[ML][CV] 車牌辨識 / 資料增強 / imgaug
===
###### tags: `ML / CV`
###### tags: `ML`, `CV`, `augment`
<br>
## 套件
- [[github] aleju/imgaug](https://github.com/aleju/imgaug)
<br>
## [Weather](https://github.com/aleju/imgaug/blob/master/imgaug/augmenters/weather.py)
- ### 效果圖
| FastSnowyLandscape<br>(lightness_multiplier=2.0) | Clouds | Fog | Snowflakes |
| -------- | -------- | -------- | -------- |
|  |  |  |  |
<br><br>
## [Weather / FastSnowyLandscape](https://github.com/aleju/imgaug/blob/master/imgaug/augmenters/weather.py)
- ### lightness_multiplier=2.0
[](https://i.imgur.com/bXBNJWu.png)
<br>
- ### lightness_threshold=50, lightness_multiplier=20.0
[](https://i.imgur.com/AkG2bOH.png)
[](https://i.imgur.com/pYYND9P.png)
<br><br>
## [Weather / Clouds](https://github.com/aleju/imgaug/blob/master/imgaug/augmenters/weather.py)
- 無參數可控制，效果可能不如預期。在 ```0003-ZT```中，16 張才出現 3~4 張想要的。
```python
iaa.Clouds().augment_image(raw)
```
[](https://i.imgur.com/QpTxZ10.png)
<br>
[](https://i.imgur.com/JoP2Bl9.png)
<br><br>
## [Weather / Fog](https://github.com/aleju/imgaug/blob/master/imgaug/augmenters/weather.py)
- 無參數可控制。
```python
iaa.Fog().augment_image(raw)
```
iaa.Clou
[](https://i.imgur.com/MPP4ZqN.png)
<br>
[](https://i.imgur.com/7ml6TGM.png)
<br><br>
## [Weather / Snowflake](https://github.com/aleju/imgaug/blob/master/imgaug/augmenters/weather.py)
- ### 效果圖
- 原圖
<span style="margin-right:1.8em"></span>
<br>
- ```flake_size=(0.4, 0.9), speed=(0.01, 0.05)``` 隨機產生
- ### 程式碼
```python
from matplotlib import image as im
from matplotlib import pyplot as plt
from imgaug import augmenters as iaa # image / aug / augmenters
```
```python
raw = im.imread('tn_2M-7933_lp.jpg')
plt.figure(figsize=(6,2))
plt.title('raw')
plt.imshow(raw)
fig = plt.figure(figsize=(20, 30))
row = 10
col = 3
applier = iaa.Snowflakes(flake_size=(0.4, 0.9), speed=(0.01, 0.05))
for idx in range(row * col):
fig.add_subplot(row, col, idx + 1) # the current canvas
aug_img = applier.augment_image(raw)
plt.title('Snowflakes - ' + str(idx + 1))
plt.imshow(aug_img)
```
<br>
- ### 參數說明
- #### density (雪花密度)
- 有效範圍 ```[0.0, 1.0]```
- 建議範圍 ```[0.01, 0.075]```
- #### flake_size (雪花大小)
- 有效範圍 ```[0.0, 1.0]```
- 建議範圍
* 影像大小 96x128 搭配 (0.1, 0.4)
* 影像大小 192x256 搭配 (0.2, 0.7)
* 影像大小 960x1280 搭配(0.7, 0.95)
- #### angle (雪花角度)
> Angle in degrees of motion blur applied to the snowflakes,
> where ``0.0`` is motion blur that points straight upwards.
> Recommended to be in the interval ``[-30, 30]``.
- 單位 degrees
- 建議範圍 ```[-30, 30]```
- #### speed (雪花的下降速度)
> This parameter controls the motion blur's kernel size. It follows roughly the form ``kernel_size = image_size * speed``. Hence, values around ``1.0`` denote that the motion blur should "stretch" each snowflake over the whole image.
