Yudi Ruan, Hao Ma, Weikai Li, Xiao Wang

Low-light image enhancement (LLIE) is critical in computer vision. Existing LLIE methods often fail to discover the underlying relationships between different sub-components, causing the loss of complementary information between multiple modules and network layers, ultimately resulting in the loss of image details. To beat this shortage, we design a hierarchical mutual Enhancement via a Cross Attention transformer (ECAFormer), which introduces an architecture that enables concurrent propagation and interaction of multiple features. The model preserves detailed information by introducing a Dual Multi-head self-attention (DMSA), which leverages visual and semantic features across different scales, allowing them to guide and complement each other. Besides, a Cross-Scale DMSA block is introduced to capture the residual connection, integrating cross-layer information to further enhance image detail. Experimental results show that ECAFormer reaches competitive performance across multiple benchmarks, yielding nearly a 3% improvement in PSNR over the suboptimal method, demonstrating the effectiveness of information interaction in LLIE.

[Paper]

Yudi Ruan*, Ling Zhang*, Liling Peng*, Weikai Li, Xin Gao
* Co-first authorship

Major depressive disorder (MDD) poses a significant challenge to global mental health, necessitating the development of sophisticated diagnostic tools for its early and accurate detection. Currently, rs-fMRI has attracted considerable attention in evaluating MDD through functional brain networks (FBNs). The advent of graph convolutional networks (GCNs) has revolutionized the analysis of FBNs by capturing the complex interregional connection patterns that underlie neurological disorders, including MDD. However, existing GCN-based methodologies have predominantly concentrated on the examination of FBNs through a single topological structure, thereby neglecting the rich, multifaceted information encoded within various connection patterns. This limitation hinders the full realization of GCNs' potential in MDD diagnosis. To address this, we propose the Multiple Functional Connection Pattern Graph Convolutional Network (MFCP), an innovative framework that integrates three distinct connection patterns—Sparse Representation (SR), Partial Correlation (PC), and Granger Causality Mapping (GCM)—to harness the synergistic insights they provide. Our preliminary investigation integrates multiple graph convolutional modules to amalgamate diverse connection information, thereby enriching the MDD diagnostic features extracted from FBNs. To evaluate the proposed MFCP, we conduct experiment on the REST-MDD dataset with 533 subjects. The experimental results indicate that our MFCP attained an accuracy rate of 87.74% and an AUC score of 0.9326, confirming the effectiveness of our MFCP.

Huaiqu Feng, Yudi Ruan, Te Xi, Yulei Pan, Dongdong Du, Yongwei Wang

This research presents a comprehensive study on the contour mapping of rice canopies for Agricultural Unmanned Ground Vehicles (Agri-UGVs) operating in complex, unknown hybrid rice fields. Accurate Mapping is crucial for these vehicles to plan their operational areas efficiently while steering clear of protected zones. The motion equations that underpin the control of Agri-UGVs tasked with intelligent impurity removal heavily rely on the precise measurements of hybrid rice height and canopy contour. Therefore, the accurate estimation of these parameters is vital for the effective perception and operation of Agricultural robots. The Agri-UGV, equipped with advanced RGB-D depth sensors, facilitates the collection of distance measurements and the perception of the surrounding hybrid rice canopy's geometrical structure. This introduces an innovative mapping estimation and algorithm tailored for the hybrid rice canopy contour, utilizing a combination of kinematic and inertial measurements to achieve proprioceptive localization of the Agri-UGV. The algorithm incorporates an inference framework for grid mapping, which leverages real-time RGB-D data to iteratively refine the probabilistic distributions of the canopy's morphological properties. This is achieved by correlating the hybrid rice height intercept with in-field data while accounting for drift and uncertainty in the state estimation process. The result is a set of probabilistic contour estimates formatted as grid-based elevation maps. The proposed algorithm has been designed to operate in real-time within the ROS-Jetson system, and its performance is rigorously evaluated using metrics such as relative and absolute pose error. The algorithm has been successfully deployed on a High- clearance Agri-UGV platform, demonstrating its ability to accurately estimate rice canopy height's morphological properties in various paddy field environments. It also details the experiments conducted to explore the Agri-UGV's mapping capabilities within the cluttered hybrid rice environment. This research significantly contributes to the field of autonomous agricultural operations by enhancing the precision and reliability of canopy contour mapping, which is essential for the effective navigation and task execution of Agri-UGVs in intricate agricultural settings.

