计算机视觉基础与实战应用指南

2.2 图像增强与滤波

为了去除噪声或突出特征，直方图均衡化和各种滤波手段必不可少。均值滤波、高斯滤波和中值滤波各有适用场景，而 Sobel 算子和 Canny 边缘检测则是提取轮廓的关键工具。

import cv2
import numpy as np

def histogram_equalization(image):
    gray_image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    equalized_image = cv2.equalizeHist(gray_image)
    return equalized_image

def mean_filter(image, kernel_size=3):
    blurred_image = cv2.blur(image, (kernel_size, kernel_size))
    return blurred_image

def gaussian_filter(image, kernel_size=3, sigma=0):
    blurred_image = cv2.GaussianBlur(image, (kernel_size, kernel_size), sigma)
    return blurred_image

def median_filter(image, kernel_size=3):
    blurred_image = cv2.medianBlur(image, kernel_size)
    return blurred_image

def sobel_edge_detection(image):
    gray_image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    sobel_x = cv2.Sobel(gray_image, cv2.CV_64F, 1, 0, ksize=3)
    sobel_y = cv2.Sobel(gray_image, cv2.CV_64F, 0, 1, ksize=3)
    sobel_combined = np.sqrt(sobel_x**2 + sobel_y**2)
    sobel_combined = np.uint8(sobel_combined / np.max(sobel_combined) * 255)
    return sobel_combined

def canny_edge_detection(image, threshold1=100, threshold2=200):
    gray_image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    edges = cv2.Canny(gray_image, threshold1, threshold2)
    return edges

三、特征提取方法

在深度学习普及之前，手工设计的特征描述符曾是主流。HOG、SIFT 和 ORB 至今仍在特定场景中发挥作用。

• HOG (Histogram of Oriented Gradients)：通过统计梯度方向直方图来描述局部形状，常用于行人检测。
• SIFT (Scale-Invariant Feature Transform)：具有尺度不变性，能检测关键点并计算描述符，抗干扰能力强。
• ORB (Oriented FAST and Rotated BRIEF)：作为 SIFT 的快速替代方案，速度更快且免费开源。

import cv2
import numpy as np

def extract_hog_features(image):
    hog = cv2.HOGDescriptor()
    gray_image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    features = hog.compute(gray_image)
    return features

def extract_sift_features(image):
    sift = cv2.SIFT_create()
    gray_image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    keypoints, descriptors = sift.detectAndCompute(gray_image, None)
    return keypoints, descriptors

def extract_orb_features(image):
    orb = cv2.ORB_create()
    gray_image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
    keypoints, descriptors = orb.detectAndCompute(gray_image, None)
    return keypoints, descriptors

四、常用模型与架构

随着深度学习的发展，模型架构不断演进。从早期的 LeNet 到经典的 AlexNet、VGG，再到解决梯度消失问题的 ResNet，以及实时目标检测的 YOLO，每一代架构都解决了特定的痛点。

对于传统机器学习，支持向量机（SVM）、决策树和随机森林也是不错的基线模型。但在处理大规模图像数据时，卷积神经网络（CNN）通常是首选。

4.1 模型训练实战 (PyTorch)

以下代码展示了如何使用 PyTorch 加载预训练的 ResNet18 并进行微调。注意数据增强策略和损失函数的选择。

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torchvision import datasets, transforms, models

def train_resnet_model(data_dir, num_classes=2, batch_size=32, num_epochs=10, lr=0.001):
    # 数据预处理
    data_transforms = {
        'train': transforms.Compose([
            transforms.RandomResizedCrop(224),
            transforms.RandomHorizontalFlip(),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
        ]),
        'val': transforms.Compose([
            transforms.Resize(256),
            transforms.CenterCrop(224),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
        ])
    }
    
    image_datasets = {x: datasets.ImageFolder(f'{data_dir}/{x}', data_transforms[x]) for x in ['train', 'val']}
    dataloaders = {x: DataLoader(image_datasets[x], batch_size=batch_size, shuffle=True, num_workers=4) for x in ['train', 'val']}
    dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']}
    class_names = image_datasets['train'].classes
    
