Coursework Title
课程作业名称:
Design and Implementation of a Custom Computer Vision Model Using Classical Deep Learning Techniques with Attention Network and Explainable AI for Image Classification
使用经典深度学习技术、注意力网络和可解释的 AI 设计和实现自定义计算机视觉模型以进行图像分类
Instruction for Coursework Writing Structure:
课程作业写作结构说明：
Follow the provided article template with the following structure:
遵循提供的文章模板，其结构如下：
Abstract:
抽象：
Write a concise and informative abstract that summarizes the key aspects of your report. Include the problem statement, the proposed custom model design, the dataset used, main findings, and key evaluation metrics. The abstract should provide readers with a clear understanding of your report at a glance.
撰写一份简洁而翔实的摘要，总结报告的关键方面。包括问题陈述、建议的自定义模型设计、使用的数据集、主要发现和关键评估指标。摘要应使读者一目了然地了解您的报告。
1.0 Introduction:
1.0 简介：
Introduce the problem of image classification, the importance of XAI in developing trustworthy AI models, and attention mechanisms. State the objectives and significant contribution of your report.
介绍图像分类问题、XAI 在开发可信 AI 模型方面的重要性以及注意力机制。说明报告的目标和重要贡献。
2.0 Literature Review:
2.0 文献综述：
Review relevant literature on image classification, ensemble models, and attention mechanisms. Discuss previous works and highlight their contributions and limitations.
查看有关图像分类、集成模型和注意力机制的相关文献。讨论以前的工作并强调他们的贡献和局限性。
3.0 Methodology:
3.0 方法：
Detail the dataset used, data preprocessing steps, model architecture, attention mechanism, and the integration of XAI techniques. Provide diagrams illustrating your model architecture and design pipeline.
详细说明使用的数据集、数据预处理步骤、模型架构、注意力机制以及 XAI 技术的集成。提供说明模型架构和设计管道的图表。
3.1 Data Preprocessing:
3.1 数据预处理：
Obtain the image dataset for the image classification task. Ensure the dataset is consistent and contains diverse images from different categories.
获取图像分类任务的图像数据集。确保数据集一致并包含来自不同类别的不同图像。
Perform data preprocessing steps, including resizing, normalization, and augmentation, to prepare the data for model training and testing.
执行数据预处理步骤，包括调整大小、规范化和增强，以准备用于模型训练和测试的数据。
3.2 Data Split:
3.2 数据拆分：
Split the dataset into three subsets: training, validation, and testing sets. Use a split ratio appropriate for the dataset size.
将数据集拆分为三个子集：训练集、验证集和测试集。使用适合数据集大小的拆分比率。
3.3 Model Construction:
3.3 模型构建：
Design the Model: Create a custom image classification model incorporating classical deep learning techniques that combines residual learning, Inception learning of varying filter sizes with bottleneck strategy, and Depthwise Separable convolutional learning as taught during the lectures. You are required to build this model from scratch. The model design should reflect a thoughtful selection of layers, activation functions, and optimization strategies. You cannot use pre-trained models.
设计模型：创建一个自定义图像分类模型，该模型结合了残差学习、不同过滤器大小的 Inception 学习与瓶颈策略以及 Depthwise Separable 卷积学习，如在 讲座。您需要从头开始构建此模型。模型设计应反映对层、激活函数和优化策略的深思熟虑的选择。您不能使用预先训练的模型。
Attention Mechanism: Implement the attention mechanism from scratch to enhance the model's focus on relevant image features during classification. You cannot use existing attention mechanisms such as Squeeze-and-Excitation and CBAM.
注意力机制：从零开始实现注意力机制，以增强模型在分类过程中对相关图像特征的关注。 您不能使用现有的注意力机制，例如 Squeeze-and-Excitation 和 CBAM。
Integrate Explainable AI: Integrate XAI techniques, such as Grad-CAM, LIME, and SHAP, into your model to provide interpretability and explainability for its decisions.
集成可解释的 AI：将 XAI 技术（如 Grad-CAM、LIME 和 SHAP）集成到您的模型中，为其决策提供可解释性和可解释性。
3.4 Model Training:
3.4 模型训练：
Train the Model: Use the training dataset to train your model. Optimize hyperparameters, such as learning rate, batch size, epochs, and others, to improve model performance.
训练模型：使用训练数据集训练模型。优化超参数，例如学习率、批量大小、纪元等，以提高模型性能。
Monitor Progress: Track and record training metrics, including loss, accuracy, and training time.
监控进度：跟踪和记录训练指标，包括损失、准确率和训练时间。
Results and Discussion:
结果与讨论：
Present your experimental results with tables, figures, and plots. Interpret the results, discuss model performance, and analyze how XAI contributed to understanding model decisions.
