Skip to content

Commit 4b3ecd7

Browse files
SongZihui-sudoSongZihui-sudo
committed
fix: Improve readme.md
1 parent fbfb479 commit 4b3ecd7

File tree

4 files changed

+108
-315
lines changed

4 files changed

+108
-315
lines changed

‎english.md‎

Lines changed: 76 additions & 180 deletions
Original file line numberDiff line numberDiff line change
@@ -1,228 +1,124 @@
1-
# DigitalFilm - Kodak Gold 200
1+
# DigitalFilm Digital Film
22

3-
DigitalFilm： Use a neural network to simulate film style.
3+
DigitalFilm: Use a neural network to simulate film style.
44

55
---
66

77
<!-- PROJECT LOGO -->
88
<br />
99

1010
<p align="center">
11-
<a href="./readme.md">
12-
<img src="images/logo.svg" alt="Logo" width="320" height="160">
13-
</a>
14-
15-
<h3 align="center">"DigitalFilm</h3>
16-
<p align="center">
17-
Use a neural network to simulate film style.
18-
<br />
19-
<a href="https://github.com/shaojintian/Best_README_template"><strong>Explore the documentation for this project »</strong></a>
20-
<br />
21-
<br />
22-
<a href="./app/digitalFilm.py">View Demo</a>
23-
·
24-
<a href="https://github.com/SongZihui-sudo/digitalFilm/issues">Report bugs</a>
25-
·
26-
<a href="https://github.com/SongZihui-sudo/digitalFilm/issues">Propose new features</a>
27-
</p>
11+
<a href="./readme.md">
12+
</a>
2813

14+
<h3 align="center">"DigitalFilm" Digital Film</h3>
15+
<p align="center">
16+
Use a neural network to simulate film style.
17+
<br />
18+
<a href="https://github.com/shaojintian/Best_README_template"><strong>Explore the documentation of this project »</strong></a>
19+
<br />
20+
<br />
21+
<a href="./app/digitalFilm.py">View the demo</a>
22+
·
23+
<a href="https://github.com/SongZihui-sudo/digitalFilm/issues">Report a bug</a>
24+
·
25+
<a href="https://github.com/SongZihui-sudo/digitalFilm/issues">Propose a new feature</a>
2926
</p>
3027

31-
This README.md is aimed at developers and users
32-
[简体中文](./readme.md)
28+
</p>
29+
30+
This README.md is for developers and users
31+
[English](./english.md)
32+
33+
## Table of Contents
3334

34-
catalogue
35+
- [DigitalFilm Digital Film](#digitalfilm-digital-film)
36+
- [Table of Contents](#table-of-contents)
37+
- [Sample](#sample)
38+
- [Run Demo](#run-demo)
39+
- [training model](#training-model)
40+
- [**Installation steps**](#installation-steps)
41+
- [Overall architecture](#overall-architecture)
42+
- [Dataset](#dataset)
43+
- [File directory description](#file-directory-description)
44+
- [Version Control](#version-control)
45+
- [Author](#author)
46+
- [Copyright](#copyright)
3547

36-
-Digital Film - Kodak Gold 200
37-
-[Catalog] (# Catalog)
38-
-[Run Demo] (# Run Demo)
39-
-[* * Installation Steps * *] (# Installation Steps)
40-
-[Overall Architecture] (# Overall Architecture)
41-
-[Dataset] (# Dataset)
42-
-[Comparison of Generated Images] (# Comparison of Generated Images)
43-
-[Color Space for Generating Images] (# Color Space for Generating Images)
44-
-[File Directory Description] (# File Directory Description)
45-
-[Version Control] (# Version Control)
46-
-[Author] (# Author)
47-
-[Copyright Notice] (# Copyright Notice)
48+
### Sample
49+
50+
![kodak_gold_200](./example/kodak_gold_200.jpg)
51+
<center style="font-size:14px;color:#C0C0C0;text-decoration:underline">Figure 1 Sample kodak gold 200</center>
52+
53+
![fuji_color_200](./example/fuji_color_200.jpg)
54+
<center style="font-size:14px;color:#C0C0C0;text-decoration:underline">Figure 2 Sample fuji color 200</center>
4855