    # 加载模型
    model = models.resnet18(pretrained=True)
    num_ftrs = model.fc.in_features
    model.fc = nn.Linear(num_ftrs, num_classes)
    
    # 定义损失函数和优化器
    criterion = nn.CrossEntropyLoss()
    optimizer = optim.SGD(model.parameters(), lr=lr, momentum=0.9)
    scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=7, gamma=0.1)
    
    # 训练模型
    for epoch in range(num_epochs):
        print(f'Epoch {epoch}/{num_epochs - 1}')
        print('-' * 10)
        for phase in ['train', 'val']:
            if phase == 'train':
                model.train()
            else:
                model.eval()
            running_loss = 0.0
            running_corrects = 0
            for inputs, labels in dataloaders[phase]:
                optimizer.zero_grad()
                with torch.set_grad_enabled(phase == 'train'):
                    outputs = model(inputs)
                    _, preds = torch.max(outputs, 1)
                    loss = criterion(outputs, labels)
                    if phase == 'train':
                        loss.backward()
                        optimizer.step()
                running_loss += loss.item() * inputs.size(0)
                running_corrects += torch.sum(preds == labels.data)
            if phase == 'train':
                scheduler.step()
            epoch_loss = running_loss / dataset_sizes[phase]
            epoch_acc = running_corrects.double() / dataset_sizes[phase]
            print(f'{phase} Loss: {epoch_loss:.4f} Acc: {epoch_acc:.4f}')
    print('Training complete')
    return model

五、实战项目：计算机视觉应用开发

理论最终要服务于实践。我们将构建一个包含图像输入、分类、检测功能的桌面应用。

5.1 环境搭建

首先确保安装了必要的依赖库。建议使用虚拟环境管理。

pip install opencv-python
pip install pillow
pip install torch torchvision
pip install tensorflow

5.2 核心功能实现

图像输入与显示

使用 Tkinter 构建界面，PIL 处理图片加载。注意 ANTIALIAS 在新版 Pillow 中已弃用，建议改用 LANCZOS。

import tkinter as tk
from tkinter import filedialog
from PIL import Image, ImageTk

class ImageInputFrame(tk.Frame):
    def __init__(self, parent, on_image_selected):
        tk.Frame.__init__(self, parent)
        self.parent = parent
        self.on_image_selected = on_image_selected
        self.create_widgets()

    def create_widgets(self):
        self.image_label = tk.Label(self)
        self.image_label.pack(pady=10, padx=10, fill="both", expand=True)
        tk.Button(self, text="选择图像", command=self.select_image).pack(pady=10, padx=10)

    def select_image(self):
        file_path = filedialog.askopenfilename(filetypes=[("Image Files", "*.png *.jpg *.jpeg *.bmp")])
        if file_path:
            image = Image.open(file_path)
            image = image.resize((400, 300), Image.LANCZOS)
            photo = ImageTk.PhotoImage(image)
            self.image_label.configure(image=photo)
            self.image_label.image = photo
            self.on_image_selected(file_path)

图像分类

利用加载好的 ResNet 模型进行分类推理。记得设置 map_location='cpu' 以保证跨设备兼容性。

import torch
from torchvision import transforms, models
from PIL import Image

def classify_image(image_path, model_path, class_names):
    data_transforms = transforms.Compose([
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ])
    image = Image.open(image_path)
    image = data_transforms(image)
    image = image.unsqueeze(0)
    
    model = models.resnet18()
    num_ftrs = model.fc.in_features
    model.fc = torch.nn.Linear(num_ftrs, len(class_names))
    model.load_state_dict(torch.load(model_path, map_location='cpu'))
    model.eval()
    
    with torch.no_grad():
        outputs = model(image)
        _, preds = torch.max(outputs, 1)
    return class_names[preds[0]]

