深度学习驱动的目标检测：现状与进展
需积分: 16 87 浏览量
更新于2024-07-14
收藏 4.66MB PDF 举报
本文档《Object Detection With Deep Learning：A Review》是一篇综述性质的文章，主要关注深度学习在目标检测领域的应用和发展。随着计算机视觉和图像理解的紧密联系，近年来，目标检测已经成为研究热点。传统的目标检测方法依赖于手工设计的特征和浅层可训练架构，其性能提升往往受限于复杂的手工特征组合和多层次的图像理解。
文章指出，随着深度学习的快速发展，更加强大的工具被引入，这些工具能够学习到语义丰富的深层特征，从而解决传统架构中的问题。深度学习模型在网络结构、训练策略和优化函数方面展现出多样性。作者对基于深度学习的目标检测框架进行了全面的概述：
1. **历史回顾**：首先，作者简要介绍了目标检测的历史背景，包括从早期的传统方法（如Haar特征和级联分类器）到深度学习技术的兴起。
2. **深度学习的优势**：深度学习通过卷积神经网络（CNN）自动提取特征，能够学习到更高级别的抽象表示，这显著提高了目标检测的精度和鲁棒性。
3. **网络架构**：文章详细讨论了各种深度学习模型，如R-CNN系列（Region-based Convolutional Networks）、Fast R-CNN、Faster R-CNN、YOLO（You Only Look Once）以及其变体，如YOLOv2和YOLOv3，它们在实时性和准确性之间寻找平衡。
4. **训练策略**：深度学习模型的训练涉及到多任务学习（例如目标检测和定位）、迁移学习（预训练模型在新任务上的微调）以及数据增强等策略，这些都对最终性能有重大影响。
5. **优化函数与损失函数**：论文探讨了如何优化深度学习模型，如使用梯度下降优化器，以及设计适应目标检测任务的特定损失函数，如边界框回归和分类损失。
6. **实时性和效率**：随着深度学习的发展，目标检测模型不仅追求精度，也开始注重速度，因此文中也讨论了轻量级模型的设计和实时目标检测的重要性。
7. **挑战与未来方向**：尽管深度学习已经在目标检测领域取得了巨大进步，但文中也提到了一些尚未解决的问题，如小目标检测、遮挡物体识别和多目标跟踪，以及对计算资源的需求。
《Object Detection With Deep Learning：A Review》为读者提供了一个深入理解深度学习在目标检测领域的现状、方法和技术的全面视角，有助于研究人员和工程师进一步探索和优化这一关键领域的技术。
3216 IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTE MS, VOL. 30, NO. 11, NOVEMBER 2019
Fig. 4. Architecture of SPP-net for object detection [64].
However, more candidate boxes are required to achieve com-
parable results to those of R-CNN.
2) SPP-Net: FC layers must take a fixed-size input. That
is why R-CNN chooses to warp or crop each region proposal
into the same size. However, th e object may exist partly in
the cropped region and unwanted geometric distortion may be
produced due to the warping operation. These content losses or
distortion s will reduce recognition accuracy, especially when
the scales of objects vary.
To solve this problem, He et al. [64] took the theory of
spatial pyramid matching (SPM) [89], [90] into consideration
and proposed a novel CNN architecture named SPP-net. SPM
takes several finer to coarser scales to partition the image into
a number of divisions and aggregates quantized local features
into mid-level representations.
The architecture of SPP-net for object detection can be
found in Fig. 4. Different from R-CNN, SPP-net reuses
feature maps of the fifth conv layer (conv5) to project region
proposals of arbitrary sizes to fixed-length feature vectors. The
feasibility of the reusability of these feature maps is due to
the fact that the feature maps not only involve the strength of
local responses but also have relationships with their spatial
positions [64]. The layer after the final conv layer is referred to
as the SPP layer. If the number of feature maps in conv5 is 256,
taking a three-level pyramid, the final feature vector for each
region proposal obtained after the SPP layer has a dimension
of 256 × (1
2
+ 2
2
+ 4
2
) = 5376.
SPP-net not only gains better results with a correct estima-
tion of different region proposals in their corresponding scales
but also improves detection efficiency in the testing period
with the sharing of computation cost before SPP layer among
different proposals.
3) Fast R-CNN: Although SPP-net has achieved impressive
improvements in both accuracy and efficiency over R-CNN,
it still has some notable drawbacks. SPP-net takes almost the
same multistage pip eline as R-CNN, including feature extrac-
tion, network fine-tuning, SVM training, and bounding-box
regressor fitting. Therefore, an additional expense on storage
space is still required. In addition, the conv layers preceding
the SPP layer cannot be updated with the fine-tuning algorithm
introduced in [64]. As a result, an accuracy drop of very deep
networks is unsurprising. To this end, Girshick [16] introduced
a multitask loss on classification and bounding box regression
and proposed a novel CNN architecture named Fast R-CNN.
