Author name cluster

Heng Li

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

22 papers

2 author rows

AAAI Conference 2026 Conference Paper

DeLightMono: Enhancing Self-Supervised Monocular Depth Estimation in Endoscopy by Decoupling Uneven Illumination

Mingyang Ou
Haojin Li
Yifeng Zhang
Ke Niu
Zhongxi Qiu
Heng Li
Jiang Liu

Self-supervised monocular depth estimation serves as a key task in the development of endoscopic navigation systems. However, performance degradation persists due to uneven illumination inherent in endoscopic images, particularly in low-intensity regions. Existing low-light enhancement techniques fail to effectively guide the depth network. Furthermore, solutions from other fields, like autonomous driving, require well-lit images, making them unsuitable and increasing data collection burdens. To this end, we present DeLightMono - a novel self-supervised monocular depth estimation framework with illumination decoupling. Specifically, endoscopic images are represented by a designed illumination-reflectance-depth model, and are decomposed with auxiliary networks. Moreover, a self-supervised joint-optimizing framework with novel losses leveraging the decoupled components is proposed to mitigate the effects of uneven illumination on depth estimation. The effectiveness of the proposed methods was rigorously verified through extensive comparisons and an ablation study performed on two public datasets.

PDF Details DOI

AAAI Conference 2026 Conference Paper

VisAssist: A Visually Impaired-Captured Video Question Answering Benchmark for Assistive Systems

Qi Gao
Heng Li
Yixin Zhou
Meixuan Zhou
Jieqiong Chen
Xinyu Chai

We present VisAssist, the first large-scale video question-answering dataset with 13,413 real-world videos captured by visually impaired users, addressing a critical gap in assistive vision research. Unlike existing benchmarks relying on third-person footage, VisAssist provides authentic first-person perspectives that uniquely capture challenges in blind photography—including unconventional framing, motion artifacts, and frequent information omission. Benchmark evaluations of SOTA multimodal models reveal systematic limitations: severe deficiencies in spatial reasoning when processing dynamic first-person viewpoints, an inability to distinguish missing information from poor capture quality leading to hazardous hallucinations, and fragile text understanding especially for non-Latin scripts under suboptimal conditions. This work establishes a vital real-world benchmark and underscores the need for specialized architectures in visual assistance systems.

PDF Details DOI

AAAI Conference 2025 Conference Paper

AIF-SFDA: Autonomous Information Filter Driven Source-Free Domain Adaptation for Medical Image Segmentation

Haojin Li
Heng Li
Jianyu Chen
Rihan Zhong
Ke Niu
Huazhu Fu
Jiang Liu

Decoupling domain-variant information (DVI) from domain-invariant information (DII) serves as a prominent strategy for mitigating domain shifts in the practical implementation of deep learning algorithms. However, in medical settings, concerns surrounding data collection and privacy often restrict access to both training and test data, hindering the empirical decoupling of information by existing methods. To tackle this issue, we propose an Adaptive Information Filter-driven Source-free Domain Adaptation (AIF-SFDA) algorithm, which leverages a frequency-based learnable information filter to autonomously decouple DVI and DII. Information Bottleneck (IB) and Self-supervision (SS) are incorporated to optimize the learnable frequency filter. The IB governs the information flow within the filter to diminish redundant DVI, while SS preserves DII in alignment with the specific task and image modality. Thus, the adaptive information filter can overcome domain shifts relying solely on target data. A series of experiments covering various medical image modalities and segmentation tasks were conducted to demonstrate the benefits of AIF-SFDA through comparisons with leading algorithms and ablation studies.

PDF Details DOI

EAAI Journal 2025 Journal Article

Detecting worker loss of balance events from point cloud sequence using unsupervised motion-pose learning

Mingyu Zhang
Lei Wang
Yinong Hu
Shuai Han
Jiawen Zhang
Heng Li

Workers' loss of balance (LB), such as slip and trip, may lead to severe injuries and even fatalities. Existing methods for detecting LB typically rely on wearable sensors and focus on specific body parts. This study introduces a novel, non-contact approach utilizing light detection and ranging (LiDAR) technology to detect LB events. By capturing full-body point cloud data, the proposed method extracts both static pose and dynamic motion features across multiple body sections and detects LB events through unsupervised learning. The high-dimensional point cloud sequence is transformed into interpretable gait features, enabling effective unsupervised learning through sequence reconstruction. A two-stream network and fusion strategy are also developed to combine pose and motion features for final LB detection. Experiments with various LB events demonstrate the method's effectiveness, achieving an F1 score of 0. 98 and a recall of 0. 98. Our analysis reveals that integrating features from multiple body parts and the fusion of pose and motion information significantly enhances detection performance. This study offers a promising alternative to traditional methods, providing effective, non-intrusive monitoring of worker safety in dynamic construction environments.

