Guangxin Wang

Guangxin Wang

Auto-generated profile, based on publicly available information

Guangxin Wang

Inspur Academy of Science and Technology, China

Research Focus & Metrics

Academic Career Length

4Years (2023-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
24 May 2024
Verified
Conference Paper
Open access
GWGuangxin Wang
Guangxin Wang
XLXue Li
Xue Li
DGDongshen Guo
Dongshen Guo
QDQiang Duan
Qiang Duan
QCQibin Chen
Qibin Chen
ZWZizhong Wei
Zizhong Wei
Identifying and preventing unsafe behaviors is an important task to protect people’s lives and property. Due to the complexity of behavioral types and the fact that many behavioral actions have small amplitude, rendering ineffective spatial and temporal variation information, zero-shot image-based detection of unsafe behaviors is a highly challenging task. The multimodal large language model shows excellent performance in the zero-shot task once it appears. However, it is often impossible to accurately locate and analyze each individual’s behavior in complex scenes solely relying on its ability. Therefore, this paper proposes an innovative zero-shot method to detect unsafe behaviors. We built a two-stage detection model. First, we focus on the target area of people through the target detection algorithm, and then analyze the behavior through the multimodal large language model combined with the construction strategy of positive and negative description prompts. This method shows excellent performance on the dataset we built. It can accurately identify unsafe behaviors in the real environment, which is very important to protect public safety and property.

17 May 2024
Verified
Conference Paper
Open access
ZWZizhong Wei
Zizhong Wei
QZQilai Zhang
Qilai Zhang
QDQiang Duan
Qiang Duan
GWGuangxin Wang
Guangxin Wang
RLRui Li
Rui Li
KJKai Jiang
Kai Jiang
In real-world user interactions, complex problems often involve multiple intents, such as retrieval, RAG (Retrieval-Augmented Generation), and generation. However, existing methods such as ReAct and COT (Chain-of-Thought) are designed for single tasks, resulting in a gap between research and practical applications. To bridge this gap, we propose a novel complex problem-solving framework called TAR (Think-Action-Reflection). The framework consists of three key stages: the Think stage conducts fine-grained, multi-dimensional analysis of the problem to form a structured problem representation; the Action stage dynamically schedules execution tools based on the output of the Think stage, flexibly adapting to the solving requirements of different problems; and the Reflection stage assesses the alignment between the outputs from the Action stage and the initial user intent, initiating iterative enhancements where necessary. To verify the effectiveness of the TAR framework, we conducted experiments on two types of datasets. Firstly, we constructed a business scenario dataset to validate the effectiveness of the proposed method in addressing real-world problems. Secondly, we evaluated the effectiveness of the Think stage using open-domain question answering datasets such as HotpotQA, TriviaQA, and Natural Questions (NQ). The results demonstrate that, compared with existing methods, the TAR framework achieves significant performance improvements across various complex problem tasks, exhibiting advantages in problem understanding, task execution, and result optimization. These findings confirm the effectiveness of the TAR framework as a paradigm for complex problem-solving in real-world user interactions.

22 December 2023
Verified
Conference Paper
Open access
KJKai Jiang
Kai Jiang
TYTong Yang
Tong Yang
XLXue Li
Xue Li
ZWZizhong Wei
Zizhong Wei
QCQibin Chen
Qibin Chen
RLRui Li
Rui Li
Despite significant advancements in document recognition software catering to the needs of smart office environments, software in the card recognition domain remains scarce, often limited to single-function capabilities, with each software variant designed to recognize only one type of card. Card recognition confronts substantial challenges due to the diversity of card types, complexity of layouts, and the lower quality of images captured in natural settings. To address these challenges, this paper introduces a generalized method for structured information extraction from cards based on template matching. This method supports recognition of various common card formats and facilitates the extraction of structured data. We have incorporated an image enhancement technique based on conditional generative adversarial networks (GANs) to mitigate interference from irrelevant information and backgrounds, and to rectify text area distortions. A card positioning approach, predicated on anchor point matching, employs a flexible mechanism to align feature points between target and template images. Character recognition of card contents is executed using a convolutional neural network adapted for local adjustments, augmented by optimization steps including table detection, offset correction, and automated character rectification, thereby enhancing the quality of the outputted card structure information. Tests on cards collected from a multitude of natural scenes demonstrate that our generalized card recognition method can handle multiple card types simultaneously and maintain high matching and recognition rates. It exhibits strong practicality, especially in handling cards with complex features.