深度学习新篇章：卷积神经网络在图像识别中的全面综述

需积分: 50 98 浏览量更新于2024-07-19 1 收藏 3.14MB PDF 举报

"深度卷积神经网络在图像分类任务中的应用与综合回顾" 深度卷积神经网络（Deep Convolutional Neural Networks，简称CNN）自20世纪80年代末期开始应用于视觉任务，但真正迎来爆发式发展是在21世纪初，这主要得益于计算能力的提升、大量标注数据的出现以及优化算法的进步。CNN如今已成为神经网络复兴的领头羊，尤其在图像分类领域取得了显著成果。 1. CNN的历史与先驱：CNN的前身可以追溯到早期的滤波器模型和局部连接神经网络。这些早期工作奠定了CNN的基础，例如Hubel和Wiesel的工作揭示了大脑皮层中视觉处理的局部和方向敏感特性，为CNN的设计提供了灵感。 2. CNN的复兴：2000年代中期，随着GPU计算能力的增强，以及如ImageNet这样的大型标注图像数据库的出现，CNN开始受到广泛关注。2012年，AlexNet在ImageNet图像识别挑战赛上的突破性表现，标志着深度学习时代的开启，CNN在此过程中起到了关键作用。 3. CNN的结构与原理：CNN的核心特点是卷积层、池化层和全连接层。卷积层通过可学习的滤波器对输入图像进行特征提取，池化层则用于降低数据维度，保持计算效率。全连接层将提取的特征映射到预定义的类别上。此外，ReLU激活函数、批量归一化和dropout等技术也对提升CNN性能至关重要。 4. CNN在图像分类的应用：CNN已经成为处理图像分类任务的标准工具，广泛应用于物体识别、人脸识别、医学图像分析、自动驾驶等多个领域。其强大的特征学习能力和对图像局部结构的捕获能力，使得CNN在处理复杂视觉问题时表现出色。 5. 近期的进展与挑战：近年来，研究者不断提出新的架构，如ResNet（残差网络）、DenseNet（稠密连接网络）和Inception系列，以解决深度网络中的梯度消失和过拟合问题。同时，模型的迁移学习和微调策略也使得CNN能在有限的训练数据下取得良好效果。然而，CNN仍面临计算效率、模型解释性和泛化能力等方面的挑战。 6. 展望未来：未来的研究可能会集中在开发更高效、可解释的CNN变体，利用无监督学习或自我监督学习来减少对标注数据的依赖，以及探索如何结合其他机器学习方法（如强化学习和生成对抗网络）以提升整体性能。深度卷积神经网络在图像分类领域的应用和研究已经取得了显著的成就，但这个领域仍有大量的探索空间，未来有望继续推动计算机视觉和人工智能的发展。

NECO_a_00990-Rawat MITjats-NECO.cls May 31, 2017 22:33

Uncorrected Proof

16 W. Rawat and Z. Wang

Figure 5: DCNN architecture split over two GPUs (Krizhevsky et al., 2012).

since become the leading architecture for most visual tasks, in particu-

lar, for image classication–related applications, as the rest of this review

shows.

Central to their success, they implemented several novel and unusual

techniques. Rather than using traditional sigmoid or hyperbolic tangent

activation functions, they were inspired by Jarrett et al. (2009) and used

ReLU (Nair & Hinton, 2010) activations, which allowed much faster train-

ing times (see section 5.2.1 for further details). Since their network was too

big to t into one GPU, they spread it over two GPUs arranged in a parallel

conguration, which was similar to the multicolumn DCNNs proposed by

Ciresan, Meier, and Schmidhuber (2012). Inspired by the local contrast nor-

malization of Jarrett et al. (2009), they applied local response normalization.

Denoted mathematically, if a kernel i,atposition(x;y)isusedtocompute

the activity of a neuron denoted b a

x,y

and the ReLU nonlinearity is than

applied, the response-normalized activit b

x,y

can be expressed as

x,y

= a

x,y



k + α

min(N−1,i+n/2)



j=max(0, i−n/2)

x,y

)



, (4.1)

where N is the total number of kernels in the layer and the sum runs over

n “adjacent” kernel maps at the same spatial position. This scheme aided

generalization and reduced their network classication error rates. They

further reduced the classication error by overlapping the network’s max

pooling layers. Figure 5 illustrates the revolutionary architecture presented

by Krizhevsky et al. (2012). It consisted of ve convolutional layers, three

of which were followed by max pooling layers, and three fully connected

layers. The various layer parts in the top half of the gure ran on one GPU,

while the layer parts at the bottom ran on the second GPU. The GPUs in-

teracted with each other only at specic layers.