Details DOI

EAAI Journal 2025 Journal Article

Dynamic flame feature-driven prediction model for basic oxygen furnace steelmaking endpoint carbon content based on three-dimensional multi-layer complex networks

Jianxun Liu
Hui Liu
FuGang Chen
YunKe Su
Heng Li
Xiaojun Xue

Accurate prediction of carbon content at the endpoint is crucial for the endpoint management of Basic Oxygen Furnace (BOF) steelmaking. The carbon content in the molten pool is closely related to the dynamic and static characteristics of the flame at the furnace’s mouth. However, the flame’s texture change exhibits multidirectional and multiscale properties, posing challenges for existing algorithms to effectively extract dynamic color texture features. To address this issue, this paper proposes a dynamic texture feature extraction model based on a three-dimensional multi-layer complex network (3D-MLCN). The model constructs an unbounded complex network for a single-frame flame picture by integrating spatiotemporal position information of the image region’s centroid with color information, thereby quantizing the single-frame image into a complex network with spatiotemporal properties. Subsequently, a multi-scale multi-direction weighted dynamic color texture complex network is built for the flame video at the furnace mouth, utilizing the temporal index of the video frames in combination with vertex color values to capture the time-varying features of the flame video. The proposed method quantifies network characteristics through vertex degree distribution features to obtain dynamic color texture feature descriptors. These descriptors are then combined with static color texture features and color features to construct dynamic and static feature descriptors for the flame video, enabling the prediction of the endpoint carbon content using a regression model. By analyzing the actual production data of BOF steelmaking, the prediction accuracy of carbon content within the error range of ±0. 02% is 87. 91%, the R 2 value is 0. 8547, and the RMSE value is 2. 0959, which verifies the effectiveness of the proposed method.

Details DOI

NeurIPS Conference 2025 Conference Paper

MaxSup: Overcoming Representation Collapse in Label Smoothing

Yuxuan Zhou
Heng Li
Zhi-Qi Cheng
Xudong Yan
Yifei Dong
Mario Fritz
Margret Keuper

Label Smoothing (LS) is widely adopted to reduce overconfidence in neural network predictions and improve generalization. Despite these benefits, recent studies reveal two critical issues with LS. First, LS induces overconfidence in misclassified samples. Second, it compacts feature representations into overly tight clusters, diluting intra-class diversity, although the precise cause of this phenomenon remained elusive. In this paper, we analytically decompose the LS-induced loss, exposing two key terms: (i) a regularization term that dampens overconfidence only when the prediction is correct, and (ii) an error-amplification term that arises under misclassifications. This latter term compels the network to reinforce incorrect predictions with undue certainty, exacerbating representation collapse. To address these shortcomings, we propose Max Suppression (MaxSup), which applies uniform regularization to both correct and incorrect predictions by penalizing the top-1 logit rather than the ground-truth logit. Through extensive feature-space analyses, we show that MaxSup restores intra-class variation and sharpens inter-class boundaries. Experiments on large-scale image classification and multiple downstream tasks confirm that MaxSup is a more robust alternative to LS.