The architecture of Fast R-CNN is exhibited in Fig. 5.
Similar to SPP-net, the wh ole image is processed with conv
layers to produce feature maps. Then, a fixed-length feature
vector is extracted from each region proposal with an RoI
Fig. 5. Architecture of Fast R-CNN [16].
pooling layer. The RoI pooling layer is a special case of the
SPP layer, which has only one pyramid level. Each feature
vector is then fed into a sequence of FC layers before finally
branching into two sibling output layers. One output layer is
responsible for producing softmax probabilities for all C + 1
categories (C object classes plus one “background” class)
and the other output layer encodes refined bounding-box
positions with four real-valued numbers. All parameters in
these procedures (except the generation o f region proposals)
are optimized via a multitask loss in an end-to-end way.
The multitasks loss L is defined in the following to jointly
train classification and bounding-box regression:
L( p, u, t
u
,v) = L
cls
( p, u) + λ[u ≥ 1]L
loc
(t
u
,v) (1)
where L
cls
( p, u) =−log p
u
calculates the log loss for ground
truth class u,andp
u
is driven from the discrete probability
distribution p = ( p
0
, ···, p
C
) over the C +1 outputs from the
last FC layer. L
loc
(t
u
,v) is defined over the predicted offsets
t
u
= (t
u
x
, t
u
y
, t
u
w
, t
u
h
) and ground-truth bounding-box regression
targets v = (v
x
,v
y
,v
w
,v
h
),wherex , y,w, and h denote
the two coordinates of the box center, width, and height,
respectively. Each t
u
adopts the parameter settings in [15] to
specify an object proposal with a log-space height/width shift
and scale-invariant translation. The Iverson bracket indicator
function [u ≥ 1] is employed to omit all background RoIs.
To provide more robustness against outliers and eliminate the
sensitivity in exploding gradients, a smooth L
1
loss is adopted
to fit bounding-box regressors as follows:
L
loc
(t
u
,v) =
i∈x ,y,w,h
smooth
L
1
t
u
i
− v
i
(2)
where
smooth
L
1
(x ) =
0.5x
2
if |x| < 1
|x|−0.5otherwise.
(3)
To accelerate the pipeline of Fast R-CNN, another two
tricks are of necessity. On the one hand, if training sam-
ples (i.e., RoIs) come from different images, backpropagation
through the SPP layer becomes highly inefficient. Fast R-CNN
samples minibatches hierarchically, n amely, N images sam-
pled randomly at first and then R/N RoIs sampled in each
image, where R represents the number of RoIs. Critically,
computation and memory are shared by RoIs from the same
image in the forward and backward pass. On the other hand,
much time is spent in computing the FC layers during the
forward pass [16]. The truncated singular value decomposition
(SVD) [91] can be utilized to compress large FC layer s an d
to accelerate the testing procedure.
In the Fast R-CNN, regardless of region proposal genera-
tion, the training of all network layers can be processed in
a single stage with a multitask loss. It saves the additional
Authorized licensed use limited to: SOUTHWEST JIAOTONG UNIVERSITY. Downloaded on October 31,2020 at 01:20:24 UTC from IEEE Xplore. Restrictions apply.
剩余20页未读，继续阅读
189 浏览量
2021-09-25 上传
2019-08-26 上传
153 浏览量
点击了解资源详情
点击了解资源详情
点击了解资源详情
点击了解资源详情
点击了解资源详情
点击了解资源详情
yuppie__1029
- 粉丝: 8
- 资源: 12
我的内容管理
展开
最新资源
- 深入浅出：自定义 Grunt 任务的实践指南
- 网络物理突变工具的多点路径规划实现与分析
- multifeed: 实现多作者间的超核心共享与同步技术
- C++商品交易系统实习项目详细要求
- macOS系统Python模块whl包安装教程
- 掌握fullstackJS：构建React框架与快速开发应用
- React-Purify: 实现React组件纯净方法的工具介绍
- deck.js：构建现代HTML演示的JavaScript库
- nunn：现代C++17实现的机器学习库开源项目
- Python安装包 Acquisition-4.12-cp35-cp35m-win_amd64.whl.zip 使用说明
- Amaranthus-tuberculatus基因组分析脚本集
- Ubuntu 12.04下Realtek RTL8821AE驱动的向后移植指南
- 掌握Jest环境下的最新jsdom功能
- CAGI Toolkit：开源Asterisk PBX的AGI应用开发
- MyDropDemo: 体验QGraphicsView的拖放功能
- 远程FPGA平台上的Quartus II17.1 LCD色块闪烁现象解析
