Title: CNN Thunderdome Showdown: Benchmarking YOLOv7, VGG-16, and GoogleNet for Recyclables Image Classification Accuracy

Date: April 8, 2023
Authors: Tom Sun, Dexuan Ren, Deepta Adhikary
University: York University

Abstract

The study focuses on image classification for recyclable sorting and classification. Three models were tested: YOLOv7, VGG-16, and GoogleNet. GoogleNet was found to be the most effective with an accuracy rate of 94% and faster training speed. The findings can be applied to reduce costs in recycling plants and create garbage collecting mini robots.

Introduction

The project explores the challenges of image classification and how transfer learning using pre-trained models like CNNs can overcome these challenges. The study focuses on recyclables and assumes that recycling plants have techniques to isolate each piece of recyclable and capture images.

Methodology

• Models Used: YOLOv7, VGG-16, and GoogleNet.
• Dataset: Kaggle dataset for recyclables.
• Design & Training Pipeline: Pre-trained VGG-16 and GoogleNet models implemented in PyTorch. YOLOv7 used for object detection. Images processed using RoboFlow platform.
• Model Training Changes for YOLOv7: Background removal and manual relabeling.

Results

• Baseline Accuracy: 70% (average from Kaggle models), Human accuracy at 99%.
• Scores:
  • YOLOv7: 70%
  • VGG-16: 88%
  • GoogleNet: 94%
• Observations: Different models have strengths and weaknesses depending on the specific materials or objects being classified.

Discussion

• Strengths and Limitations: YOLOv7 required significant overhead, while VGG-16 and GoogleNet were easier to train. GoogleNet was concluded as the better model for this task.
• Future Directions: Further testing on various datasets and exploring other areas like facial or shape recognition.
• Peer Evaluation: Feedback received and changes made to improve the report.

Updated YOLO V8 and latest model

The project has been updated to use the latest YOLO V8, which enhances the performance and efficiency of the models. You can find more details about YOLO V8 model here.

Conclusion

The study provides a foundation for automated garbage sorting, potentially reducing costs and human effort in waste management. GoogleNet was identified as the most suitable model for this task, with potential applications in recycling industries and the creation of waste-collecting robots.

Full Report

Implemented the Automation on Site for two stores, Using REST API to sync with cloud.
All devices will be auto on or off depending on the user’s Booking, also support remote control.
Store will automaticly close and open in midnight and morning.
Supported Device including, PC, lights, sensors, Survilance cameras, Montion IR cameras.

Source code is an IP to Mr.Leaves Electronics and Networking Inc. thus its not shown here.

Click this to Learn More about the store: ClubHouse Golf

Introduction

We have integrated the Pterodactyl panel for these new nodes, a move that streamlines management and operations while enhancing automation capabilities.
I also participates in the egg (Docker image and Installation scripts) developement of Pterodactyl in the major open source repositories.

My Github Page.

Technical Design

Architecture:

• Distributed System: The nodes are part of a distributed system that ensures high availability and fault tolerance.
• Microservices: The system is built using a microservices architecture, allowing for scalability and ease of maintenance.

Containerization with Docker:

• Isolation: Docker containers encapsulate each game server, ensuring isolation and consistent environments.
• Docker Compose: Used for defining and running multi-container Docker applications, simplifying deployment.

Pterodactyl Panel:

• Integration: The panel is integrated with the Docker ecosystem, providing a unified interface for managing game servers.
• Customization: Allows for extensive customization of server settings and configurations.

Security:

• MFA: Multi-Factor Authentication (MFA) using the Google Authenticator app enhances security.
• Firewalls & Network Isolation: Implementation of firewalls and network segmentation to protect sensitive data.

Automation & Monitoring:

• CI/CD Pipeline: Continuous Integration and Continuous Deployment (CI/CD) pipeline for automated testing and deployment.
• Monitoring Tools: Integration with monitoring tools like Prometheus and Grafana for real-time insights.

Deployment Strategy

Environment Setup:

• Development, Staging, and Production Environments: Ensuring a smooth transition from development to production.

Scalability:

• Horizontal Scaling: Ability to add or remove nodes based on demand, ensuring optimal resource utilization.

Backup & Recovery:

• Regular Backups: Scheduled backups of critical data.
• Disaster Recovery Plan: A robust plan to ensure data integrity and availability in case of failure.

Accessing the New Panel

The new panel can be accessed at Mr.Leaves Server Group.

Conclusion

The technical design and deployment strategy of our game server hosting system reflect our commitment to excellence. By leveraging cutting-edge technologies and best practices, we provide a robust, scalable, and secure solution that meets the needs of demanding users. Explore the new panel and experience the benefits of our advanced hosting solution.