EDPS 50th Game_KV.jpeg

This project aims to develop an integrated approach to control and predict blood glucose levels in patients with type 1 diabetes (T1D) by combining an insulin delivery method with advanced artificial intelligence (AI) algorithms. The project consists of two main parts: (1) the implementation of an insulin delivery system using a proportional-integral-derivative (PID) controller, and (2) the utilization of continuous glucose monitoring (CGM) data and AI algorithms for predicting future blood glucose levels, a capability beyond the scope of the PID controller. This combined system offers personalized treatment and with prediction capability it can give early warning to medical staff to manage T1D patients more effectively.&nbsp;&nbsp;Introduction:&nbsp;Type 1 diabetes (T1D) is a chronic condition characterized by insufficient insulin production, leading to unstable blood glucose levels. Maintaining optimal blood glucose control is crucial for T1D patients to prevent complications and ensure a&nbsp;healthy lifestyle. This project aims to address this challenge by integrating an insulin delivery method and AI-based glucose level prediction.&nbsp;&nbsp;Insulin Delivery System with PID Control:&nbsp;In the first part of the project, an innovative insulin delivery system is introduced. A virtual patient model is employed to simulate real patients’ blood glucose levels. The system utilizes a PID controller to regulate the amount of insulin delivered and achieve optimal blood glucose control. The PID controller is carefully tuned to prioritize the prevention of low blood glucose levels (hypoglycemia) while maintaining a balance to avoid high glucose levels (hyperglycemia), ensuring patient safety.&nbsp;The following graphs show the main concept and result of the first part of the project.&nbsp; <img src="https://strapi.edps.com.hk/uploads/sample1_e35ba963e3.jpg" alt="sample1.jpg"> AI Prediction Model with CGM Data:&nbsp;The second part of the project focuses on leveraging CGM data and advanced AI algorithms to predict patients’ future blood glucose levels. By using continuous and extensive data provided by CGM, a recurrent neural network (RNN) model is trained to offer alternative predictions and personalized treatment plans for T1D patients. To mitigate the impact of error accumulation over time, the AI model is designed to predict glucose levels within a specific near-future timeframe of approximately 5 to 15 minutes. This narrower prediction window enhances accuracy and reliability, enabling healthcare professionals to make timely interventions and adjustments.&nbsp;The following graphs show the main concept and result of the second part of the project.&nbsp; <img src="https://strapi.edps.com.hk/uploads/sample1_2_ca24cb4944.png" alt="sample1.2.png"> Advantages and Implications:&nbsp;The combination of the insulin delivery system with PID control and the AI prediction model provides several significant advantages. Firstly, it facilitates personalized treatment by considering individual variations in insulin requirements and metabolic responses. This approach ensures better glucose control to each patient’s specific needs. Secondly, the system offers early warning signs of high and low blood glucose episodes, enabling medical staff to intervene promptly and prevent potential complications.&nbsp;Conclusion:&nbsp;The Glucose Control Systems with PID Control and AI Prediction project presents an integrated approach to manage blood glucose levels in patients with T1D. By combining an insulin delivery system with PID control and AI prediction using CGM data, the project offers personalized treatment and early warning capabilities, ultimately contributing to improved glucose control and enhanced patient well-being.&nbsp;

Glucose Control Systems with PID Control and AI Prediction Abstract: 