Huaiqu Feng, Yudi Ruan, Dongfang Li, Te Xi, Yulei Pan, Yongwei Wang, Jun Wang

Hybrid-rice seed production demands rapid removal of heterologous plants. We present a decentralized nUGV- 1UAV framework that couples low-altitude remote sensing with on-board swarm planning to accomplish this task in large paddy fields. A single UAV performs one-off high-resolution mapping; thereafter, multiple UGVs rely solely on the downloaded map and peer-to-peer communication to execute impurity removal. A topology-guided hybrid A* planner generates homotopy-consistent routes, while a decoupled space–time optimizer refines trajectories for curvature and collision constraints. Field experiments covering 12.7 acres with 73 impurity targets show that a fleet of six UGVs finishes the task in 1.21 h, attaining an individual UGV efficiency of 6 989 m2/h (≈10.5 acres/h). The optimal UGV-to-impurity ratio is 0.47: 5.75: 1 (UGV: impurities: acre). Simulations up to 200 acres demonstrate linear scalability with 5 % deviation from the analytical model. Even when the UAV is disabled, UGVs maintain 92 % task completion using offline maps, confirming robust decentralization.

Huaiqu Feng, Yudi Ruan, Dongfang Li, Te Xi, Yulei Pan, Yongwei Wang,

In contemporary smart agriculture, multi-machine collaboration among agricultural robotic is increasingly essential for navigating vast hybrid paddy fields. This paper presents the Energy-Constrained and Charging Station Low-altitude Remote Sensing-based Motion Planning (ECS-LRSMP), a novel approach for cooperative UGV-UAV systems in agricultural field exploration. The 1UGV-nUAV algorithm effectively addresses the challenges of energy constraints and path optimization through a five-step process: computing Voronoi Tessellations for charging station placement, solving the Hitting Set problem to select an optimal charging station set, applying a gravitational optimization algorithm to minimize UGV travel distance, and solving the Traveling Salesman Problem (TSP) for both UGV and UAV. This approach combines traditional combinatorial techniques with modern evolutionary algorithms to optimize travel distance. Simulation results demonstrate the algorithm's effectiveness in generating a coordinated plan that optimizes the performance of the hybrid UGV- UAV system in exploration scenarios. The findings highlight the significance of key parameters such as the maximum UAV flight distance 𝑅 and the number of clusters 𝛿 in designing efficient ECS-LRSMP instances. For example, increasing 𝑅 by 300% (from 3 to 9) raises UAV travel distance by 35% and UGV travel distance by 15%, establishing a positive correlation between R and total mission distance. Reducing 𝛿 from 15 to 5 cuts UGV travel distance by 28% and total mission distance by 32%, confirming a positive correlation between 𝛿 and total mission distance. Future research will explore incorporating environmental constraints, online UAV path planning, and refining the algorithm for broader applicability.

Yudi Ruan, Di Wang, Yijing Yuan, Shixin Jiang, Xianyi Yang

As the demands for ensuring bridge safety continue to rise, crack detection technology has become more crucial than ever. In this context, deep learning methods have been widely applied in the field of intelligent crack detection for bridges. However, existing methods are often constrained by complex backgrounds and computational limitations, struggling with issues such as weak crack continuity and insufficient detail representation. Inspired by biological mechanisms, a dynamic snake convolution (DSC) with tubular offsets is incorporated to tackle these challenges ef- fectively. Additionally, a channel-wise self-attention (CWSA) mechanism is introduced to efficiently fuse multi-scale features in U-Net, significantly enhancing the ability of the model to capture fine details. In the classification head, the traditional linear layer is replaced with a Kolmogorov-Arnold network (KAN) structure, which strengthens the robustness and generalization capacity of the model. Experimental results demonstrate that the proposed model im- proves detection accuracy, achieving a mean intersection over union (mIoU) of 0.877, while maintaining almost the same number of parameters, showcasing exceptional performance and practical applicability. Our project is released at https://github.com/ruanyudi/KanSeg-Bi.

