深度学习与CNN在计算机视觉实践中的应用详解（Python+TensorFlow+Kivy）

需积分: 9 201 浏览量更新于2024-07-17 收藏 8.78MB PDF 举报

《实用计算机视觉应用：深度学习与卷积神经网络》是一本由Ahmed Fawzy Gad编写的深入指南，专为希望在实际场景中应用深度学习进行计算机视觉任务的专业人士设计。本书结合了理论与实践，以Python为例，特别是利用TensorFlow和Kivy这两个强大的开源工具。书中详细介绍了如何利用卷积神经网络（CNNs）处理各种计算机视觉问题，如图像分类、物体检测、人脸识别等。该书的主要内容围绕以下几个方面展开： 1. **入门介绍**：首先，读者将了解到计算机视觉的基础概念，包括CNNs的工作原理和它们在深度学习中的核心地位，以及为何选择Python作为主要编程语言。 2. **深度学习基础**：章节中会涵盖深度学习的基本原理，如神经网络结构、反向传播算法以及优化方法（如Adam或SGD），为理解CNNs做准备。 3. **CNN架构解析**：作者将深入讲解各种经典的CNN模型，如LeNet、AlexNet、VGG、ResNet和Inception，以及它们在特定任务上的优化版本，如YOLO（You Only Look Once）和Faster R-CNN用于目标检测。 4. **Python实战示例**：书中提供了大量的代码示例，通过实际操作演示如何在Python环境中搭建CNN模型，从数据预处理到模型训练、验证和评估，再到部署和应用。 5. **Kivy集成**：特别强调了如何将深度学习技术与Kivy这个跨平台的图形库结合，以便开发用户友好的交互式应用程序，比如实时图像分析或移动端视觉应用。 6. **案例研究**：书中还包含几个实用案例，展示了如何解决实际生活中的计算机视觉问题，如自动驾驶、医疗图像分析和社交媒体内容识别，帮助读者理解如何将所学知识应用于实际项目。 7. **版权和授权**：最后，版权信息明确，确保读者在合法范围内使用本书内容，同时强调了对知识产权的尊重。《实用计算机视觉应用：深度学习与CNNs》是一本实用且全面的教程，不仅适合计算机视觉领域的初学者，也对有一定经验的开发者提供深入了解和提升技能的机会。通过阅读这本书，读者将能够掌握如何在实际工作中利用深度学习的威力来解决复杂的计算机视觉问题。

placeholders, variables, dataflow graphs, and TensorBoard are discussed based on some

examples building a simple linear model and an ANN for simulating the XOR gate. By

the end of this chapter, a CNN is created using tensorflow.nn module for classifying the

CIFAR10 dataset.

Chapter 7 deploys the trained models into a web server for being accessed by

Internet users using a web browser. A web application is created using the Flask

microframework. HTML, CSS, and JavaScript are used to build the front pages for

accessing the web server. The HTML pages send HTTP requests to the server with an

image, and the server responds to such requests with the predicted class.

Chapter 8 builds cross-platform applications using the Kivy open source library.

By linking Kivy to NumPy, it is possible to build data science applications that work

unchanged on different platforms. This removes the overhead of customizing the code

for a specific platform. An Android application is created to read an image and execute

the CNN layers implemented using NumPy in Chapter 5.

To benefit from the projects created, it is preferred to push them online for other

people to use and benefit from. An appendix discusses how to package Python projects

and distribute them into the Python package index (PyPI) repository.

Before starting, let’s take a brief overview of the Python environment used in the

book.

All code in the book is implemented using Python. Because native Python is

complex for handling images, multiple libraries are used to help to produce an efficient

implementation for applications across the chapters.

At first, native Python could be downloaded from this link (www.python.org/

downloads). The book uses Python 3. Just install the version of Python that is suitable

for your system. The next step is to prepare all libraries required across the entire book.

Rather than installing individual libraries, it is recommended to use Anaconda Python

distribution. It is available for download from this link (www.anaconda.com/download).

It supports Windows, Mac, and Linux and packages more than 1,400 data science

libraries. A list of all supported packages can be accessed from this page (https://repo.

anaconda.com/pkgs). By just installing Anaconda on your machine, all of the supported

libraries will be ready for use. This is helpful to avoid the challenges of preparing the

Python environment.

The required libraries in this book are NumPy, SciPy, Matplotlib, scikit-image,

scikit-learn, TensorFlow, Flask, Werkzeug, Jinja, Pickle, Pillow, and Kivy. All of these

libraries, except for Kivy, are supported by Anaconda. Later in this chapter, we will

InTroduCTIon

xxi

see the function of each of them. Note that such libraries can be installed easily. After

installing the native Python, we can use the pip installer to download and install a library

based on this command: “pip install <lib-name>”. Just type the name of the library. Some

installations are not straightforward and might change if the system changes. Thus,

we can’t cover the different installations. For such reasons, Anaconda is better than

installing each library individually. Let’s discuss the libraries needed.

Python supports a number of built-in data structures: list, tuple, dictionary, set, and

string. Unfortunately, no data structure provides flexibility in data science applications.

These data structures support working with different data types at the same time.

The same data structure might contain numbers, characters, objects, and more. String

is an exception, in which only characters are supported. Moreover, string and tuple are

immutable, which means it is impossible to change their values after they are created.

Dictionary adds a key to each item. Saving image pixels using a dictionary requires

adding a key to each pixel which enlarges the amount of data saved. Set is restricted to

just set operations and images are not restricted to just such operations.

