision, 46.7% reduction in parameters, and 53% decrease in computational complexity. These results indicate that the GD-YOLOv8 model not only reduces network computational costs but also provides superior detection performance.
Lightweight human action recognition system based on LSTM and deformable convolutionYin, Xu; Wang, Shiming
doi: 10.1117/12.3061169pmid: N/A
To address the challenges of insufficient spatiotemporal feature extraction and the complexity of feature design in current human action recognition, this study proposes an improved dual-stream convolutional model. The model effectively integrates the LSTM's global feature extraction capability for temporal action sequence data with the Adaptive Convolutional Network (ACNet)'s strength in spatial feature extraction. By employing Deformable Convolutional Networks (DCN) and Convolutional Neural Networks (CNN) as the backbone of the ACNet model, human skeletal spatial motion postures are extracted and fused. This approach enhances the adaptability of convolutional kernels to input data shapes while constructing a deep network. Experimental results on the publicly available DSL-46 datasets demonstrate the superiority of this method in recognizing complex and dynamic action features, achieving a recognition accuracy of 95.4%, providing a lightweight solution for the development of embedded intelligent interaction systems.
Tooth semantic segmentation method based on DPC algorithmYu, Hao; Liu, Longdu; Chen, Shuangmin; Xin, Shiqing
doi: 10.1117/12.3061281pmid: N/A
The development of 3D computer graphic technology has brought a major change to traditional medicine area, and the combination of stomatology and computational geometry has become a new point of study. The oral model of patients obtained by three-dimensional scanning usually requires segment process before used in further treatment. Most of the traditional methods are based on surface curvature, thus it relies heavily on the quality of the input mesh. To solve this, a new method based on DPC algorithm is proposed in this paper according to the common features of teeth model, where points on the boundary of tooth model are clustered to find segment points. The points are matched and connected by geodesic curve, then used to generate the semantic segmentation of tooth. Statistics show that the method of this paper is efficient, robust and extensible, even facing complex malocclusions.
Research on path tracking algorithms for intelligent vehicles in complex dynamic environmentsKang, Haonan
doi: 10.1117/12.3061824pmid: N/A
In this paper, a framework for local path planning is proposed based on discrete optimization methods, aiming to provide safe and efficient driving paths for intelligent vehicles. By constructing a road environment model, generating candidate paths and selecting the optimal path, combined with a fractional-order PID controller for path tracking, this paper realizes the accurate control of vehicle heading deviation. The experimental results show that the particle swarm optimization-based fractional-order PID controller outperforms the traditional PID controller and fractional-order PID controller in terms of path tracking accuracy, overshooting amount and response time. In addition, the path planning effectiveness experiments show that the path planning algorithm based on the artificial potential field performs well in dynamic environments with the shortest path length, the lowest number of collisions and the highest path smoothness.
Citrus leaf disease detection based on improved YOLO11 with C3K2Chen, Xiaoqin; Jiang, Ni; Yu, Zhanbo; Qian, Wei; Huang, Tao
doi: 10.1117/12.3061501pmid: N/A
Citrus planting plays an important role in China's agricultural planting. It is not only a characteristic industry with remarkable economic benefits, but also an important way to help rural revitalization and increase farmers' income. China has the largest citrus planting area in the world, but its output ranks only second, among which pest and disease infection is one of the main reasons leading to the decline of citrus output. Aiming at the problem of citrus leaf disease detection, this paper proposes a C3K2E module combining multi-scale edge information enhancement module, which is embedded into the backbone feature extraction network of YOLOv11 to optimize its disease detection ability. By improving the module design, the model not only enhances the ability to identify diseased areas, but also enhances the ability to capture small targets and edge features, achieving more efficient and accurate disease detection. The test results on the citrus leaf disease data set show that the map accuracy of the method reaches 0.952, which is 3 percentage points higher than 0.922 of the basic model YOLOv11n, significantly verifying the superiority of the proposed method.
