Template Matching

Object Detection Project Using Limited Images

This project aims to detect objects in images using a very limited amount of input data. I have started this project with four different images of mice with varying colors and patterns. One of the features of this project is the use of diverse and varied inputs, which enhances the diversity of the training dataset and improves the model's accuracy.

In this project, I have used ten different images as backgrounds. These backgrounds are combined with various input images to create a comprehensive and diverse dataset for training the YOLOv8 model. To achieve this, several libraries and tools for image processing, data augmentation, and model training have been utilized.

This project allows me to create a rich and diverse training dataset with minimal input data.

Table of Contents

1. Project Setup
2. Directory Structure
3. Downloading Data
4. Generating Backgrounds
5. Image Augmentation
6. Creating Composite Images
7. Preparing the Dataset
8. Training YOLOv8
9. Model Evaluation
10. Model Inference
11. Results and Applications
12. Requirements
13. Contribution
14. Acknowledgements
15. Contact
16. References

Project Setup

First, clone the required repository and install necessary libraries. This involves using Git to clone the GradientGenerator repository and installing Python packages like ultralytics, gdown, Augmentor, and tqdm.

Directory Structure

Create the necessary directories for storing background images, training data, validation data, and test data.

Downloading Data

Download the required project data, including input images, background images, and test images from Google Drive. This ensures you have all the necessary resources for the project.

Generating Backgrounds

1. Creating Gradient Backgrounds Initially, gradient background images are generated using the create_back(r_channel, g_channel, b_channel) function. This function utilizes predefined color channels (r_channel, g_channel, b_channel) and fixed colors such as white, black, yellow, red, and blue to produce gradient background images sized at 640x640 pixels. For each predefined color, a gradient background image is created, and its name and path are updated accordingly.

2. Incorporating Downloaded Background Images In this stage, alongside gradient color backgrounds, 10 challenging images are used as additional backgrounds. These images are downloaded from other sources and moved to the backgrounds_images folder.

Image Augmentation

Random transformations are applied to input images to increase dataset diversity using the Augmentor library. These transformations include rotation, horizontal and vertical flips, contrast and brightness adjustments, zooming, distortion, and random cropping.

Creating Composite Images

Method 1:

Image Size 50x50 pixels

For each input image resized to 50x50 pixels, the final image is placed on each available background. This process is repeated for 15 random positions per background, resulting in 15 final images for each background.

Image Size 110x110 pixels

For each input image resized to 110x110 pixels, the final image is placed on each available background. This process is repeated for 10 random positions per background, resulting in 10 final images for each background.

Method 2: Random Number of Input Images on Backgrounds

In this method, a random number of input images (between 2 to 5 images) are placed on each background. This adds variability to the dataset by allowing a different number of input images to be combined on each background.

Overlapping Check of Images

To ensure that the input images do not overlap on the backgrounds:

• A function is utilized to check the coordinates of each image.
• If overlap is detected, the function determines a new position for the image to avoid overlap.

These methods aim to generate a diverse dataset for training the YOLO network, enhancing its ability to detect objects effectively across various backgrounds and scenarios. Each method offers a different approach to combining input images with backgrounds, contributing to the robustness and generalization of the trained model.

Generating Labels for YOLO

For each final image, a label file in text format is generated that specifies the normalized position and size of each object in the image relative to the dimensions of the background image. Generally, this format includes: class x_center y_center width height.

Preparing the Dataset

Split the dataset into training, validation, and test sets. 70% of the data is used for training, 20% for validation, and 10% for testing. This ensures that the model can be properly evaluated on unseen data.

Training YOLOv8

The YOLOv8 model is trained using the training dataset for 7 epochs. This includes configuring various training parameters such as the number of epochs, image size, and settings for data augmentation to optimize the model. Data augmentation settings involve changes in color, rotation, translation, scaling, and more.

Model Evaluation

After training, the model is evaluated using validation data to assess its performance.

Model Inference

The trained model is used for predicting objects in new images. Prediction results are displayed, and some samples of predicted images are shown as output.

Results and Applications

This project enables us to generate a rich and diverse training set of images using a limited number of input images, thereby improving the quality and accuracy of object detection models. The trained model can accurately detect desired objects in new images.

Requirements

• Python 3.x
• Ultralytics
• Augmentor
• YOLO
• tqdm
• NumPy
• Pandas
• gdown
• random

Contribution

Contributions are welcome! Please fork the repository, make your changes, and submit a pull request.

To enhance the project and add new capabilities, you can consider the following approaches:

1. Utilizing ImageNet and other images as background:

   • Using ImageNet to check images and extract suitable backgrounds for blending with input images.
   • Implementing an automated system to select appropriate backgrounds based on categories available in ImageNet.

2. Use image captioning for create background

3. Implementing YOLOv10:

• Testing and applying YOLOv10 instead of YOLOv8 for better performance evaluation and accuracy.
• New features and potential improvements in the YOLOv10 architecture, such as higher object detection accuracy, faster processing speed, and network training.

4. Upscaling Images using Seamless Cloning:
   • Implementing more advanced methods for upscaling images using Seamless Cloning in OpenCV.
   • Improving the accuracy and quality of images produced by this method through process and method optimizations.

Acknowledgements

• GradientGenerator Repository
• YOLOv8 by Ultralytics

Contact

For any inquiries or support, please contact [email protected].

References

• Blending images using OpenCV
• Augmentation settings
• (https://rumn.medium.com/yolo-data-augmentation-explained-turbocharge-your-object-detection-model-94c33278303a)