PDF Details

EAAI Journal 2025 Journal Article

Monocular three-dimensional object detection for proximity monitoring in human-machine collision warning systems on construction sites

Yuexiong Ding
Qiong Liu
Ankang Ji
Heng Li
Xiaowei Luo

Monitoring workers’ proximities to avoid struck-by hazards has aroused great concern in construction safety management. Existing methods are either too laborious and costly to apply extensively or lack spatial perception for accurate monitoring. This study proposes a novel framework for proximity monitoring using only an ordinary two-dimensional (2D) camera to realize human-machine collision warning, which integrates a monocular three-dimensional (3D) object detection model and a post-processing classification module to identify four proximity categories: Dangerous, Potentially Dangerous, Concerned, and Safe. A new dataset containing 22, 500 virtual and real-world construction images with 3D bounding box annotations has been created and publicly released to facilitate system development and evaluation. Experiments show that the implemented system is rapid-response and camera carrier-independent, achieving promising proximity detection performance on both virtual and real-world data, with mean precision, recall, and F1 scores of approximately 0. 8, 0. 7, and 0. 8, respectively, within a range of 50 meters. This study preliminarily reveals the potential and feasibility of proximity monitoring using only a 2D surveillance camera, providing a new, promising, and affordable way for early warning of human-machine collisions.

Details DOI

ICRA Conference 2025 Conference Paper

OG-Gaussian: Occupancy Based Street Gaussians for Autonomous Driving

Yedong Shen
Xinran Zhang
Yifan Duan
Shiqi Zhang
Heng Li
Yilong Wu
Jianmin Ji
Yanyong Zhang

Accurate and realistic 3D scene reconstruction enables the lifelike creation of autonomous driving simulation environments. With advancements in 3D Gaussian Splatting (3DGS), previous studies have applied it to reconstruct complex dynamic driving scenes. These methods typically require expensive LiDAR sensors and pre-annotated datasets of dynamic objects. To address these challenges, we propose OG-Gaussian, a novel approach that replaces LiDAR point clouds with Occupancy Grids (OGs) generated from surround-view camera images using Occupancy Prediction Network (ONet). Our method leverages the semantic information in OGs to separate dynamic vehicles from static street background, converting these grids into two distinct sets of initial point clouds for reconstructing both static and dynamic objects. Additionally, we estimate the trajectories and poses of dynamic objects through a learning-based approach, eliminating the need for complex manual annotations. Experiments on Waymo Open dataset demonstrate that OG-Gaussian is on par with the current state-of-the-art in terms of reconstruction quality and rendering speed, achieving an average PSNR of 35. 13 and a rendering speed of 143 FPS, while significantly reducing computational costs and economic overhead.

Details

NeurIPS Conference 2025 Conference Paper

PanTS: The Pancreatic Tumor Segmentation Dataset

Wenxuan Li
Xinze Zhou
Qi Chen
Tianyu Lin
Pedro R. A. S. Bassi
Xiaoxi Chen
Chen Ye
Zheren Zhu

PanTS is a large-scale, multi-institutional dataset curated to advance research in pancreatic CT analysis. It contains 36, 390 CT scans from 145 medical centers, with expert-validated, voxel-wise annotations of over 993, 000 anatomical structures, covering pancreatic tumors, pancreas head, body, and tail, and 24 surrounding anatomical structures such as vascular/skeletal structures and abdominal/thoracic organs. Each scan includes metadata such as patient age, sex, diagnosis, contrast phase, in-plane spacing, slice thickness, etc. AI models trained on PanTS achieve significantly better performance in pancreatic tumor detection, localization, and segmentation than those trained on existing public datasets. Our analysis indicates that these gains are directly attributable to the 16× larger-scale tumor annotations and indirectly supported by the 24 additional surrounding anatomical structures. As the largest and most comprehensive resource of its kind, PanTS offers a new benchmark for developing and evaluating AI models in pancreatic CT analysis.

PDF Details

NeurIPS Conference 2025 Conference Paper

SeePhys: Does Seeing Help Thinking? – Benchmarking Vision-Based Physics Reasoning

Kun Xiang
Heng Li
Terry Jingchen Zhang
Yinya Huang
Zirong Liu
Peixin Qu
Jixi He
Jiaqi Chen

We present SeePhys, a large-scale multimodal benchmark for LLM reasoning grounded in physics questions ranging from middle school to PhD qualifying exams. The benchmark covers 7 fundamental domains spanning the physics discipline, incorporating 21 categories of highly heterogeneous diagrams. In contrast to prior works where visual elements mainly serve auxiliary purposes, our benchmark features a substantial proportion of vision-essential problems (75%) that mandate visual information extraction for correct solutions. Through extensive evaluation, we observe that even the most advanced visual reasoning models (e. g. , Gemini-2. 5-pro and o4-mini) achieve sub-60% accuracy on our benchmark. These results reveal fundamental challenges in current large language models' visual understanding capabilities, particularly in: (i) establishing rigorous coupling between diagram interpretation and physics reasoning, and (ii) overcoming their persistent reliance on textual cues as cognitive shortcuts. Project Page: github. com/SeePhys/seephys-projectHugging Face: huggingface. co/datasets/SeePhys/SeePhys