用表格、图形和绘图来展示您的实验结果。解释结果，讨论模型性能，并分析 XAI 如何有助于理解模型决策。
4.1 Model Testing and Evaluation:
4.1 模型测试和评估：
Evaluate the Model: Test the model using the testing dataset. Evaluate the performance using metrics such as loss, accuracy, recall, precision, sensitivity, specificity, confusion matrix, F1-score, and ROC-AUC curve, precision-recall curve.
评估模型：使用测试数据集测试模型。使用损失、准确率、召回率、精度、敏感性、特异性、混淆矩阵、F1 分数和 ROC-AUC 曲线、精度-召回曲线等指标评估性能。
Apply XAI Methods: Use the XAI techniques integrated into your model to interpret the classification results. Analyze which features were most influential in making predictions and how the model's decisions can be explained.
应用 XAI 方法：使用集成到模型中的 XAI 技术来解释分类结果。分析哪些特征对进行预测最有影响，以及如何解释模型的决策。
Conclusion, Limitation and Future Work:
结论、局限性和未来工作：
Summarize your findings, discuss any limitations encountered, and propose directions for future research or model improvements.
总结您的发现，讨论遇到的任何限制，并为未来的研究或模型改进提出方向。
References:
引用：
Cite all the sources used in your report writing. The minimum number of references is 10 and should not exceed 15.
引用报告写作中使用的所有来源。最小引用数为 10 个，且不应超过 15 个。
There are two styles of citation in Computer Science: IEEE or ACM. We recommend using IEEE style. IEEE citation style is used primarily for electronics, engineering, telecommunications, software engineering, computer science, and information technology reports. The three main parts of a reference are as follows:
计算机科学有两种引用方式：IEEE 或 ACM。我们建议使用 IEEE 样式。IEEE 引文格式主要用于电子、工程、电信、软件工程、计算机科学和信息技术报告。参考的三个主要部分如下：
Author’s name listed as first initial of first name, then full last.
作者姓名列为名字的第一个首字母，然后是完整的姓氏。
Title of article, patent, conference paper, etc., in quotation marks.
文章、专利、会议论文等的标题，用引号括起来。
Title of journal or book in italics
斜体字的期刊或书籍标题
Each reference number should be enclosed in square brackets on the same line as the text, before any punctuation, with a space before the bracket.
每个参考编号应用方括号括在与文本相同的行上，在任何标点符号之前，括号前有一个空格。
Examples of in-text citation:
文内引用示例：
“. . .end of the line for my report [1].”
“. . .我的报告行结束 [1]。
“The theory was first put forward in 1987 [2].”
“该理论于 1987 年首次提出 [2]。”
“Scholtz [3] has argued. . . .” “For example, see [4].”
“朔尔茨 [3] 争论过......”“例如，参见 [4]。”
“Several recent studies [3, 4, 15, 22] have suggested that. . . .”
“最近的几项研究 [3， 4， 15， 22] 表明......
Reference
参考
[1] S. Bhanndahar. ECE 4321. Class Lecture, Topic: “Bluetooth can’t help you.” School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, Jan. 9, 2008.
[1] S. Bhanndahar.欧洲委员会 4321。课堂讲座，主题：“蓝牙帮不了你。佐治亚理工学院电气与计算机工程学院，佐治亚州亚特兰大，2008 年 1 月 9 日。
Dataset:
Dataset：
This coursework is designed for you to work independently and to ensure the uniqueness of your report in order to avoid collusion. No two students should have the same or similar CNN model. The dataset will be randomly allocated to every student, and you will have to download the dataset from the link provided to you. You must use and stick to the dataset assigned to you. Any violation of this instruction will result in a reduction of 20 points because the aim is to encourage students to work individually and to avoid collusion and cheating.
本课程旨在让您独立工作并确保报告的唯一性，以避免串通。任何两个学生都不应拥有相同或相似的 CNN 模型。数据集将随机分配给每个学生，您必须从提供给您的链接下载数据集。您必须使用并坚持使用分配给您的数据集。任何违反此指示的行为都将被扣 20 分，因为其目的是鼓励学生单独工作并避免串通和作弊。
Submission:
提交：
Submit your report following the single-column report template provided by the tutor. Include step-by-step descriptions of the tasks you performed, and the results obtained during the experiment. Ensure that your report is well-organized, clearly written, and includes all the necessary evaluation metrics and graphs as specified in the coursework requirements. The length of the report should not be less than 2000 words and MUST not exceed 5000 words for the coursework to contribute towards the development of writing skills; therefore, it is required to complete your coursework following these instructions. The submission deadline is week 12, December, 2024, by 15:30. Late submissions may incur penalties of up to 20 marks reduction, so make sure to plan your work accordingly. Failure to submit your coursework will result in a Zero Mark. In the case of exceptional circumstances, contact the Award Administrator in advance.
