如何获取最优图像输入大小

# 导入库
import pandas as pd
import matplotlib.pyplot  as plt
from PIL import Image
from pathlib import Path
import imagesize
import numpy as np

# 获取图像路径
root=r"D:/Test/YOLOv5/data/coco128_fire/images/val2017/"
print( Path(root).iterdir())
imgs = [img.name for img in Path(root).iterdir() if img.suffix == ".png"]
img_meta = {}
for f in imgs: img_meta[str(f)] = imagesize.get(root+f)

# 将其转换为 Dataframe 并计算纵横比
img_meta_df = pd.DataFrame.from_dict([img_meta]).T.reset_index().set_axis(['FileName', 'Size'], axis='columns', inplace=False)
img_meta_df[["Width", "Height"]] = pd.DataFrame(img_meta_df["Size"].tolist(), index=img_meta_df.index)
img_meta_df["Aspect Ratio"] = round(img_meta_df["Width"] / img_meta_df["Height"], 2)

print(f'Total Nr of Images in the dataset: {len(img_meta_df)}')
img_meta_df.head()


# 可视化图像分辨率
fig = plt.figure(figsize=(8, 8))
ax = fig.add_subplot(111)
points = ax.scatter(img_meta_df.Width, img_meta_df.Height, color='blue', alpha=0.5, s=img_meta_df["Aspect Ratio"]*100, picker=True)
ax.set_title("Image Resolution")
ax.set_xlabel("Width", size=14)
ax.set_ylabel("Height", size=14)

检查错误的数据

1.如果我们可以在检查图中的点的同时看到底层图像，那不是很好吗？这可以帮助我们确定以下潜在问题。

• 错误标记的数据：可能会混淆模型。

• 某些类图像的分辨率很高或很低：可能会使模型产生偏差。

2.检查后，如果我们认为需要删除某些数据点，则使用Lasso Selector来实现。

# # 导入库
# import pandas as pd
# import matplotlib.pyplot  as plt
# from PIL import Image
# from pathlib import Path
# import imagesize
# import numpy as np

# # 获取图像路径
# root=r"D:/Test/YOLOv5/data/coco128_fire/images/val2017/"
# print( Path(root).iterdir())
# imgs = [img.name for img in Path(root).iterdir() if img.suffix == ".png"]
# img_meta = {}
# for f in imgs: img_meta[str(f)] = imagesize.get(root+f)

# # 将其转换为 Dataframe 并计算纵横比
# img_meta_df = pd.DataFrame.from_dict([img_meta]).T.reset_index().set_axis(['FileName', 'Size'], axis='columns', inplace=False)
# img_meta_df[["Width", "Height"]] = pd.DataFrame(img_meta_df["Size"].tolist(), index=img_meta_df.index)
# img_meta_df["Aspect Ratio"] = round(img_meta_df["Width"] / img_meta_df["Height"], 2)

# print(f'Total Nr of Images in the dataset: {len(img_meta_df)}')
# img_meta_df.head()


# # 可视化图像分辨率
# fig = plt.figure(figsize=(8, 8))
# ax = fig.add_subplot(111)
# points = ax.scatter(img_meta_df.Width, img_meta_df.Height, color='blue', alpha=0.5, s=img_meta_df["Aspect Ratio"]*100, picker=True)
# ax.set_title("Image Resolution")
# ax.set_xlabel("Width", size=14)
# ax.set_ylabel("Height", size=14)




# 导入库
import pandas as pd
import matplotlib.pyplot  as plt
from PIL import Image
from pathlib import Path
import imagesize
import numpy as np

# 获取图像路径
root=r"D:/Test/YOLOv5/data/coco128_fire/images/val2017/"
print( Path(root).iterdir())
imgs = [img.name for img in Path(root).iterdir() if img.suffix == ".png"]
img_meta = {}
for f in imgs: img_meta[str(f)] = imagesize.get(root+f)

# 将其转换为 Dataframe 并计算纵横比
img_meta_df = pd.DataFrame.from_dict([img_meta]).T.reset_index().set_axis(['FileName', 'Size'], axis='columns', inplace=False)
img_meta_df[["Width", "Height"]] = pd.DataFrame(img_meta_df["Size"].tolist(), index=img_meta_df.index)
img_meta_df["Aspect Ratio"] = round(img_meta_df["Width"] / img_meta_df["Height"], 2)

print(f'Total Nr of Images in the dataset: {len(img_meta_df)}')
img_meta_df.head()


# # 可视化图像分辨率
# fig = plt.figure(figsize=(8, 8))
# ax = fig.add_subplot(111)
# points = ax.scatter(img_meta_df.Width, img_meta_df.Height, color='blue', alpha=0.5, s=img_meta_df["Aspect Ratio"]*100, picker=True)
# ax.set_title("Image Resolution")
# ax.set_xlabel("Width", size=14)
# ax.set_ylabel("Height", size=14)



# Import libraries
from matplotlib.widgets import LassoSelector
from matplotlib.path import Path as mplPath

# Lasso Selection of data points
class SelectFromCollection:
  def __init__(self, ax, collection, alpha_other=0.3):
    self.canvas = ax.figure.canvas
    self.collection = collection

    self.xys = collection.get_offsets()
    self.lasso = LassoSelector(ax, onselect=self.onselect)
    self.ind = []

  def onselect(self, verts):
    path = mplPath(verts)
    self.ind = np.nonzero(path.contains_points(self.xys))[0]
    self.canvas.draw_idle()

  def disconnect(self):
    self.canvas.draw_idle()

# Show the original image upon picking the point
def on_pick(event):
  ind = event.ind[0]
  w, h = points.get_offsets().data[ind]
  img_file = Path(img_meta_df.iloc[ind, 0])
  if Path(root,img_file).is_file():
    print(f"Showing: {img_file}")
    img = Image.open(Path(root,img_file))
    figs = plt.figure(figsize=(5, 5))
    axs = figs.add_subplot(111)
    axs.set_title(Path(img_file).name, size=14)
    axs.set_xticks([])
    axs.set_yticks([])
    axs.set_xlabel(f'Dim: {round(w)} x {round(h)}', size=14)
    axs.imshow(img)
    figs.tight_layout()
    figs.show()

# Save selected image filenames  
def save_selected_imgs(df, fileName = Path("Images to discard.csv")):
  if fileName.is_file():
    orgData = pd.read_csv(fileName)
    df = pd.concat([orgData, df])
  df.set_axis(['FileName'], axis='columns').to_csv(fileName, index=False)

# Store selected points upon pressing "enter"
def accept(event):
  if event.key == "enter":
    selected_imgs = img_meta_df.iloc[selector.ind, 0].to_frame()
    save_selected_imgs(selected_imgs)
    print("Selected images:")
    print(selected_imgs)
    selector.disconnect()
    fig.canvas.draw()
        
# Plot the image resolutions        
fig = plt.figure(figsize=(8, 8))
ax = fig.add_subplot(111)
points = ax.scatter(img_meta_df.Width, img_meta_df.Height, color='blue', alpha=0.5, s=img_meta_df["Aspect Ratio"]*100, picker=True)
ax.set_title("Press enter to after selecting the points.")
ax.set_xlabel("Width", size=14)
ax.set_ylabel("Height", size=14)

# Add interaction
selector = SelectFromCollection(ax, points)
fig.canvas.mpl_connect("key_press_event", accept)
fig.canvas.mpl_connect('pick_event', on_pick)
plt.show()

import cv2
cv2.waitKey(0)

link