PDF Details

AAAI Conference 2025 Conference Paper

Universal Features Guided Zero-Shot Category-Level Object Pose Estimation

Wentian Qu
Chenyu Meng
Heng Li
Jian Cheng
Cuixia Ma
Hongan Wang
Xiao Zhou
Xiaoming Deng

Object pose estimation, crucial in computer vision and robotics applications, faces challenges with the diversity of unseen categories. We propose a zero-shot method to achieve category-level 6-DOF object pose estimation, which exploits both 2D and 3D universal features of input RGB-D image to establish semantic similarity-based correspondences and can be extended to unseen categories without additional model fine-tuning. Our method begins with combining efficient 2D universal features to find sparse correspondences between intra-category objects and gets initial coarse pose. To handle the correspondence degradation of 2D universal features if the pose deviates much from the target pose, we use an iterative strategy to optimize the pose. Subsequently, to resolve pose ambiguities due to shape differences between intra-category objects, the coarse pose is refined by optimizing with dense alignment constraint of 3D universal features. Our method outperforms previous methods on the REAL275 and Wild6D benchmarks for unseen categories.

PDF Details DOI

EAAI Journal 2024 Journal Article

A hierarchical federated learning framework for collaborative quality defect inspection in construction

Hai-Tao Wu
Heng Li
Hung-Lin Chi
Wei-Bin Kou
Yik-Chung Wu
Shuai Wang

Recent advancements in robotics and deep learning (DL) have made it possible to implement robots in civil infrastructures' quality defect inspection. Robots can reduce human inspectors' workloads and enhance inspection results' reliability by collecting data and automatically identifying quality defects from the raw data. However, current methods for training DL models rely on centralized strategies that require the aggregation of defect data (e. g. , uploading to a cloud server), posing concerns about data privacy and security. Thus, this study proposes a three-fold federated learning (FL) framework for training DL models collaboratively, without the need to share local data among construction robots. The framework is specifically applied to image-based crack segmentation, critical for ensuring infrastructures’ safety and serviceability. A lightweight DL model is developed to enable easy implementation on resource-constrained construction robots and to reduce communication costs during federated training. Experimental results show that the proposed FL method outperforms traditional centralized methods. The critical contribution of this study is the hierarchical FL framework, which enables construction robots to leverage big data in a privacy-preserving manner.

Details DOI

NeurIPS Conference 2024 Conference Paper

Human-Aware Vision-and-Language Navigation: Bridging Simulation to Reality with Dynamic Human Interactions

Heng Li
Minghan Li
Zhi-Qi Cheng
Yifei Dong
Yuxuan Zhou
Jun-Yan He
Qi Dai
Teruko Mitamura

Vision-and-Language Navigation (VLN) aims to develop embodied agents that navigate based on human instructions. However, current VLN frameworks often rely on static environments and optimal expert supervision, limiting their real-world applicability. To address this, we introduce Human-Aware Vision-and-Language Navigation (HA-VLN), extending traditional VLN by incorporating dynamic human activities and relaxing key assumptions. We propose the Human-Aware 3D (HA3D) simulator, which combines dynamic human activities with the Matterport3D dataset, and the Human-Aware Room-to-Room (HA-R2R) dataset, extending R2R with human activity descriptions. To tackle HA-VLN challenges, we present the Expert-Supervised Cross-Modal (VLN-CM) and Non-Expert-Supervised Decision Transformer (VLN-DT) agents, utilizing cross-modal fusion and diverse training strategies for effective navigation in dynamic human environments. A comprehensive evaluation, including metrics considering human activities, and systematic analysis of HA-VLN's unique challenges, underscores the need for further research to enhance HA-VLN agents' real-world robustness and adaptability. Ultimately, this work provides benchmarks and insights for future research on embodied AI and Sim2Real transfer, paving the way for more realistic and applicable VLN systems in human-populated environments.