- 有效範圍 ```[0.0, 1.0]```
- 建議範圍
* 影像大小 96x128 搭配 (0.01, 0.05)
* 影像大小 192x256 搭配 (0.007, 0.03)
* 影像大小 960x1280 搭配 (0.001, 0.03)
- 當值為 1 時，雪花軌跡會形成白線條，橫跨整張圖
- ### 案例說明
- 即便參數都是固定一個數值，產生的效果也是隨機的
```density=0.1, flake_size=0.0001, angle=-6, speed=0.5```
<br>
- density 超過 0.5 以上的例子，超過 9 成都無法使用。必須將 speed 控制為 0
- density = 0.5
[](https://i.imgur.com/eB094Ph.png)
- density = 0.7
[](https://i.imgur.com/EcB5pHj.png)
- density = 1.0
[](https://i.imgur.com/orCMcU2.png)
<br>
## 測試實驗
- ### case1 - base
- train
- dataset/trainset/G4_Archives_7monthes.new
- test
- dataset/trainset/G4_Archives_7monthes.new
- Total: 9292
- Wrong: 9
- Time: 0.07271468539305714
- ACC: 0.999
- dataset/testset/T1_20191210_tn_renew/
- Total: 68
- Wrong: 8
- Time: 0.1030008126707638
- ACC: 0.88
- ### case2 - with snowflakes
- train
- dataset/trainset/G4_Archives_7monthes.new.Snowflakes
- test
- dataset/trainset/G4_Archives_7monthes.new.Snowflakes
- dataset/testset/T1_20191210_tn_renew/
<!--
from matplotlib import image as im
from matplotlib import pyplot as plt
import matplotlib._png as png
from imgaug import augmenters as iaa # image / aug / augmenters
import shutil
import random
applier_list = [
iaa.Snowflakes(density=0.1, flake_size=0.9, speed=0),
iaa.Snowflakes(density=0.2, flake_size=0.7, speed=0),
iaa.Snowflakes(density=0.3, flake_size=0.5, speed=0),
iaa.Snowflakes(density=0.4, flake_size=0.3, speed=0),
iaa.Snowflakes(density=0.5, flake_size=0.1, speed=0),
iaa.Snowflakes(density=0.1, flake_size=0.3, speed=0.04),
iaa.Snowflakes(density=0.2, flake_size=0.6, speed=0.04),
iaa.Snowflakes(density=(0.05, 0.075), flake_size=(0.8, 0.9), speed=(0.03, 0.045))
]
src_dataset_folder = '/tmp/TJ/dataset/trainset/G4_Archives_7monthes.new/'
new_dataset_folder = '/tmp/TJ/dataset/trainset/G4_Archives_7monthes.new.Snowflakes/'
filename_list = os.listdir(src_dataset_folder)
filename_list.sort()
row = 5
col = 3
aug_size_per_img = row * col
plt_limit = 0
for filename in filename_list:
filepath = os.path.join(src_dataset_folder, filename)
new_filepath = os.path.join(new_dataset_folder, filename)
shutil.copyfile(filepath, new_filepath)
#raw_img = im.imread(file_path)
raw_img = png.read_png_int(filepath)
#plt.title('raw: ' + filename)
#plt.imshow(raw_img)
#plt.show()
#fig = plt.figure(figsize=(20, 12))
for idx in range(aug_size_per_img):
#fig.add_subplot(row, col, idx + 1)
index = int(random.random() * 100 // 5)
if index < len(applier_list):
applier = applier_list[index]
else:
applier = applier_list[-1]
aug_img = applier.augment_image(raw_img)
tokens = os.path.splitext(filename)
new_filename = tokens[0] + '_' + '{:02d}.jpg'.format(idx)
new_filepath = os.path.join(new_dataset_folder, new_filename)
im.imsave(new_filepath, aug_img)
#plt.imshow(aug_img)
#plt.title('aug: ' + str(idx))
#plt.show()
#plt_limit += 1
#if plt_limit > 5:
# break
-->
Sign in with Wallet
Connect another wallet