4956
### Run Demo
5057

58+
> The length and width of the input photo need to be divisible by **32**.
59+
5160
```bash
5261
python digitalFilm.py [-v/-h/-g] -i <input> -o <ouput> -m <model>
5362
```
54-
-- v Print version information
55-
-- h Help Information
56-
-- g Graphically select images
57-
-- i Input the directory of the image
58-
-- o Directory for outputting images
59-
-- m model directory
63+
- -v print version information
64+
- -h help information
65+
- -g graphical image selection
66+
- -i input image directory
67+
- -o output image directory
68+
- -m model directory
69+
70+
### training model
71+
72+
training model directly use cyclegan.ipynb.
73+
But you need to download the pre-trained model of resnet18 in advance.
74+
Prepare digital photos and film photos in two folders.
75+
The Kodak Gold 200 model and Fuji c200 model are included in the `app` file.
6076

6177
###### **Installation steps**
6278

6379
```sh
64-
git clone https://github.com/SongZihui-sudo/digitalFilm.git
80+
git clone https://github.com/SongZihui-sudo/digitalFilm.git
6581
```
6682

67-
It's best to create an environment in conda now and install various dependencies.
83+
It is best to create an environment in conda now and then install various dependencies.
6884

6985
```sh
7086
pip install -r requirement.txt
7187
```
7288

7389
### Overall architecture
7490

75-
The overall architecture first trains the data generator through manually annotated digital analog film image pairs, and then generates digital labels using the generator. Finally, the model was trained using digital simulated film images, and then fine tuned using the generated digital real film photo dataset.
76-
77-
```txt
78-
----------------------------------------------------------------
79-
Layer (type) Output Shape Param #
80-
================================================================
81-
Conv2d-1 [-1, 32, 200, 320] 896
82-
BatchNorm2d-2 [-1, 32, 200, 320] 64
83-
LeakyReLU-3 [-1, 32, 200, 320] 0
84-
Conv2d-4 [-1, 64, 200, 320] 18,496
85-
BatchNorm2d-5 [-1, 64, 200, 320] 128
86-
LeakyReLU-6 [-1, 64, 200, 320] 0
87-
AdaptiveAvgPool2d-7 [-1, 64, 1, 1] 0
88-
Conv2d-8 [-1, 4, 1, 1] 256
89-
ReLU-9 [-1, 4, 1, 1] 0
90-
Conv2d-10 [-1, 64, 1, 1] 256
91-
AdaptiveMaxPool2d-11 [-1, 64, 1, 1] 0
92-
Conv2d-12 [-1, 4, 1, 1] 256
93-
ReLU-13 [-1, 4, 1, 1] 0
94-
Conv2d-14 [-1, 64, 1, 1] 256
95-
Sigmoid-15 [-1, 64, 1, 1] 0
96-
ChannelAttention-16 [-1, 64, 1, 1] 0
97-
Conv2d-17 [-1, 128, 1, 1] 73,856
98-
BatchNorm2d-18 [-1, 128, 1, 1] 256
99-
LeakyReLU-19 [-1, 128, 1, 1] 0
100-
AdaptiveAvgPool2d-20 [-1, 128, 1, 1] 0
101-
Conv2d-21 [-1, 8, 1, 1] 1,024
102-
ReLU-22 [-1, 8, 1, 1] 0
103-
Conv2d-23 [-1, 128, 1, 1] 1,024
104-
AdaptiveMaxPool2d-24 [-1, 128, 1, 1] 0
105-
Conv2d-25 [-1, 8, 1, 1] 1,024
106-
ReLU-26 [-1, 8, 1, 1] 0
107-
Conv2d-27 [-1, 128, 1, 1] 1,024
108-
Sigmoid-28 [-1, 128, 1, 1] 0
109-