NECO_a_00990-Rawat MITjats-NECO.cls May 31, 2017 22:33

Uncorrected Proof

Deep Convolutional Neural Networks for Image Classication 17

To overcome overtting, the authors employed a regularization tech-

nique known as Dropout (Hinton et al., 2012). Specically, when each train-

ing case was presented to the network during the training phase, each

hidden neuron was randomly omitted from the network with a probability

of 0.5. Thus, hidden neurons could not rely on other hidden neurons be-

ing present, and this prevented complex coadaptations of features on the

training data. At test time, all of the hidden neurons were used, but their

outputs were multiplied by 0.5 to compensate for the fact that double the

number of neurons were now active. The result of this was a strong regular-

ization effect that signicantly reduced overtting (Krizhevsky et al., 2012;

Hinton et al., 2012; Srivastava, Hinton, Krizhevsky, Sutskever, & Salakhut-

dinov, 2014). Figure 11 (in section 5.4.2) shows the effect of Dropout on a

standard feedforward network, with two hidden layers, while section 5.4.1

gives a more formal description of the technique and introduces several of

its variants.

Overtting was further reduced by applying data augmentation, a pop-

ular procedure to articially enlarge a data set (LeCun et al., 1998; Simard

et al., 2003; Ciresan et al., 2011, 2012). In particular, they created more im-

ages by applying translations and horizontal reections to the training im-

ages, altering the intensities of their color channels, and performing princi-

pal component analysis (PCA) on their pixel values, which led to improved

classication performance. Krizhevsky et al.’s (2012) model has been used

extensively for various purposes since its development. Vast amounts of re-

search have used it to benchmark their models against or as a base model

to test new algorithms. Furthermore, their model has inspired DCNN work

and has become one of the major contributors to the recent rise in DCNN

technology for image classication–related applications.

Notably, preceding the pioneering work of Krizhevsky et al. (2012) was

the series of work proposed by Ciresan et al. (2011, 2012). They presented

deep hierarchical CNNs, trained in a fully supervised fashion, that achieved

the best published results on the NORB (LeCun et al., 1998; Krizhevsky,

2009; Coates, Lee, & Ng, 2011) and MNIST (LeCun et al., 1998; Ciresan,

Meier, Gambardella, & Schmidhuber, 2010) classication benchmarks (Cire-

san et al., 2011). By stacking these DCNNs into columns, they further im-

proved the state of the art for these benchmarks and, in particular, reached

human-level performance on the MNIST data set. Furthermore, on the Ger-

man trafc sign recognition benchmark (GTSRB; Stallkamp, Schlipsing,

Salmen, & Igel, 2011), they surpassed human performance by a factor of

two.

Table 1 summarizes the key attributes of the image classication data

sets introduced thus far. Although other classication data sets exist (see

sections 5.1.2.3 and 5.3.2), these are the most commonly used for DCNN

evaluation and benchmarking. Among them, the MNIST data set (LeCun

et al., 1998) has stood the test of time and become the most popular,

NECO_a_00990-Rawat MITjats-NECO.cls May 31, 2017 22:33

Uncorrected Proof

Deep Convolutional Neural Networks for Image Classication 19

although modern classication systems are judged by their success on the

ILSVRC (Russakovsky et al., 2015), as the next section shows.

4.4 Representative Improvements Exemplify DCNN Dominance.

Since the groundbreaking work of Krizhevsky et al. (2012), DCNNs have

dominated image classication tasks, in particular the ILSVRC (Rus-

sakovsky et al., 2015). In fact, they were triumphant in every ImageNet

classication challenge since 2012 (Simonyan & Zisserman, 2014; Zeiler &

Fergus, 2014; Szegedy, Liu et al., 2014; He, Zhang, Ren, & Sun, 2015b). In an

attempt to understand them and derive ways to improve their performance,

Zeiler and Fergus (2014) introduced a new visualization technique using a

multilayered deconvolutional network (Zeiler, Taylor, & Fergus, 2011) that

provided vision into the intermediate feature extraction layers of the net-

work. They used this in a diagnostic role to improve the DCNN architec-

ture and performance of Krizhevsky et al. (2012). Thus, when compared

to Krizhevsky et al. (2012), their model achieved better results on the Ima-

geNet classication benchmark, and multiple models were averaged to win

the ILSVRC 2013 (Russakovsky et al., 2015). Furthermore, their model gen-

eralized excellently, and they demonstrated this by achieving the best pub-

lished results on the CALTECH-101 (Fei-Fei et al., 2006) and CALTECH-

256 data sets (Grifn et al., 2007). Although their visualization technique

worked well on relatively higher-dimensional color images, it would be

interesting to test its applicability on the popular MNIST data set (LeCun

et al., 1998), since it is conceivable that the deconvolutional network may not

be able to reproduce the lower-dimensional gray-scale MNIST images with

the same accuracy. Another interesting direction from using the extracted

features in a diagnostic role will be to investigate its applicability to solve

some of the remaining DCNN challenges, specically like those mentioned

in sections 6.2 and 6.4.