PDF Details DOI

ICRA Conference 2024 Conference Paper

OCC-VO: Dense Mapping via 3D Occupancy-Based Visual Odometry for Autonomous Driving

Heng Li
Yifan Duan
Xinran Zhang
Haiyi Liu
Jianmin Ji
Yanyong Zhang

Visual Odometry (VO) plays a pivotal role in autonomous systems, with a principal challenge being the lack of depth information in camera images. This paper introduces OCC-VO, a novel framework that capitalizes on recent advances in deep learning to transform 2D camera images into 3D semantic occupancy, thereby circumventing the traditional need for concurrent estimation of ego poses and landmark locations. Within this framework, we utilize the TPV-Former to convert surround view cameras’ images into 3D semantic occupancy. Addressing the challenges presented by this transformation, we have specifically tailored a pose estimation and mapping algorithm that incorporates Semantic Label Filter, Dynamic Object Filter, and finally, utilizes Voxel PFilter for maintaining a consistent global semantic map. Evaluations on the Occ3D-nuScenes not only showcase a 20. 6% improvement in Success Ratio and a 29. 6% enhancement in trajectory accuracy against ORB-SLAM3, but also emphasize our ability to construct a comprehensive map. Our implementation is open-sourced and available at: https://github.com/USTCLH/OCC-VO.

Details

EAAI Journal 2024 Journal Article

Semi-supervised anomaly detection with contamination-resilience and incremental training

Liheng Yuan
Fanghua Ye
Heng Li
Chenhao Zhang
Cuiying Gao
Chengqing Yu
Wei Yuan
Xinge You

Anomaly detection plays a vital role in various realistic applications, including fraud detection, network traffic analysis, medical diagnosis, and so on. Semi-supervised anomaly detection methods have recently attracted increasing attention, owing to their low requirement for labeled anomalous samples. However, existing semi-supervised methods suffer from performance degradation when training data are contaminated with anomalies, and cannot well support incremental training required in scenarios where original training data are hard to obtain. To overcome these limitations, we propose SAE-CRIT, a lightweight semi-supervised anomaly detection method with contamination resilience and incremental training. SAE-CRIT effectively mitigates the negative impact of contaminated data through differentially weighting samples, and leverages a three-layer neural network to detect anomalies, allowing for efficient incremental training by updating only the last layer with new data. We compare SAE-CRIT with eight anomaly detection methods over four datasets. Extensive experiments demonstrate the advantages of SAE-CRIT in contamination resistance, incremental training, and training costs. More specifically, the state-of-the-art detection method GOAD achieved an F1-score of 89. 3% and 90. 6% on the contaminated datasets KDDCUP and KDDCUP-Rev, respectively. Under the same settings, however, SAE-CRIT exhibited an F1-score of 92. 4% and 96. 9%, respectively. In addition, the training time of SAE-CRIT is less than 20 s on these two datasets. The time spent by SAE-CRIT on these two datasets only accounts for 0. 26% and 1. 8% of the total time spent by GOAD, respectively.

Details DOI

AAAI Conference 2024 Conference Paper

TDeLTA: A Light-Weight and Robust Table Detection Method Based on Learning Text Arrangement

Yang Fan
Xiangping Wu
Qingcai Chen
Heng Li
Yan Huang
Zhixiang Cai
Qitian Wu

The diversity of tables makes table detection a great challenge, leading to existing models becoming more tedious and complex. Despite achieving high performance, they often overfit to the table style in training set, and suffer from significant performance degradation when encountering out-of-distribution tables in other domains. To tackle this problem, we start from the essence of the table, which is a set of text arranged in rows and columns. Based on this, we propose a novel, light-weighted and robust Table Detection method based on Learning Text Arrangement, namely TDeLTA. TDeLTA takes the text blocks as input, and then models the arrangement of them with a sequential encoder and an attention module. To locate the tables precisely, we design a text-classification task, classifying the text blocks into 4 categories according to their semantic roles in the tables. Experiments are conducted on both the text blocks parsed from PDF and extracted by open-source OCR tools, respectively. Compared to several state-of-the-art methods, TDeLTA achieves competitive results with only 3.1M model parameters on the large-scale public datasets. Moreover, when faced with the cross-domain data under the 0-shot setting, TDeLTA outperforms baselines by a large margin of nearly 7%, which shows the strong robustness and transferability of the proposed model.