目标检测

使用 Faster R-CNN 进行物体检测，并在原图上绘制边界框。

import cv2
import numpy as np
import torch
from torchvision import transforms, models
from PIL import Image

def detect_objects(image_path, model_path, class_names):
    image = cv2.imread(image_path)
    image_rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
    image_pil = Image.fromarray(image_rgb)
    
    data_transforms = transforms.Compose([
        transforms.Resize((416, 416)),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ])
    image_tensor = data_transforms(image_pil)
    image_tensor = image_tensor.unsqueeze(0)
    
    model = models.detection.fasterrcnn_resnet50_fpn(pretrained=False)
    in_features = model.roi_heads.box_predictor.cls_score.in_features
    model.roi_heads.box_predictor = models.detection.faster_rcnn.FastRCNNPredictor(in_features, len(class_names))
    model.load_state_dict(torch.load(model_path, map_location='cpu'))
    model.eval()
    
    with torch.no_grad():
        outputs = model(image_tensor)
    
    boxes = outputs[0]['boxes'].cpu().numpy()
    scores = outputs[0]['scores'].cpu().numpy()
    labels = outputs[0]['labels'].cpu().numpy()
    
    for i in range(len(boxes)):
        if scores[i] > 0.5:
            box = boxes[i].astype(int)
            label = class_names[labels[i]]
            score = scores[i]
            cv2.rectangle(image, (box[0], box[1]), (box[2], box[3]), (0, 255, 0), 2)
            cv2.putText(image, f"{label}: {score:.2f}", (box[0], box[1]-10), cv2.FONT_HERSHEY_SIMPLEX, 0.9, (0, 255, 0), 2)
    return image

结果可视化与主程序

整合上述模块，构建统一的 GUI 入口。

import tkinter as tk
from tkinter import ttk, messagebox, filedialog
from PIL import Image, ImageTk
from image_input_frame import ImageInputFrame
from result_frame import ResultFrame
from cv_functions import classify_image, detect_objects

class CVApp:
    def __init__(self, root):
        self.root = root
        self.root.title("计算机视觉应用")
        self.class_names = ['猫', '狗']
        self.model_path = 'model.pth'
        self.create_widgets()

    def create_widgets(self):
        self.image_input_frame = ImageInputFrame(self.root, self.process_image)
        self.image_input_frame.pack(pady=10, padx=10, fill="both", expand=True)
        
        function_frame = tk.LabelFrame(self.root, text="功能选择")
        function_frame.pack(pady=10, padx=10, fill="x")
        self.function_var = tk.StringVar()
        self.function_var.set("图像分类")
        tk.Radiobutton(function_frame, text="图像分类", variable=self.function_var, value="图像分类").grid(row=0, column=0, padx=5, pady=5)
        tk.Radiobutton(function_frame, text="目标检测", variable=self.function_var, value="目标检测").grid(row=0, column=1, padx=5, pady=5)
        
        self.result_frame = ResultFrame(self.root)
        self.result_frame.pack(pady=10, padx=10, fill="both", expand=True)
        self.output_image_label = tk.Label(self.root)
        self.output_image_label.pack(pady=10, padx=10, fill="both", expand=True)

    def process_image(self, image_path):
        function = self.function_var.get()
        try:
            if function == "图像分类":
                result = classify_image(image_path, self.model_path, self.class_names)
                self.result_frame.display_result(result)
            elif function == "目标检测":
                result_image = detect_objects(image_path, self.model_path, self.class_names)
                result_image = cv2.cvtColor(result_image, cv2.COLOR_BGR2RGB)
                result_image_pil = Image.fromarray(result_image)
                result_image_pil = result_image_pil.resize((400, 300), Image.LANCZOS)
                photo = ImageTk.PhotoImage(result_image_pil)
                self.output_image_label.configure(image=photo)
                self.output_image_label.image = photo
            else:
                raise ValueError("未知功能")
        except Exception as e:
            messagebox.showerror("错误", f"处理失败：{str(e)}")

if __name__ == "__main__":
    root = tk.Tk()
    app = CVApp(root)
    root.mainloop()

六、总结

本文从计算机视觉的基本概念出发，逐步深入到图像处理、特征提取及深度学习模型的应用。通过解析 HOG、SIFT 等传统方法与 ResNet、YOLO 等现代架构的差异，并结合 PyTorch 与 OpenCV 的完整代码示例，希望能帮助你建立起系统的知识框架。最后的实战项目演示了如何将算法封装为可用的桌面应用，这不仅是技术的验证，更是工程落地的第一步。掌握这些技能后，你便具备了开发各类智能视觉应用的能力。