Szegedy, Liu et al. (2014) introduced a DCNN architecture that they

called the Inception model. A particular incarnation of this model,

GoogLeNet, produced outstanding image classication and object detec-

tion results, winning both the ImageNet classication and detection chal-

lenges in 2014 (Russakovsky et al., 2015). Their success was brought about

by using a very large network, consisting of 22 layers. Since the cost of this

is a larger number of parameters, which makes the network more prone

to overtting and a considerably larger computational burden, they used

a carefully engineered design, based on Hebbian principles, that allowed

them to move from a fully to sparsely connected convolutional architecture,

which was motivated by the ndings of Arora, Bhaskara, Ge, and Ma (2014).

Specically, their architecture used 1 × 1 convolutions heavily, inspired by

Lin et al. (2013), to perform two functions. Most signicant, they served as

dimension-reduction blocks prior to the more computationally costly 3 ×3

and 5 ×5 convolutions, and they included the use of rectied linear activa-

tions (Nair & Hinton, 2010), thus making them dual purpose. Subsequently,

NECO_a_00990-Rawat MITjats-NECO.cls May 31, 2017 22:33

Uncorrected Proof

20 W. Rawat and Z. Wang

they were able to increase the depth and width of their network, while only

marginally increasing the computational cost. Figure 8 (in section 5.1.1.1)

illustrates an Inception module, which incorporates the dimension reduc-

tion 1 ×1 convolution lters, illustrated by the bevel in the diagram. These

modules are the building blocks of the Inception model, which has since

been improved several times, as discussed in section 5.1.1.2.

Similar to Szegedy, Liu et al. (2014), Simonyan and Zisserman (2014), the

runners-up in the same ILSVRC 2014 classication contest (Russakovsky

et al., 2015), also used a very deep DCNN, which consisted of 19 layers

compared to the 22 of their competitors. However, asserting that the Incep-

tion model was too complex, they kept all the parameters of their DCNN

architecture constant and steadily increased the depth alone. This was made

feasible by using smaller-sized convolutional (3 × 3) lters throughout the

network, which was inspired by Ciresan et al. (2011), who already used

smaller kernels, albeit for shallower networks applied to simpler tasks.

The winners of the ILSVRC 2015 (Russakovsky et al., 2015), He et al.

(2015b), used an even deeper DCNN, when compared to Simonyan and

Zisserman (2014) and Szegedy, Liu et al. (2014). In fact, their model was

ultra-deep in that it consisted of 152 layers. Since deeper models are harder

to train and suffer from degradation (of training and thus test accuracy)

(He et al., 2015b; He & Sun, 2015; Srivastava, Greff, & Schmidhuber, 2015a,

2015b), they introduced a new residual learning framework.

They refor-

mulated the layers of the network and forced them to learn residual func-

tions with reference to their preceding layer inputs rather than learning

unreferenced functions. This allowed errors to be propagated directly to

the preceding units, and thus made these networks easier to optimize and,

although they were ultra-deep, easier to train. They tried different residual

module congurations and network architecturesand found that optimized

residual modules worked more optimally compared to their initial mod-

ules. Figure 6 compares the difference between the original residual module

and its optimized successor, which resulted in faster computation. As illus-

trated, ReLUs (Nair & Hinton, 2010) feature heavily in both versions; how-

ever, optimized residual connections make use of the dimension-reducing

1 × 1 lters to cushion computation. A more formal description of the orig-

inal residual learning technique, as well its improvements, is discussed in

section 5.5.4.

Table 2 illustrates the performance of DCNNs in the ILSVRC since its in-

ception. Signicantly, the table highlights DCNN domination over earlier

methods that used feature extraction and compression, followed by clas-

sication with a shallow classier (Perronnin, Sánchez, & Mensink, 2010;

Degradation is caused by the poor propagation of activations and gradients because

of stacking several nonlinear transformations on top of each other.

剩余97页未读，继续阅读

RoaringKitty

粉丝: 6w+
资源: 26

深度学习新篇章：卷积神经网络在图像识别中的全面综述

深度卷积神经网络详解

深度学习之卷积神经网络CNN

《动手学习深度学习》之三：1.卷积神经网络（CNN）基础（打卡2.4）

深度卷积神经网络.zip_卷积神经网络_深度学习 检索_深度神经网络_脑 深度学习_语音识别深度

深度卷积神经网络.pptx

卷积神经网络与深度卷积神经网络

深度卷积神经网络属于卷积神经网络吗

深度卷积神经网络与卷积神经网络的区别

深度卷积神经网络和普通卷积神经网络的对比

卷积神经网络与深度卷积神经网络有什么区别

最新资源

深度卷积神经网络.zip_卷积神经网络_深度学习检索_深度神经网络_脑深度学习_语音识别深度