PDF Details DOI

EAAI Journal 2023 Journal Article

A fast end-to-end method for automatic interior progress evaluation using panoramic images

Xin Fang
Heng Li
Haitao Wu
Lang Fan
Ting Kong
Yue Wu

Interior construction makes up a large portion of project budget and time and is more prone to schedule delays. Most research efforts on progress management focus on exterior environment, while few on interior construction. Although progress monitoring methods based on laser point clouds and computer vision are investigated before, the problems of costly acquisition and creation of point clouds and images are still open, which impede the study of progress evaluation, particularly in interior construction environments where clutters and occlusions are universal. This paper introduces a method based on 360° panoramic images and deep learning for fast end-to-end interior progress evaluation in room units. The method takes only one or two 360° panoramic images as input, estimates key corners, generates and registers room layouts, and semantically segments sparse point cloud. With the extracted corners and segmentation results, the progress states of interior trades can be evaluated. The experimental results show that the proposed method based on deep learning techniques achieves comparable performance against those on public data sets with 3D Intersection over Union (3D IoU) of 83. 69% vs 84. 23%, Corner Error (CE) of 0. 4% vs 0. 69%, and mean class Interaction over Union (mIoU) of 70. 28% vs 53. 5%. A case study of an interior decoration project of a hotel is adopted to demonstrate the feasibility and practical capabilities of the proposed method.

Details DOI

EAAI Journal 2023 Journal Article

Integration of ROV and vision-based underwater inspection for Limnoperna fortunei in water conveyance structure

Xin Fang
Heng Li
Sherong Zhang
Jikang Zhang
Chao Wang
Xiaohua Wang
Ziao Ma
He Jia

The invasion of Limnoperna fortunei (L. fortunei) has been identified as one major biofouling in the operation of hydraulic engineering, which not only corrodes the concrete structures but also reduces the pipe diameter and increases the surface roughness, leading to the decrease of water conveyance capacity and the increase of the project operation cost. To better cope with this problem, an automated underwater inspection analysis scheme for the biofouling of L. fortune is provided in this study, which innovatively integrates the underwater remote operating rover (ROV) and computer vision techniques to inspect and evaluate the invasion of L. fortunei in water conveyance structure. This scheme first presents an image enhancement approach based on the fusion strategy to improve the quality of images extracted from underwater robot inspection videos. Then, the L. fortunei is segmented by U-Net in the enhanced underwater images, and the definition of adherent area ratio quantitatively assesses the biofouling severity. At last, the underwater inspection analysis scheme is implemented in a typical aqueduct, and the automatic analysis results are compared with the field investigation during the emptying maintenance of the aqueduct. In this study, the dataset of real ROV inspection video sequences was first used to evaluate the effectiveness of the proposed method for inspecting L. fortunei invasions in realistic scenarios, and then for the comparison with state-of-the-art methods. The results show that the proposed automated inspection scheme is capable of efficiently improving the underwater imaging quality and accurately detecting the L. fortunei.

Details DOI

JBHI Journal 2023 Journal Article

Multi-Learner Based Deep Meta-Learning for Few-Shot Medical Image Classification

Hongyang Jiang
Mengdi Gao
Heng Li
Richu Jin
Hanpei Miao
Jiang Liu

Few-shot learning (FSL) is promising in the field of medical image analysis due to high cost of establishing high-quality medical datasets. Many FSL approaches have been proposed in natural image scenes. However, present FSL methods are rarely evaluated on medical images and the FSL technology applicable to medical scenarios need to be further developed. Meta-learning has supplied an optional framework to address the challenging FSL setting. In this paper, we propose a novel multi-learner based FSL method for multiple medical image classification tasks, combining meta-learning with transfer-learning and metric-learning. Our designed model is composed of three learners, including auto-encoder, metric-learner and task-learner. In transfer-learning, all the learners are trained on the base classes. In the ensuing meta-learning, we leverage multiple novel tasks to fine-tune the metric-learner and task-learner in order to fast adapt to unseen tasks. Moreover, to further boost the learning efficiency of our model, we devised real-time data augmentation and dynamic Gaussian disturbance soft label (GDSL) scheme as effective generalization strategies of few-shot classification tasks. We have conducted experiments for three-class few-shot classification tasks on three newly-built challenging medical benchmarks, BLOOD, PATH and CHEST. Extensive comparisons to related works validated that our method achieved top performance both on homogeneous medical datasets and cross-domain datasets.