ChannelAttention-29 [-1, 128, 1, 1] 0
110-
Conv2d-30 [-1, 256, 1, 1] 295,168
111-
BatchNorm2d-31 [-1, 256, 1, 1] 512
112-
LeakyReLU-32 [-1, 256, 1, 1] 0
113-
AdaptiveAvgPool2d-33 [-1, 256, 1, 1] 0
114-
Conv2d-34 [-1, 16, 1, 1] 4,096
115-
ReLU-35 [-1, 16, 1, 1] 0
116-
Conv2d-36 [-1, 256, 1, 1] 4,096
117-
AdaptiveMaxPool2d-37 [-1, 256, 1, 1] 0
118-
Conv2d-38 [-1, 16, 1, 1] 4,096
119-
ReLU-39 [-1, 16, 1, 1] 0
120-
Conv2d-40 [-1, 256, 1, 1] 4,096
121-
Sigmoid-41 [-1, 256, 1, 1] 0
122-
ChannelAttention-42 [-1, 256, 1, 1] 0
123-
Conv2d-43 [-1, 128, 1, 1] 295,040
124-
BatchNorm2d-44 [-1, 128, 1, 1] 256
125-
LeakyReLU-45 [-1, 128, 1, 1] 0
126-
AdaptiveAvgPool2d-46 [-1, 128, 1, 1] 0
127-
Conv2d-47 [-1, 8, 1, 1] 1,024
128-
ReLU-48 [-1, 8, 1, 1] 0
129-
Conv2d-49 [-1, 128, 1, 1] 1,024
130-
AdaptiveMaxPool2d-50 [-1, 128, 1, 1] 0
131-
Conv2d-51 [-1, 8, 1, 1] 1,024
132-
ReLU-52 [-1, 8, 1, 1] 0
133-
Conv2d-53 [-1, 128, 1, 1] 1,024
134-
Sigmoid-54 [-1, 128, 1, 1] 0
135-
ChannelAttention-55 [-1, 128, 1, 1] 0
136-
Conv2d-56 [-1, 64, 1, 1] 73,792
137-
BatchNorm2d-57 [-1, 64, 1, 1] 128
138-
LeakyReLU-58 [-1, 64, 1, 1] 0
139-
AdaptiveAvgPool2d-59 [-1, 64, 1, 1] 0
140-
Conv2d-60 [-1, 4, 1, 1] 256
141-
ReLU-61 [-1, 4, 1, 1] 0
142-
Conv2d-62 [-1, 64, 1, 1] 256
143-
AdaptiveMaxPool2d-63 [-1, 64, 1, 1] 0
144-
Conv2d-64 [-1, 4, 1, 1] 256
145-
ReLU-65 [-1, 4, 1, 1] 0
146-
Conv2d-66 [-1, 64, 1, 1] 256
147-
Sigmoid-67 [-1, 64, 1, 1] 0
148-
ChannelAttention-68 [-1, 64, 1, 1] 0
149-
Conv2d-69 [-1, 32, 1, 1] 18,464
150-
BatchNorm2d-70 [-1, 32, 1, 1] 64
151-
LeakyReLU-71 [-1, 32, 1, 1] 0
152-
AdaptiveAvgPool2d-72 [-1, 32, 1, 1] 0
153-
Conv2d-73 [-1, 2, 1, 1] 64
154-
ReLU-74 [-1, 2, 1, 1] 0
155-
Conv2d-75 [-1, 32, 1, 1] 64
156-
AdaptiveMaxPool2d-76 [-1, 32, 1, 1] 0
157-
Conv2d-77 [-1, 2, 1, 1] 64
158-
ReLU-78 [-1, 2, 1, 1] 0
159-
Conv2d-79 [-1, 32, 1, 1] 64
160-
Sigmoid-80 [-1, 32, 1, 1] 0
161-
ChannelAttention-81 [-1, 32, 1, 1] 0
162-
Conv2d-82 [-1, 3, 1, 1] 867
163-
Conv2d-83 [-1, 3, 200, 320] 99
164-
Conv2d-84 [-1, 3, 1, 1] 195
165-
================================================================
166-
Total params: 805,161
167-
Trainable params: 805,161
168-
Non-trainable params: 0
169-
----------------------------------------------------------------
170-
Input size (MB): 0.73
171-
Forward/backward pass size (MB): 142.14
172-
Params size (MB): 3.07
173-
Estimated Total Size (MB): 145.94
174-
----------------------------------------------------------------
175-
```
91+
Converting digital photos to film style can be regarded as an image style conversion task. Therefore, the overall architecture adopts the cycleGAN network.
92+
[pytorch-CycleGAN-and-pix2pix](https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix)
93+
In addition, it is difficult to obtain large-scale digital photos and film-style photos, so an unsupervised approach is adopted to use unpaired data for training.
17694