Huaiqu Feng, Te Xi, Yudi Ruan, Dunhong Yang, Yulei Pan, Rongkai Shi, Bo Chen, Yongwei Wang, Jun Wang

Hybrid-rice seed production technology is a significant contributor to the high yield. The way to improve seed production's yield and quality is purification. With the functional requirements of the intelligent Undesired-rice removal process, the environmental elements of field operation and the morphological parameters of Desired-rice and Undesired-rice are collected and measured. The hybrid rice phenotypic of the multiple growth stages was collected to explore topological organization unveiling the hybrid rice in-field for variety purification. The phenological growth stages include the elongation, booting, heading, and filling stages. Expert knowledge on identifying abnormal mature varieties within hybrid rice was incorporated to ensure accuracy in variety purification. Then, the data was collected using the self-supervised learning visualization method for deep cluster analysis. The proposed method, the t-SimCLR algorithm, uses Contrastive Learning and neighbor embeddings to visualize high- dimensional data. The parametric mapping is trained from the high-dimensional pixel space into two dimensions and achieves classification accuracy by the t-SimCLR algorithm. The semantic relationships of Undesired-rice and Desired-rice data are faithfully captured. Finally, the preliminary results from the expert's prior knowledge are compared with the deep cluster analysis results to explore the topological organization of the hybrid rice in-field. Last, the integration of expert prior knowledge in the detection of panicles was showcased. The study culminates in demonstrating the integration of expert knowledge in the detection of panicles, which is crucial for accurately identifying and detecting abnormal mature varieties. This study contributes to the advancement of hybrid rice purification techniques and sets the stage for developing more sophisticated intelligent systems for agricultural applications.

Weikai Li, Hongfeng Wei, Yanlai Wu, Jie Yang, Yudi Ruan, Yuan Li, Ying Tang

Few-shot object detection (FSOD) aims to extract semantic knowledge from limited object instances of novel categories within a target domain. Recent advances in FSOD focus on fine-tuning the base model based on a few objects via meta-learning or data augmentation. Despite their success, the majority of them are grounded with parametric readjustment to generalize on novel objects, which face considerable challenges in Industry 5.0, such as 1) a certain amount of fine-tuning time is required and 2) the parameters of the constructed model being unavailable due to the privilege protection, making the fine-tuning fail. Such constraints naturally limit its application in scenarios with real-time configuration requirements or within black-box settings. To tackle the challenges mentioned above, we formalize a novel FSOD task, referred to as test-time few-shot detection (TIDE), where the model is un-tuned in the configuration procedure. To that end, we introduce an asymmetric architecture for learning a support-instance-guided dynamic category classifier. Further, a cross-attention module and a multiscale resizer are provided to enhance the model performance. Experimental results on multiple FSOD platforms reveal that the proposed TIDE significantly outperforms existing contemporary methods.

Jie Yang*, Xiaowen Xu*, Mingxiang Sun*, Yudi Ruan, Chenhao Sun, Weikai Li, Xin Gao
* Co-first authorship

Functional connectome has revealed remarkable potential in the diagnosis of neurological disorders, e.g. autism spectrum disorder. However, existing studies have primarily focused on a single connectivity pattern, such as full correlation, partial correlation, or causality. Such an approach fails in discovering the potential complementary topology information of FCNs at different connection patterns, resulting in lower diagnostic performance. Consequently, toward an accurate autism spectrum disorder diagnosis, a straight-forward ambition is to combine the multiple connectivity patterns for the diagnosis of neurological disorders. To this end, we conduct functional magnetic resonance imaging data to construct multiple brain networks with different connectivity patterns and employ kernel combination techniques to fuse information from different brain connectivity patterns for autism diagnosis. To verify the effectiveness of our approach, we assess the performance of the proposed method on the Autism Brain Imaging Data Exchange dataset for diagnosing autism spectrum disorder. The experimental findings demonstrate that our method achieves precise autism spectrum disorder diagnosis with exceptional accuracy (91.30%), sensitivity (91.48%), and specificity (91.11%).

Wenhao Zhou*, Mingxiang Sun*, Xiaowen Xu*, Yudi Ruan, Chenhao Sun, Weikai Li, Xin Gao
* Co-first authorship

The early diagnosis of autism spectrum disorder (ASD) has been extensively facilitated through the utilization of resting-state fMRI (rs-fMRI). With rs-fMRI, the functional brain network (FBN) has gained much attention in diagnosing ASD. As a promising strategy, graph convolutional networks (GCN) provide an attractive approach to simultaneously extract FBN features and facilitate ASD identification, thus replacing the manual feature extraction from FBN. Previous GCN studies primarily emphasized the exploration of topological simultaneously connection weights of the estimated FBNs while only focusing on the single connection pattern. However, this approach fails to exploit the potential complementary information offered by different connection patterns of FBNs, thereby inherently limiting the performance. To enhance the diagnostic performance, we propose a multipattern graph convolution network (MPGCN) that integrates multiple connection patterns to improve the accuracy of ASD diagnosis. As an initial endeavor, we endeavored to integrate information from multiple connection patterns by incorporating multiple graph convolution modules. The effectiveness of the MPGCN approach is evaluated by analyzing rs-fMRI scans from a cohort of 92 subjects sourced from the publicly accessible Autism Brain Imaging Data Exchange database. Notably, the experiment demonstrates that our model achieves an accuracy of 91.1% and an area under ROC curve score of 0.9742.

