基于Python与TensorFlow的非数据集图像导入及识别问题求助

Hey there! Let's tackle the two issues you're facing with your Python/TensorFlow MNIST project—custom image prediction failures and PNG support in your GUI. Here's how to fix both:

1. Fixing Custom Image Prediction Problems #

Your model works great on MNIST test data because those images follow strict formatting rules. When you use hand-drawn images (like from Paint), they almost certainly don't match that format. Let's standardize them:

• Resize to 28x28: MNIST images are exactly 28x28 pixels. Any other size will break the input shape your model expects.
• Convert to Grayscale: Most drawing tools save RGB images, but your model was trained on single-channel grayscale data.
• Normalize Pixel Values: If you scaled MNIST pixels to 0-1 (e.g., with img / 255.0) during training, your custom images need the same treatment.
• Invert Colors (If Needed): MNIST uses black backgrounds with white digits. If your drawing is white background with black digits, reverse the pixel values (255 - img) to match.
• Add Required Dimensions: Your model probably expects input in the shape (batch_size, 28, 28, 1)—you'll need to add batch and channel dimensions to your single image.

Here's a code snippet to handle all this:

import cv2
import numpy as np
import tensorflow as tf

# Load your trained model
model = tf.keras.models.load_model("your_mnist_model.h5")

def preprocess_custom_image(image_path):
    # Read image as grayscale
    img = cv2.imread(image_path, cv2.IMREAD_GRAYSCALE)
    # Resize to MNIST's 28x28 size
    img = cv2.resize(img, (28, 28))
    # Invert colors (adjust this if your model was trained on white background)
    img = 255 - img
    # Normalize to match training preprocessing
    img = img / 255.0
    # Add batch and channel dimensions
    img = np.expand_dims(img, axis=(0, -1))
    return img

# Test with your custom image
processed_img = preprocess_custom_image("your_drawing.png")
prediction = model.predict(processed_img)
print(f"Predicted Digit: {np.argmax(prediction)}")

2. Adding PNG Support to Your GUI #

PNG files often include an alpha (transparency) channel, which many default GUI image handlers can't process. Plus, some GUI libraries (like Tkinter's built-in PhotoImage) don't natively support PNGs. Here's how to fix this:

• Use PIL/Pillow for Image Handling: Pillow can read PNGs, strip the alpha channel, and convert images to a format your GUI can display.
• Convert to Grayscale (For Prediction): If you're using the GUI to feed images to your model, convert the PNG to grayscale first (like we did above).

Example for a Tkinter-based GUI:

from tkinter import Tk, Label, Button, filedialog
from PIL import Image, ImageTk

root = Tk()
root.title("MNIST Digit Predictor")

def load_and_display_png():
    file_path = filedialog.askopenfilename(filetypes=[("PNG Files", "*.png")])
    if not file_path:
        return
    
    # Open PNG, strip alpha channel, convert to grayscale
    img = Image.open(file_path).convert("L")
    # Resize for display (keep original for prediction if needed)
    display_img = img.resize((200, 200))
    # Convert to Tkinter-compatible image
    tk_img = ImageTk.PhotoImage(display_img)
    
    # Update label with the image (keep reference to avoid garbage collection)
    image_label.config(image=tk_img)
    image_label.image = tk_img

image_label = Label(root, text="No Image Loaded")
image_label.pack(pady=10)

load_btn = Button(root, text="Load PNG Image", command=load_and_display_png)
load_btn.pack(pady=5)

root.mainloop()

Make sure you have the required libraries installed:

pip install opencv-python pillow tensorflow

Let me know if you hit any snags—these fixes should cover most common cases for your issues!

内容的提问来源于stack exchange，提问作者wicwik