Yiyun Chen Final Project

CSYE_7370/Final Project/Google_Recaptcha_Recognition-YOLOv5/License.txt

@@ -0,0 +1,8 @@
+License
+Copyright 2022 Yiyun Chen
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

CSYE_7370/Final Project/Google_Recaptcha_Recognition-YOLOv5/README.md

@@ -0,0 +1,77 @@
+# Google Recaptcha Recognition using YOLOv5
+
+## Introduction
+
+### Team Member
+Yiyun Chen 
+NUID:001557532
+
+### About Capcha Recognition
+It can be really hard to specify the label of capcha pictures: “Select all images with a bus/stop signs/a bicycle”, and sometimes we human may be judged as AI as we failed to choose all images correctly. I wonder how good a real AI would perform at this job, so I planned to use deep learning algorithms to recognize the label of capcha pictures.
+
+### Dataset
+#### Original Dataset
+Google Recaptcha Image Dataset
+https://www.kaggle.com/datasets/mikhailma/test-dataset
+Almost 12000 images used in Google Recaptcha V2 collected by category more than 500 of which with manual markup for training object detection model such as YOLO.
+One can use those two txt files and Data1.yaml in folder about_datasets to train it with yolo.
+#### My own datasets
+https://github.com/yiyundotchen/Google-recaptcha-data-for-yolo
+##### Dataset1
+Using Data1.ipynb in folder about_datasets, I selected useful images from original dataset, and made it suitable for yolo training.
+It's smaller than original dataset.
+##### Dataset2
+As shown in results part, the original data is not good for yolo training. No matter how I tune it, the mAP can't be better than 0.5
+So I made labels for bicycle and bus images myself, and made this dataset2 for yolo training.
+I used Data2.ipynb in folder about_datasets to adjust this dataset.
+One can use Data2.yaml in folder about_datasets to train it with yolo.
+
+### Evaluations
+To evaluate object detection models like R-CNN and YOLO, the mean average precision (mAP) is used. The mAP compares the ground-truth bounding box to the detected box and returns a score. The higher the score, the more accurate the model is in its detections.
+Yolo training function offers mAP and other evaluations. I put the evaluations of my final result(the best model) in folder best_model.
+
+
+## Methods
+I'm using YOLOv5, which means "You Only Look Once".
+Yolo is an algorithm that uses convolutional neural networks for object detection. It improves the detection time given that it predicts objects in real-time and provides accurate results with minimal background errors.
+I used yolo both using Jupyter and Google Colab.
+When using Jupyter to run yolo, remember to install cuda, so it can be run on GPU instead of CPU, which is much faster.
+It's much easier and faster to use Colab. I created more Google accounts so that I can use Colab for free.(Highly recommended!!!!!!!)
+
+## Results
+The "img" in forms below means to change the width and height into any number which is divisible by 32.
+Those are results for Data1, details are in Data1_using_Jupyter.ipynb and Data1_using_Colab.ipynb in folder trainings and results.
+|  img   | batch  |  epochs  |  mAP  |
+|  :----:  | :----:  | :----:  | :----:  |
+| 120 | 10 | 500 | 0.437 |
+| 640 | 2 | 60 | 0.354 |
+| 640 | 10 | 60 | 0.36 |
+| 640 | 3 | 300 | 0.372 |
+| 640 | 10 | 300 | 0.417 |
+| 640 | 15 | 300 | 0.444 |
+| 480 | 15 | 300 | 0.441 |
+| 128 | 15 | 300 | 0.398 |
+
+Those are results for Data2.
+|  img   | batch  |  epochs  |  mAP_bicycle  |  mAP_chimney  |  mAP_bus  |  mAP  |
+|  :----:  | :----:  | :----:  | :----:  | :----:  | :----:  | :----:  |
+| 640 | 2 | 3 | 0.26 | 0.00321 | 0.468 | 0.244 |
+| 640 | 16 | 400 | 0.759 | 0.586 | 0.906 | 0.75 |
+
+Those are results for the best model.
+![image](best_model/image/results.png)
+![image](best_model/image/val_batch1_pred.jpg)
+
+## Prospect
+The images used for training are 120*120, which is really unclear. It must have affected the final result.
+With those images, my best mAP is 0.763, still lower than 0.8.
+I'm planning to preprocess those images to improve the model.
+
+
+## Reference
+
+https://github.com/ultralytics/yolov5/issues/6916
+https://medium.com/analytics-vidhya/understanding-yolo-and-implementing-yolov3-for-object-detection-5f1f748cc63a
+https://www.youtube.com/watch?v=GRtgLlwxpc4
+https://github.com/ultralytics/yolov5
+