Details DOI

AIIM Journal 2022 Journal Article

Intelligent and strong robust CVS-LVAD control based on soft-actor-critic algorithm

Te Li
Wenbo Cui
Nan Xie
Heng Li
Haibo Liu
Xu Li
Yongqing Wang

Left ventricular assist device (LVAD) is an effective method to treat ventricular failure. According to the physiological conditions of different patients, the device adaptively adjusts its rotation speed to change LVAD output. In this study, a physiological control system for LVAD based on deep reinforcement learning (DRL) is proposed. The system estimates the amount of blood required by LVAD based on a Starling-like method. The DRL controller regulates LVAD to adjust the speed and quickly approach the target value. The changes of vascular resistance, myocardial contractility, and the transition from rest to exercise were simulated, and the single factor and mixed factor experiments were carried out to compare the effects of DRL controller and proportional integral derivative (PID) controller, which controls the system according to the difference between measured variables and expected values. Two metrics are used to illustrate the regulation effect: the sum of absolute error (SAE) and the response time of the two controllers, where SAE is the difference between the estimated required pumped blood flow LVADQ e and the actual measured blood flow LVADQ m. The experimental result shows that the SAE of the DRL controller is 47. 6% of that of the PID controller, and the response time of the DRL controller is 38. 6% of that of the PID controller. This study demonstrates that the LVAD based on the DRL controller can respond more quickly and more effectively to the different physiological needs of a variety of patients than a PID controller.

Details DOI

IJCAI Conference 2022 Conference Paper

Recent Advances in Concept Drift Adaptation Methods for Deep Learning

Liheng Yuan
Heng Li
Beihao Xia
Cuiying Gao
Mingyue Liu
Wei Yuan
Xinge You

In the ``Big Data'' age, the amount and distribution of data have increased wildly and changed over time in various time-series-based tasks, e. g weather prediction, network intrusion detection. However, deep learning models may become outdated facing variable input data distribution, which is called concept drift. To address this problem, large number of samples are usually required to update deep learning models, which is impractical in many realistic applications. This challenge drives researchers to explore the effective ways to adapt deep learning models to concept drift. In this paper, we first mathematically describe the categories of concept drift including abrupt drift, gradual drift, recurrent drift, incremental drift. We then divide existing studies into two categories (i. e. , model parameter updating and model structure updating), and analyze the pros and cons of representative methods in each category. Finally, we evaluate the performance of these methods, and point out the future directions of concept drift adaptation for deep learning.

PDF Details DOI

AIIM Journal 2018 Journal Article

Image processing strategies based on saliency segmentation for object recognition under simulated prosthetic vision

Heng Li
Xiaofan Su
Jing Wang
Han Kan
Tingting Han
Yajie Zeng
Xinyu Chai

Background and objective Current retinal prostheses can only generate low-resolution visual percepts constituted of limited phosphenes which are elicited by an electrode array and with uncontrollable color and restricted grayscale. Under this visual perception, prosthetic recipients can just complete some simple visual tasks, but more complex tasks like face identification/object recognition are extremely difficult. Therefore, it is necessary to investigate and apply image processing strategies for optimizing the visual perception of the recipients. This study focuses on recognition of the object of interest employing simulated prosthetic vision. Method We used a saliency segmentation method based on a biologically plausible graph-based visual saliency model and a grabCut-based self-adaptive-iterative optimization framework to automatically extract foreground objects. Based on this, two image processing strategies, Addition of Separate Pixelization and Background Pixel Shrink, were further utilized to enhance the extracted foreground objects. Results i) The results showed by verification of psychophysical experiments that under simulated prosthetic vision, both strategies had marked advantages over Direct Pixelization in terms of recognition accuracy and efficiency. ii) We also found that recognition performance under two strategies was tied to the segmentation results and was affected positively by the paired-interrelated objects in the scene. Conclusion The use of the saliency segmentation method and image processing strategies can automatically extract and enhance foreground objects, and significantly improve object recognition performance towards recipients implanted a high-density implant.

Details DOI