In the realm of art and design, the color of images holds great significance. Its purposefulness is multi-faceted, being paramount for attracting viewers and users in a variety of applications such as flat media, advertisements, and wallpapers. Often, the colors need to be deliberately changed to align with specific design themes or to achieve targeted visual goals. This project introduces a novel multi-modal image re-coloring pipeline, dynamically adjusting colors in images based on natural text, precise user clicks or reference images. By using Vision Transformers, semantic hints are effectively propagated across image and Sub-Pixel Convolution is utilized for real-time inference. The project aims to bring improvement in existing field of image re-colorization while giving end-users a semiautomated process of image re-colorization without specific needs for domain knowledge.&nbsp;&nbsp;Keywords — Interactive Colorization, Vision Transformer,&nbsp;Computational Photography, Image Manipulation, Neural&nbsp;Networks&nbsp;&nbsp;I.&nbsp;INTRODUCTION This project introduces a multi-modal interactive colorization pipeline with real time inference. Its key contributions are:&nbsp;&nbsp;<ul><li>Multi-modal (text, clicks and exemplar image) conditional image colorization using hint points as intermediate representation.&nbsp;</li><li>Vision Transformer with Sub-Pixel convolution achieving real time inference without requiring extensive computational resources.&nbsp;</li></ul>&nbsp;II.&nbsp;METHODOLOGY&nbsp;<ol><li>A.&nbsp;User Hint Generation&nbsp;</li></ol>In the first step, our model converts different types of user input to hint point masks which acts as an input alongside the image for colorization operation. The model supports three types of input. For textual input, we use&nbsp; a special pipeline utilizing CLIP [1] to convert&nbsp; to hint points. For reference image, we use warping and deep learning based [2] operation to generate hint points. Lastly, users can also manually select their preferred points and desired color through precise clicks and color selection through color gamut.&nbsp;<ol><li>B.&nbsp;Colorization Operation&nbsp;</li></ol>As seen in Figure 1, Once Hint points are generated, patches of input image are created which is encoded in a positional layer and passed into a Vision Transformer [3] backbone.&nbsp;This backbone effectively propagates the user hint points respecting semantic context. Once the features are generated the image is up sampled using a sub-pixel convolution&nbsp;technique, which reshapes channel dimension to increase&nbsp;spatial resolution. However, for larger ratio, artefacts can be produced. Because of this, an additional Stabilizing layer (which is a modified Convolutional Neural Network) isadded to group and scale neighboring pixels to achieve&nbsp;smoother upscaling. This allows our model to achieve real time inference without requiring extensive computational resources.&nbsp; &nbsp;<img src="https://strapi.edps.com.hk/uploads/sample2_c06a5160bd.jpg" alt="sample2_.jpg"> Fig. 1. &nbsp;&nbsp;&nbsp;&nbsp;&nbsp; The model takes as input an image and user hint (generated through text, reference image or precise clicks), and passes it into a vision transformer encoder and performs stabilized sub-pixel convolution to generate precise color with rich semantic understanding.&nbsp;&nbsp;III.&nbsp;METHODOLOGY&nbsp;TABLE&nbsp;I&nbsp;:&nbsp;INFERENCE SPEED OF CAPSTONE MODEL AGAINST OTHER BASE LINE MODELS. <img src="https://strapi.edps.com.hk/uploads/sample2_2_f862551426.png" alt="sample2.2.png"> Table 1 highlights, Capstone model’s efficient inference, particularly enabled by its sub-pixel convolution, which is significantly faster than traditional decoder based models.&nbsp; IV.&nbsp;CONCLUSION&nbsp; In conclusion, our interactive multi-modal image recolorization pipeline marks a good advancement in art and&nbsp;design applications, blending user input with cutting-edge AI to deliver real-time, semantically aware colorization.&nbsp; REFERENCES&nbsp;<ol><li>A. Radford et al., “Learning Transferable Visual Models From Natural Language Supervision,” CVPR, pp. 8748–8763, Jul. 2021.&nbsp;</li><li>B. &nbsp;&nbsp;&nbsp; Zhang et al.,&nbsp;“Deep Exemplar-Based Video Colorization,” Conference &nbsp;on Computer Vision and Pattern Recognition, Jun. 2019, doi: 10.1109/cvpr.2019.00824.&nbsp;</li><li>Z. Huang, N. Zhao, and J. Liao, “UniColor: A Unified Framework for&nbsp;</li></ol>Multi-Modal &nbsp;&nbsp;&nbsp;&nbsp; Colorization &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; with &nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp;&nbsp; Transformer,” arXiv &nbsp;&nbsp; (Cornell University), Jan. 2022, doi: 10.48550/arxiv.2209.11223.&nbsp; <img src="https://strapi.edps.com.hk/uploads/sample2_3_741f93d728.jpg" alt="sample2.3.jpg"><img src="https://strapi.edps.com.hk/uploads/sample2_4_777e8efa5b.jpg" alt="sample2.4.jpg"> &nbsp;&nbsp;