按照导师提供的单列报告模板提交您的报告。包括您执行的任务的分步描述，以及在实验期间获得的结果。确保您的报告组织良好、书写清晰，并包含课程作业要求中指定的所有必要的评估指标和图表。报告的长度不应少于 2000 字，且不得超过 5000 字，以便课程作业有助于写作技巧的发展;因此，需要按照这些说明完成您的课程作业。提交截止日期为 2024 年 12 月 12 周 15：30 之前。逾期提交可能会招致最高 20 分的罚款，因此请务必相应地计划您的工作。未能提交课程作业将导致零分。如遇特殊情况，请提前联系奖项管理员。
Submission Format:
提交格式：
The coursework assignment submitted should be compressed into a .zip or .rar file, the following files should be contained in the compressed file：
提交的课程作业应压缩成 .zip 或 .rar 文件，压缩文件中应包含以下文件：
A .zip file containing the report experiments: all the program’s sources, including the code, graphs, model architecture, results, and diagrams from the experiments. All plots\figures generated should be 600dpi in resolution and the labels (text in the figure) should have font-size between 20-25pt for clarity.
一个。包含报告 Experiments 的 zip 文件：程序的所有源，包括代码、图形、模型架构、结果和实验图。为清楚起见，所有生成的绘图\图的分辨率应为 600dpi，标签（图中的文本）的字体大小应在 20-25pt 之间。
Grading Criteria and Rubric:
评分标准和评分量规：
The total weight for this coursework is 50%, and you will be assessed based on the following criteria:
本课程的总权重为 50%，您将根据以下标准进行评估：
1. Abstract (3%)
1. 摘要 （3%）
Assessment: Clarity and conciseness in summarizing the problem, custom model design, dataset, main findings, and key evaluation metrics. The abstract should give a clear snapshot of the report.
评估：总结问题、自定义模型设计、数据集、主要发现和关键评估指标的清晰度和简洁性。摘要应提供报告的清晰快照。
2. Introduction (3%)
2. 简介 （3%）
Assessment: Introduction of the image classification problem, importance of Explainable AI (XAI), attention mechanisms, and clear articulation of the report’s objectives and contributions.
评估：介绍图像分类问题、可解释 AI （XAI） 的重要性、注意力机制，并清楚地阐明报告的目标和贡献。
3. Literature Review (4%)
3. 文献综述 （4%）
Assessment: Depth of literature review on image classification, ensemble models, and attention mechanisms. Evaluation of how well the review contextualizes your approach, highlighting contributions and limitations of existing works.
评估：关于图像分类、集成模型和注意力机制的文献综述深度。评估审查将您的方法置于背景中的程度，突出现有工作的贡献和局限性。
4. Methodology (15%)
4. 方法 （15%）
Data Preprocessing (2%): Correct implementation and thorough documentation of data preprocessing steps, including dataset acquisition, consistency checks, resizing, normalization, and augmentation.
数据预处理 （2%）：正确实施并全面记录数据预处理步骤，包括数据集采集、一致性检查、调整大小、规范化和增强。
Data Split (2%): Appropriateness of the dataset split into training, validation, and testing subsets, and justification of the chosen split ratio.
数据拆分 （2%）：数据集拆分为训练、验证和测试子集的适当性，以及所选拆分比率的合理性。
Model Construction (8%): Creative and sound design of the model architecture, implementing the attention mechanism, and integrating XAI techniques (e.g., Grad-CAM, LIME, SHAP).
模型构建 （8%）：模型架构的创意和合理设计，实现注意力机制，并集成 XAI 技术（例如 Grad-CAM、LIME、SHAP）。
Model Training (3%): Effective training of the model with appropriate hyperparameter optimization. Accurately report the training metrics including loss, accuracy, and training time.
模型训练 （3%）：通过适当的超参数优化对模型进行有效训练。准确报告训练指标，包括损失、准确率和训练时间。
5. Results and Discussion (10%)
5. 结果与讨论 （10%）
Model Testing and Evaluation (6%): Accurate and thorough evaluation of the model using metrics like loss, accuracy, recall, precision, sensitivity, specificity, confusion matrix, F1-score, ROC-AUC curve, and precision-recall curve. Correct application of XAI methods to interpret and analyze the model’s decisions.
