Automated Defect Report Generation for Enhanced Industrial Quality Control
Authors: Jiayuan Xie, Zhiping Zhou, Zihan Wu, Xinting Zhang, Jiexin Wang, Yi Cai, Qing Li
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct a thorough evaluation of multiple mainstream methods for image description as an initial benchmark on our datasets... Experimental results show that our model design meets the requirements of two characteristics to a certain extent, i.e., capture valid knowledge for report generation. |
| Researcher Affiliation | Academia | 1Department of Computing, The Hong Kong Polytechnic University, Hong Kong SAR, China 2School of Software Engineering, South China University of Technology, Guangzhou, China 3Key Laboratory of Big Data and Intelligent Robot (SCUT), MOE of China 4 Department of Mathematics, The University of Hong Kong, Hong Kong SAR, China |
| Pseudocode | No | The paper describes the model architecture and mathematical formulations but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | No | No explicit statement about releasing source code or a link to a code repository was found. |
| Open Datasets | Yes | Thus, we select to construct NEU-r and MVTec-r datasets based on the NEU (Bao et al. 2021) and MVTec (Bergmann et al. 2019) datasets, respectively. |
| Dataset Splits | No | The paper describes training and test sets for different scenarios (e.g., zero-shot) but does not explicitly mention a separate validation set for hyperparameter tuning with specific split percentages or counts. |
| Hardware Specification | Yes | We implement all models in Pytorch and train them with two P100 GPUs. |
| Software Dependencies | No | The paper mentions "Pytorch" but does not provide specific version numbers for any software libraries or dependencies. |
| Experiment Setup | Yes | We divide each image into K = 196 patches for CLIP and Visual BERT. The decoder employs 350 hidden units, and dropout layers with a dropout probability of Pdrop=0.4. During the training process, we fine-tune the model s performance by minimizing the cross-entropy loss function. This optimization is achieved using the gradient descent algorithm with the Adam optimizer (Kingma and Ba 2015), initialized with a learning rate of 0.0001. |