Sifu Zeng, Jie Yang, Wang Luo, Yudi Ruan

Establishing the relationship between a limited number of samples and segmented objects in diverse scenarios is the primary challenge in few-shot segmentation. However, many previous works overlooked the crucial support-query set interaction and the deeper information that needs to be explored. This oversight can lead to model failure when confronted with complex scenarios, such as ambiguous boundaries. To solve this problem, a duplex network that utilizes the suppression and focus concept is proposed to effectively suppress the background and focus on the foreground. Our network includes dynamic convolution to enhance the support-query interaction and a prototype match structure to fully extract information from support and query. The proposed model is called dynamic prototype mixture convolutional networks (DPMC). To minimize the impact of redundant information, we have incorporated a hybrid attentional module called double-layer attention augmented convolutional module (DAAConv) into DPMC. This module enables the network to concentrate more on foreground information. Our experiments on PASCAL-5i and COCO-20i datasets suggested that DPMC and DAAConv outperform traditional prototype-based methods by up to 5–8% on average.

Rui Zhou*, Chenhao Sun*, Mingxiang Sun*, Yudi Ruan, Weikai Li, Xin Gao
* Co-first authorship

A neurodevelopmental illness termed as the autism spectrum disorder (ASD) is described by social interaction impairments. Previous studies employing resting-state functional imaging (rs-fMRI) identified both hyperconnectivity and hypoconnectivity patterns in ASD people. However, specific patterns of connectivity within and between networks linked to ASD remain largely unexplored. Methods: We utilized a meticulously selected subset of high-quality data, comprising 45 individuals diagnosed with ASD and 47 HCs, obtained from the ABIDE dataset. The pre-processed rs-fMRI time series signals were partitioned into ninety regions of interest. We focused on eight intrinsic connectivity networks and further performed intra- and inter-network analysis. Finally, support vector machine was used to discriminate ASD from HC. Results: Through different sparsities, ASD exhibited significantly decreased intra-network connectivity within default mode network and dorsal attention network, increased connectivity between limbic network and subcortical network, and decreased connectivity between default mode network and limbic network. Using the classifier trained on altered intra- and inter-network connectivity, multivariate pattern analyses classified the ASD from HC with 71.74% accuracy, 70.21% specificity and 75.56% sensitivity in 10% sparsity of functional connectivity. Conclusions: ASD showed characteristic reorganization of the brain networks and this provided new insight into the underlying process of the functional connectome dysfunction in ASD.

It integrates NVIDIA Jetson series development kits and Intel Realsense cameras, equipped with 16 high-precision servos. The robot's body is designed and cut from carbon fiber plates. Currently, it is capable of autonomous mapping and navigation, as well as target tracking, making it suitable for exploring complex terrains and performing various tasks. The future work involves integrating LLMs to create an embodied AI.

This project is an AI-based 3D medical imaging data analysis system that utilizes state-of-the-art machine vision models. It intelligently analyzes multimodal medical imaging data, including CT and MRI scans, to automatically segment and visualize relevant areas, such as 15 abdominal organs and tumors. The project employs a Client/Server (C/S) architecture, with visualization handled on the client side via ITK, and data augmentation and model inference managed on the server side. The server side utilizes the industry-leading Kubernetes (K8S) distributed architecture for cloud computing, deploying inference engines in containers to achieve automatic scaling and fault tolerance, resulting in high performance and high availability.

It utilizes NVIDIA Jetson series development kits and Intel Realsense series cameras. It aims to measure the real-time velocity and location of urban trains through visual (monocular/stereo) and accelerometer data, constructing a visual inertial odometry system to replace traditional wheel encoders.

An integrated platform that integrates image enhancement models and high-level vision models, aimed at assisting with dataset annotation and enhancement. It uses ONNX for model deployment.

Built a customizable AI agent based on RAG technology, which helps to mitigate LLM hallucinations and accelerate the deployment of LLMs. It is inspired by NVIDIA's online course "Building RAG Agent with LLMs".