模型测试和评估 （6%）：使用损失、准确率、召回率、精确率、敏感性、特异性、混淆矩阵、F1 分数、ROC-AUC 曲线和精确率召回曲线等指标对模型进行准确和全面的评估。正确应用 XAI 方法来解释和分析模型的决策。
Discussion (4%): Critical analysis and interpretation of the results, particularly focusing on how XAI techniques help in understanding the model’s decision-making process.
讨论 （4%）：对结果进行批判性分析和解释，特别关注 XAI 技术如何帮助理解模型的决策过程。
6. Conclusion, Limitation, and Future Work (3%)
6. 结论、限制和未来工作 （3%）
Assessment: Clear and concise summary of the findings, thoughtful discussion of encountered limitations, and realistic propositions for future research or model improvements.
评估：对研究结果进行清晰简洁的总结，对遇到的局限性进行深思熟虑的讨论，并为未来的研究或模型改进提出现实的主张。
7. References (3%)
7. 参考资料 （3%）
Assessment: Correct and consistent use of IEEE referencing style with appropriate and relevant citation sources, ensuring that the minimum number of references is met.
评估：正确一致地使用 IEEE 参考文献风格，并提供适当且相关的引文来源，确保满足最少的参考文献数量。
8. Presentation and Formatting (4%)
8. 表示和格式 （4%）
Assessment: Adherence to the provided report template, clarity and coherence in writing, and proper formatting of figures, tables, and diagrams. The quality of plots (600dpi) and readability of labels (font size 18-20pt) will also be evaluated.
评估：遵守提供的报告模板，写作清晰连贯，以及图形、表格和图表的正确格式。还将评估绘图的质量 （600dpi） 和标签的可读性（字体大小 18-20pt）。
9. Overall Report Quality (5%)
9. 整体报告质量 （5%）
Assessment: Overall coherence, clarity, and professionalism in the report’s writing and structure. Effectively communicates ideas and demonstrates critical thinking; showing a solid grasp of the subject matter.
评估：报告写作和结构的整体连贯性、清晰度和专业性。有效地传达想法并展示批判性思维;表现出对主题的扎实把握。
Note:
注意：
Make sure to start the coursework early, as it involves several tasks that require time and effort. Seek help from the tutor if you encounter any difficulties during the process. Good luck with your image classification experiment and report article writing!
确保尽早开始课程作业，因为它涉及几项需要时间和精力的任务。如果您在此过程中遇到任何困难，请向导师寻求帮助。祝您的图像分类实验和报告文章写作好运！
Module Name: Machine Vision
模块名称：机器视觉
Module Code: CHC6781
模块代码：CHC6781
Assessment Title: (Insert the assessment title)
评估标题：（插入评估标题）
Student Number: (Insert you student number – make sure it is correct)
学生编号：（输入您的学生编号 - 确保正确）
Word Count: (insert your total word counted excluding cover page, contents pages, reference list and appendices)
字数：（插入您的总字数，不包括封面、目录页、参考文献列表和附录）
AI Declaration:
AI 声明：
Delete as appropriate.
根据需要删除。
I have utilised / have not utilised the use of AI tool(s) in this assessment.
我在此评估中使用/未使用人工智能工具。
I have used the following AI tool(s): please provide the name of the AI tool(s) you have used and provide the exact prompt(s) you provided in
我使用了以下 AI 工具：请提供您使用的 AI 工具的名称，并提供您在
For example:
例如： AI Tool: CHAT GPT – Prompt: Find information on what are the impacts of utilising AI Tools for academic Purposes and career prospects?
AI 工具：CHAT GPT – 提示：查找有关将 AI 工具用于学术目的和职业前景有什么影响的信息？ Baidu translator: I have written the abstract, section 1, section 2, section 3, section 4, and section 5 in Chinese language and used Baidu Translate to covert these tasks to English.
百度翻译：我用中文编写了摘要、第 1 节、第 2 节、第 3 节、第 4 节和第 5 节，并使用百度翻译将这些任务转换为英文。 |
The box below:
下面的框：
If the declaration has not been made, and your tutors suspect use of AI, you will be called into do a viva voce and it will be considered academic misconduct if you fail the viva voce. This will be the same for the use of translation software which will also requires you to declare the use of.
如果尚未做出声明，并且您的导师怀疑使用 AI，您将被要求进行 viva voce，如果您未通过 viva voce，将被视为学术不端行为。这与使用翻译软件相同，也需要您声明使用。
Full disclosure will not result in an academic penalty or a lower score, so please be honest and fill in the declaration when submitting your assignment(s).
完全披露不会导致学术处罚或降低分数，因此请在提交作业时诚实并填写声明。