Research on instance segmentation of agricultural greenhouses in remote sensing images with a modified Mask_RCNN algorithmLiu, Hang; Cui, Hongxia
doi: 10.1117/12.3061242pmid: N/A
Real-time recognition of agricultural plastic greenhouses is crucial for efficient management of intelligent agriculture. To improve recognition efficiency and accuracy of agricultural plastic greenhouses in remote sensing images, the ECA-MobileNetV3 is used to replace the ResNet-101 as the main feature extraction network to reduce the FLOPs and parameters of Mask R-CNN. The experiments demonstrate that the modified Mask R-CNN model reaches an accuracy of 91.80%. Meanwhile, compared to MobileNetV3-Mask R-CNN and ECA-MobileNetV2-Mask R-CNN, the FLOPs of ECA-MobileNetV3-Mask R-CNN are reduced by 1.83% and 28.37%, respectively. In addition, ECA-MobileNetV3-Mask RCNN improves inference speed by 8.38% and 10.03% compared to MobileNetV3-Mask R-CNN and ECA-MobileNetV2- Mask R-CNN, respectively. The modified model used in this paper achieves the recognition results with less FLOPs and parameters, which can be used in real-time recognition of agricultural plastic greenhouses in remote sensing images.
Landslide detection and identification based on YOLOv11 network modelZhao, Fujun; Yang, He; Gao, Lekai
doi: 10.1117/12.3061689pmid: N/A
In this paper, the landslide detection and recognition method based on YOLOv11 deep learning model is studied. Landslide disasters are seriously destructive, and traditional manual detection methods are inefficient and difficult to meet the needs of rapid response. YOLOv11 is the latest model released by Ultralytics, which improves on YOLOv8. In view of the characteristics of irregular shape, discontinuous position and variable size of landslides, this study uses instance segmentation for landslide identification and detection, which can identify the edge of the landslide area more accurately. The experiment uses the public landslide data set of the Chinese Academy of Sciences to test. The results show that YOLOv11 is superior to YOLOv8 and YOLOv9 in detection accuracy, speed and adaptability. The model not only improves the efficiency of landslide recognition, but also provides strong technical support for disaster early warning and emergency rescue.
3D human mesh recovery with learned gradientSong, Yuanyuan
doi: 10.1117/12.3061407pmid: N/A
3D human body reconstruction is a research hotspot in computer vision. In this article, we propose a method for 3D human mesh recovery with learned gradient. We use gradient descent networks to predict model parameter updates. Previous studies have focused on 3D pose estimation, often ignoring the importance of body shape reconstruction. In order to improve the accuracy of human body model shape estimation, we introduce silhouette information that can represent human body shape, using silhouette and 2D joints as inputs to the neural network. In addition, throughout the entire training process, we only need the SMPL parameterized human pose dataset, without any image to 3D correspondence. The network learns effective pose and shape subspaces from this data, and performs optimization more effectively in that subspace. It has been proven that our method can achieve advanced performance on public datasets. Even on challenging datasets in the wild, competitive results can be obtained.
An accelerated improvement of neural network pruning based on multi-layer pruning algorithmYang, Jie; Jun, Guo
doi: 10.1117/12.3061695pmid: N/A
When the network structure has a large number of layers, the progressive pruning algorithm based on multi-layer combination pruning uses the data set to verify the redundant layers of parameters and their combinations, which is a huge computational task, which is bad for the pruning speed and affects the overall efficiency. According to pruning process of MultiAdapt, about 58% of the pruning solutions used are pruning two or more layers in the same submodule. In this paper, an improved MultiAdapt algorithm named MultiAdapt V2 is proposed based on the characteristics of CNN substructure modules. During each progressive pruning process, MultiAdapt V2 considers only intra-submodule combinations when evaluating two - or multi-layer simultaneous pruning schemes. In this method, the accuracy loss is controlled to 3% in VGG-16 and ResNet with 50% sparsity, the floating-point operation is reduced to 43%, and the pruning time is reduced by 1/3.