Chess_Recognition-and-FEN

Root Directory

Chess
├─icon
├─images
├─model

• model Directory

rename.py is optional, used for renaming the images.

Data_Aug.py is optional, used for data augmentation. Referrence: https://blog.csdn.net/qq_38973721/article/details/128700920

add_Data.py is optional, user for add images from other folders.

dataloader.py overwrites the Dataset class for this dataset.

model.py is the network architecture, employing modules from modules.py. This network fine-tuned the ResNet50. The architecture can be seen in the figure 1.

train.py and valid.py is used for training the test the model.

This project used channels = 1 to train the model to improve the recognition accuracy for black and white pieces. It's curious that the fine-tuned ResNet has poor performance to distinguish the black and white. I has tried to turn the images to the gray images with 3 channels. Compared with the grey image with single channel, the 3-channel images has no ability to recognize the black and white pieces. I'm curious about this phenomenon.

recognize.py contains the function of the chessboard division into 8*8 pieces for recognition by caliing the Board class from read_contour.py to get the square chessboard. Load the training weight and generate the recognition results.

Compare.py is used to generate changes in the state of the chessboard before and after.

---

rename.py 是可选的，用于重命名图像。

Data_Aug.py 是可选的，用于数据增强。参考：https://blog.csdn.net/qq_38973721/article/details/128700920

add_Data.py 是可选的，用户用于从其他文件夹添加图像。

dataloader.py 覆盖此数据集的 Dataset 类。

model.py是网络架构，使用modules.py中的模块。该网络对 ResNet50 进行了微调。其架构如图1所示。

train.py 和 valid.py 用于训练和测试模型。

本项目使用channels = 1来训练模型，以提高黑白棋子的识别精度。奇怪的是，经过微调的 ResNet 在区分黑白方面的性能很差。我尝试将RGB图像转换为具有 3 个通道的灰色图像。与单通道的灰度图像相比，三通道图像不具备识别黑白棋子的能力。我很不理解。

recognize.py包含将棋盘分割成8*8块进行识别的功能，通过调用read_contour.py中的Board类得到正方形棋盘。加载训练权重并生成识别结果。

Compare.py 用于生成前后棋盘状态的变化。

棋盘数组为:[[0, 12, 2, 4, 12, 2, 12, 12], [5, 5, 12, 12, 1, 5, 12, 0], [1, 12, 12, 12, 12, 12, 12, 5], [12, 7, 5, 12, 11, 12, 5, 12], [12, 12, 12, 12, 12, 11, 11, 12], [12, 12, 12, 12, 11, 12, 12, 8], [11, 11, 12, 12, 12, 12, 12, 11], [6, 12, 8, 10, 12, 12, 7, 6]]
具体棋盘为:[['rook', ' ', 'bishop', 'king', ' ', 'bishop', ' ', ' '], ['soldier', 'soldier', ' ', ' ', 'knight', 'soldier', ' ', 'rook'], ['knight', ' ', ' ', ' ', ' ', ' ', ' ', 'soldier'], [' ', 'KNIGHT', 'soldier', ' ', 'SOLDIER', ' ', 'soldier', ' '], [' ', ' ', ' ', ' ', ' ', 'SOLDIER', 'SOLDIER', ' '], [' ', ' ', ' ', ' ', 'SOLDIER', ' ', ' ', 'BISHOP'], ['SOLDIER', 'SOLDIER', ' ', ' ', ' ', ' ', ' ', 'SOLDIER'], ['ROOK', ' ', 'BISHOP', 'KING', ' ', ' ', 'KNIGHT', 'ROOK']]
棋盘数组为:[[0, 12, 12, 4, 12, 2, 12, 12], [5, 5, 12, 12, 1, 5, 12, 0], [1, 12, 12, 12, 12, 12, 12, 5], [12, 7, 5, 12, 11, 12, 5, 12], [12, 12, 12, 12, 12, 11, 2, 12], [12, 12, 12, 12, 11, 12, 12, 8], [11, 11, 12, 12, 12, 12, 12, 11], [6, 12, 8, 10, 12, 12, 7, 6]]
具体棋盘为:[['rook', ' ', ' ', 'king', ' ', 'bishop', ' ', ' '], ['soldier', 'soldier', ' ', ' ', 'knight', 'soldier', ' ', 'rook'], ['knight', ' ', ' ', ' ', ' ', ' ', ' ', 'soldier'], [' ', 'KNIGHT', 'soldier', ' ', 'SOLDIER', ' ', 'soldier', ' '], [' ', ' ', ' ', ' ', ' ', 'SOLDIER', 'bishop', ' '], [' ', ' ', ' ', ' ', 'SOLDIER', ' ', ' ', 'BISHOP'], ['SOLDIER', 'SOLDIER', ' ', ' ', ' ', ' ', ' ', 'SOLDIER'], ['ROOK', ' ', 'BISHOP', 'KING', ' ', ' ', 'KNIGHT', 'ROOK']]
['c8 -> g4']

FEN.py is imcomplete, used for converting the chessboard data into FEN string.

• Dataset Directory

All the images are divided into 13 classes, representing 12 types of pieces and empty status. Images are collected from the chess.com screenshots.

所有图像分为13类，代表12种棋子和空状态。图片是从“chess.com”屏幕截图中收集的。

Origin folder stores the original images.

Rename folder stores the renamed images.

ImageDataset folder stores the training image data after Dataset Augmentation: ./model/Data_Aug.

valid_ImageDataset stores the test image data.

├─ImageDataset
│  ├─00rook black
│  ├─01knight black
│  ├─02bishop black
│  ├─03queen black
│  ├─04king black
│  ├─05sodier black
│  ├─06rook white
│  ├─07knight white
│  ├─08bishop white
│  ├─09queen white
│  ├─10king white
│  ├─11sodier white
│  └─12empty
├─Origin
│  ├─...
├─Rename
│  ├─...
└─valid_ImageDataset

All the datasets the training weight (.pth) can be downloaded from:

BaiduNetDisk

URL:https://pan.baidu.com/s/1gI27GXrSjcB9CejoJAcN0w?pwd=qhzv

CODE: qhzv

Note that the new2.py contains some methods to crawl web data: "chess.com". Instead of using conventional mathods, I screenshot some icon images and write a program to achieve icon matching and then implement multiple functions. This is a half-effort idea, but subject to company requirements, I have to do this.