计算机视觉基础与实战应用指南

2.2 图像增强

为了改善视觉效果或提高算法鲁棒性，我们需要对图像进行增强。直方图均衡化能显著提升对比度，而平滑滤波则用于去除噪声。

这里展示几种常用的滤波实现。均值滤波简单快速，高斯滤波效果更自然，中值滤波对椒盐噪声有奇效。

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

2.3 边缘检测

边缘是图像结构的重要特征。Sobel 算子通过计算梯度幅值来检测边缘，而 Canny 算法则是工业界最常用的多阶段边缘检测方法。

import cv2
import numpy as np

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

3.1 HOG 特征

HOG（方向梯度直方图）通过统计局部区域的梯度方向分布来描述物体形状，常用于行人检测等任务。

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

3.2 SIFT 与 ORB 特征

SIFT（尺度不变特征变换）具有优秀的尺度和旋转不变性，但计算量较大。ORB 则是 SIFT 的快速替代方案，结合了 FAST 关键点检测和 BRIEF 描述符，适合实时应用。

import cv2
import numpy as np

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

4.1 传统机器学习模型

在深度学习普及前，支持向量机（SVM）、决策树和随机森林常被用于图像分类。它们依赖人工设计的特征，泛化能力相对有限。

4.2 深度学习模型

现代计算机视觉主要依赖卷积神经网络（CNN）。

• LeNet：早期经典，奠定了 CNN 基础
• AlexNet：引入 ReLU 激活函数，推动深度学习爆发
• VGG：使用小卷积核堆叠，结构简洁优雅
• ResNet：残差连接解决了深层网络梯度消失问题
• YOLO：单阶段目标检测器，兼顾速度与精度

4.3 模型训练实战

以 ResNet 为例，我们使用 PyTorch 进行迁移学习。关键步骤包括数据增强、预训练权重加载及全连接层替换。

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

5.1 需求与设计

我们要构建一个本地桌面应用，支持用户上传图像并进行分类或检测。架构上分为 UI 交互、业务逻辑、图像处理及数据存储四层。

5.2 环境搭建

首先安装必要的依赖库。

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

5.3 核心功能实现

图像输入模块

使用 Tkinter 创建窗口，支持图片选择与预览。

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.ANTIALIAS)
            photo = ImageTk.PhotoImage(image)
            self.image_label.configure(image=photo)
            self.image_label.image = photo
            self.on_image_selected(file_path)

图像分类模块

加载预训练模型并对输入图像进行推理。

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))
    model.eval()
    
    with torch.no_grad():
        outputs = model(image)
        _, preds = torch.max(outputs, 1)
    return class_names[preds[0]]

目标检测模块

使用 Faster R-CNN 检测图像中的物体并绘制边界框。

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))
    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

整合各模块，提供友好的交互界面。

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.ANTIALIAS)
                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()

5.4 运行与测试

启动程序后，依次执行以下步骤：

1. 点击'选择图像'加载本地图片
2. 在功能区选择'图像分类'或'目标检测'
3. 查看下方的文字结果或可视化图像

建议准备包含常见物体的测试集验证模型效果，并根据反馈调整阈值参数。

六、总结

本文系统讲解了计算机视觉的技术体系。从底层的图像预处理、特征提取，到上层的深度学习模型训练，再到最终的应用封装，我们完成了一次完整的技术闭环。掌握这些技能后，你不仅能理解 AI 如何'看'世界，更能亲手构建出解决实际问题的智能工具。