Image Re-Colorization for Art & Design 

Abstract:&nbsp;The project presents a high performance online flash order (Seckill) system – OrderFusion. It is built with intelligent design, queuing and caching techniques. The test result shows the solution of Queue + Cache outperformed the tradition MySQL solution, especially under high concurrency situation. The project is through 7 mouths, with the best final result of 77% decrease in delay of access, 3 times throughput and 99% decrease in denial of service incidents, comparing with tradition solution.&nbsp;Keywords:&nbsp;Seckill, Springboot, Redis, High-Performance, Online Order System&nbsp;&nbsp;<h3>I.INTRODUCTION</h3>With the popularity of online Seckill campaigns, order platforms are under tremendous operational pressure. The full solution, named with OrderFusion, is targeted to solve this problem.&nbsp;OrderFusion&nbsp;aims to work as a high performance/ concurrency online Seckill ordering system with Ultra high availability, which leverages strategies along with intelligent design, queuing, caching mechanisms and advanced security measures. Additionally, the system contains a lot of add on solution, like Chatbot powered by GPT4, automated workflows, blockchain logging function and single point authentication (SSO).&nbsp;The key challenges involve processing heavy traffic loads effectively, handling inventory discrepancies, resolving order processing issues during peak periods, and preventing performance declines ensuring security and reliability.&nbsp;In the evaluating with Apache JMeter, the result shows the solution of Queue + Cache outperformed the tradition&nbsp;MySQL solution, especially under high concurrency situation.&nbsp;&nbsp;<h3>II. DESIGN/METHODOLOGY/IMPLEMENTATION</h3>The core of OrderFusion revolves around Seckill Component, which represents high concurrency. This module follows on first-come-first-serve basis, where users join Seckill events and are put into waiting lists.&nbsp; The system uses RabbitMQ, an open-source message broker, to manage the high volume of Seckill requests by queuing them for asynchronous handling.&nbsp;Redis plays a role in caching data such, as Seckill outcomes and event specifics, which helps reduce server load and ensures rapid data retrieval. The Seckill process comprises user involvement tracking, inventory validation and updating, database transactions and user input. Notable features include inventory monitoring and locking mechanisms.&nbsp; <img src="https://strapi.edps.com.hk/uploads/SAMPLE_3_1_36b41f7d0f.png" alt="SAMPLE3.1.png"> Figure: Entire process of Seckill in OrderFusion<h3>&nbsp;</h3><h3>III.EVALAUATION AND RESULTS</h3>The report evaluates the performance of the OrderFusion system using traditional MySQL and Queue + Cache architectures. Apache JMeter was used for stress testing, providing various result views.&nbsp;Throughput, measured as requests processed per unit time, was tested under different concurrent client thread scenarios. The Queue + Cache approach consistently outperformed MySQL, especially when concurrent threads exceeded 500. However, throughput initially decreased with an increase in threads, then increased before decreasing again due to limitations in Redis Cache Pool, RabbitMQ Queue, and Docker container specifications. As concurrent threads increased beyond a certain point, the Queue + Cache's parallel processing capabilities led to a rise in throughput.&nbsp; <img src="https://strapi.edps.com.hk/uploads/SAMPLE_3_2_35ebf0e081.png" alt="SAMPLE3.2.png"> Figure: Throughput Testing Result&nbsp; The Queue + Cache approach also showed better service guarantees in high concurrency scenarios, making it suitable for high-performance online systems like OrderFusion. IV.CONCLUSION The OrderFusion project, a seven-month software engineering endeavor, successfully incorporated advanced technologies like Redis, JWT, and RabbitMQ queues, achieving enhanced performance and security. It significantly improved system efficiency with a 77% reduction in access latency, a threefold increase in throughput, and a 99% reduction in denial of service.&nbsp;This accomplishment illustrates the importance of strategic planning, innovation, and diligent execution in software development, providing insights for future projects.&nbsp;&nbsp;&nbsp;<h3>Highlight Preview&nbsp;</h3>The main task of this project is the development work. For the developers, it was&nbsp;<ul><li>A complete software engineering journey lasting 7 months.</li><li>A comprehensive experience utilizing the knowledge &amp; skills of PolyU undergraduate learning.</li></ul><img src="https://strapi.edps.com.hk/uploads/SAMPLE_3_3_ad8b5a0d3d.png" alt="SAMPLE3.3.png">Figure: OrderFusion Technology Ecology&nbsp; Named OrderFusion, the solution integrates cutting-edge technologies and patterns from the online ordering industry, such as Redis, JWT, RabbitMQ queues, and more, like Figure 2. The public service is available now, for public evaluation, at&nbsp;<a href="https://orderfusion.hirsun.tech/">https://orderfusion.hirsun.tech.</a>&nbsp; <img src="https://strapi.edps.com.hk/uploads/SAMPLE_3_4_050bd6ef23.png" alt="SAMPLE3.4.png"> Figure: OrderFusion UI Design&nbsp;

Technical Corner

Glucose Control Systems with PID Control and AI Prediction Abstract:

Image Re-Colorization for Art & Design

SecKill: High-Performance Online Flash Order System