17795
### Dataset
17896

179-
At the stage of creating the dataset. Firstly, based on data augmentation technology, simulation film effect images are generated using digital filters applied by Fimo. A preliminary training set is constructed for model pre training, and a digital photo label generation model is synchronously developed. In the second stage, the performance of the model is optimized for real-world scenarios. A dataset is constructed by collecting publicly available Kodajin 200 film samples from the network, and the pre trained label generation model is used to automatically create corresponding digital labels. Ultimately, a two-stage training framework was adopted: first, the model was initialized based on simulation data, and then fine tuned and optimized using a real sample dataset. This progressive training method effectively improved the model's ability to capture real image features while mitigating the risk of overfitting caused by insufficient raw data.
180-
181-
The dataset consists of dual source image data, with the main body collected from high-quality digital photos taken by Xiaomi 13 Ultra smartphones, and the rest selected from professional HDR image datasets [1]. The data annotation system consists of two parallel dimensions: (1) manual annotation group: 1517 pairs of accurately registered digital analog film image pairs, where the film effect is achieved through Fimo professional filters; (2) Automatic generation group: 363 Kodak Gold 200 professional film samples and their corresponding digital labels are automatically generated through a pre trained model. In the data preprocessing stage, a dynamic data augmentation strategy is adopted to apply real-time random spatial transformations to the input image, effectively improving the geometric invariance of the model. The dataset is divided according to the principle of stratified sampling, with 80% (1517 × 0.8+363 × 0.8=1504 images) as the training set and 20% (376 images) as the testing set, ensuring the consistency of the distribution of digital/film samples in the training and testing set.
182-
183-
### Comparison of generated images
184-
185-
![alt text](./images/image_1.png)
186-
187-
<center style="font size: 14px; color: # C0C0C0; text decoration: underline">Figure 1 comparison</center>
188-
189-
190-
![alt text](./images/image_2.png)
191-
<center style="font size: 14px; color: # C0C0C0; text decoration: underline">Figure 2 comparison</center>
192-
193-
![alt text](./images/image_3.png)
194-
<center style="font size: 14px; color: # C0C0C0; text decoration: underline">Figure 3 comparison</center>
195-
196-
197-
###Generate color space for images
198-
199-
![alt text](./images/image_5.png)
200-
<center style="font size: 14px; color: # C0C0C0; text decoration: underline">Figure 4 The RGB channel in Figure 1</center>
201-
202-
203-
![alt text](./images/image_4.png)
204-
<center style="font size: 14px; color: # C0C0C0; text decoration: underline">Figure 4 The color space in Figure 1</center>
97+
The dataset consists of dual-source image data, the main part of which is collected from high-quality digital photos taken by Xiaomi 13 Ultra mobile phone, and the rest is selected from professional HDR image dataset.
98+
Film samples are collected from the Internet.
20599

206100
### File directory description
207101

208-
-DigitalFilm.ipynb is used to train models
209-
-An app demo
210-
- digitalFilm.py
102+
- DigitalFilm.ipynb is used to train the model
103+
- app is a demo
104+
- digitalFilm.py
211105
- mynet.py
212-
- kodark_gold_200.pt
106+
- mynet2.py
107+
- kodark_gold_200.pth
108+
- fuji_color_200.pth
213109

214-
### Version control
110+
### Version Control
215111

216-
This project uses Git for version management. You can refer to the currently available versions in the repository.
112+
This project uses Git for version management. You can view the currently available version in the repository.
217113

218114
### Author
219115

220-
151122876@qq.com SongZihui-sudo
116+
151122876@qq.com SongZihui-sudo
221117

222-
Zhihu: Dr.who &ensp; qq:1751122876
118+
Zhihu:Dr.who &ensp; qq:1751122876
223119

224-
*You can also see all the developers involved in the project in the list of contributors*
120+
*You can also view all the developers involved in the project in the list of contributors. *
225121

226-
### Copyright Notice
122+
### Copyright
227123

228-
This project has signed a GPLv3 license, please refer to [LICENSE. txt] (./LICENSE. txt) for details
124+
This project is licensed under GPLv3. For details, please refer to [LICENSE.txt](./LICENSE.txt)

‎example/fuji_color_200.jpg‎

136 KB
Loading

‎example/kodak_gold_200.jpg‎

750 KB
Loading

0 commit comments

Comments
 (0)