Talking about images, which are the main concern of the book, list is the suitable

data structure. It is a mutable data type that is able to hold matrices. Unfortunately,

working with lists makes the process complex. We have to make sure everything is

numeric, of a certain specific type because different numeric data types can be saved

in the same list. To apply a simple operation such as adding a number to the image, we

have to write loops for visiting each element and apply such operations individually.

In data science applications, it is recommended to use the tools that make applying

the operations easier. There are some challenging tasks to conquer when building an

application, and we do not need to add another challenge in programming such tasks.

For such reasons, the NumPy (Numeric Python) library is used. Its basic role is to

support a new data structure in Python, which is array. Working with NumPy arrays is

simpler than working with lists. For example, using just the addition operator (+), we

can add a number to each element in the image after it is converted into a NumPy array.

Many other libraries have their functions accept and return a NumPy array.

Some operations are supported inside the NumPy array, but it is not meant to apply

operations. The SciPy (Scientific Python) library supports the same operations in the

NumPy arrays and more. It also supports working with the n-dimensional NumPy arrays

(e.g., images) using the scipy.ndimage submodule. For more advanced operations on

images, the scikit-image library is used. For example, image features can be extracted

using this library.

InTroduCTIon

xxii

After reading the image and applying some operations, Matplotlib is used for

displaying the images. It is mainly used for 2D visualization, but it also supports some 3D

features.

After reading the images, extracting features, and making visualizations, we can start

building ML models using the scikit-learn library. It supports different types of models

that are ready for use. Just feed it with inputs, outputs, and their parameters to have a

trained model.

After training a ML model, we can save it for later use using the pickle library, which

serializes and deserializes the objects. Up to this point, we can build and save an ML

model. We then move to building and saving a DL model using TensorFlow. It is the

most commonly used DL library, as it supports different APIs that match the needs

of professionals and beginners alike. TensorFlow has its own ways to save the trained

models.

In order to deploy the trained models, Flask is used. It is a microframework for

building web applications. By deploying the trained models to the web server, clients can

access such applications using a web browser. They can upload test images to the server

and receive the classification label. Flask uses the Jinja2 template engine and WSGI for

building the applications. For such reasons, the libraries Jinja and werkzeug must be

installed.

In order to build a data science mobile application that runs on-device, Kivy is used.

It is a Python library that allows the Python code to run cross-platform. In this book, Kivy

is used to build rich data science applications running for Android. The APK generated

by Kivy can be used in the market exactly as if it were created normally using Android

Studio.

Kivy uses the python-for-android packager, which allows adding the required

dependencies in the Android application. Because scikit-image is not supported by

python-for-android, images are read using Pillow, which is supported to run on Android

devices.

Implementations of the projects are available at this GitHub account: github.com/

AhmedFGad

InTroduCTIon

A. F. Gad, Practical Computer Vision Applications Using Deep Learning with CNNs,

https://doi.org/10.1007/978-1-4842-4167-7_1

CHAPTER 1

Recognition inComputer

Vision

Most computer science research tries to build a human-like robot that is able to function

exactly as humans. Even emotional properties are not impossible for such robots. Using

a sensor, the robot feels the temperature in the surrounding environment. Using facial

expressions, it is possible to know whether a person is sad or happy. Even things that

seem impossible might eventually only be challenging.

At the current time, a very challenging application is object recognition. Recognition

can be based on different types of data such as audio, image, and text. But image

recognition is a very efficient way due to the plenitude of information that can be helpful

in the task. Thus, it is regarded as the most popular application in computer vision.

There is a massive number of objects existing in the world, and differentiating them

is a complex task. Different objects might have similar visual appearance except for subtle

details. Moreover, the same object appears differently based on its surrounding environment.

For example, based on the light, viewing angle, distortion, and occlusion, the same object

appears differently in the image. Depending on the native image, pixels may not be a good

option for image recognition. This is because a minor change in each pixel leads to a major

change in the image, and thus the system is unable to recognize the objects correctly. The

target is to find a set of unique properties or features that do not change even with changing

pixel locations or values, as long as the structure of the object appears somewhere in the

image. Manually extracting features from images is a big challenge in image recognition.

This is why automatic approaches to feature extraction are becoming the alternative.

Because recognition of any object in any environment is complex at the current time,

the alternative way is to restrict the environment or the objects targeted. For example,

rather than recognizing all types of animals, we can just target a group of them. Rather than

working indoors and outdoors, we might restrict the environment to just indoor images.

剩余420页未读，继续阅读

swang09

粉丝: 0
资源: 19

深度学习与CNN在计算机视觉实践中的应用详解（Python+TensorFlow+Kivy）

Practical Computer Vision Applications Using Deep Learning with CNNs epub

Practical Computer Vision Applications Using Deep Learning with CNNs pdf

2019年新书-Practical Computer Vision Applications Using Deep Learning

Practical Computer Vision App Using D L with CNNs: With Detailed Examples in Py

Deep-Learning-For-Computer-Vision-全三册PDF以及start章节代码

Practical Machine Learning with Python(pdf书+sourcecode)

Udemy - Deep Learning Convolutional Neural Networks in Python

Bishop Pattern Recognition and Machine Learning

The Impact of OpenCV and Python Versions in Computer Vision Applications: A Case Study Exploring ...

【Advanced】Image Depth Estimation in MATLAB: Using Deep Learning for Image Depth Estimation

最新资源