IOP Conference Series: Materials Science and Engineering

PAPER • OPEN ACCESS You may also like

- Vehicle and Pedestrian Detection Based
Research on Daily Objects Detection Based on on Improved YOLOv4-tiny Model
Lingzhe Ma, Yu Chen and Jilin Zhang
Deep Neural Network - A novel Ship detection method from SAR
image with reduced false alarm
J Anil Raj, Sumam Mary Idicula and Binu
To cite this article: Sheng Ding and Kun Zhao 2018 IOP Conf. Ser.: Mater. Sci. Eng. 322 062024 Paul

- Small sample vehicle target recognition

method for unmanned aerial vehicle
system based on deep learning
Tingping Zhang and Di Wan
View the article online for updates and enhancements.

This content was downloaded from IP address 182.0.232.8 on 25/11/2021 at 01:02

SAMSE IOP Publishing
 
IOP Conf. Series: Materials Science and Engineering 322 (2018) 062024 doi:10.1088/1757-899X/322/6/062024
1234567890‘’“”

Research on Daily Objects Detection Based on Deep Neural

Network

Sheng Ding1,*, Kun Zhao2

1
Key Laboratory of Fiber Optic Sensing Technology and Information Processing,
Ministry of Education, Wuhan University of Technology, Wuhan, China
2
Fiberhome Telecommunication Technologies Co.,Ltd, Wuhan, China

*Corresponding author e-mail: [email protected]

Abstract. With the rapid development of deep learning, great breakthroughs have
been made in the field of object detection. In this article, the deep learning algorithm
is applied to the detection of daily objects, and some progress has been made in this
direction. Compared with traditional object detection methods, the daily objects
detection method based on deep learning is faster and more accurate. The main
research work of this article: 1. collect a small data set of daily objects; 2. in the
TensorFlow framework to build different models of object detection, and use this data
set training model; 3. the training process and effect of the model are improved by
fine-tuning the model parameters.

1. Introduction
Object detection is a very popular research direction in the vision field. Launched in 70s, object
detection began to be on track until 90s when computers became powerful and application plentiful. It
is easy for us as human to recognize objects in the images, however, things become difficult for
computers. Adding the different posture of objects and the complex environment around, object
detection is more ambiguity.
As we know, the evolution of detection algorithm is divided into two stages. Stage one is based on
the traditional features of the solution, and the second stage is the deep learning algorithm. Before
2013, most of the researches was based on the traditional feature optimization detection method. After
that, both academia and industry turned to deep learning algorithm.
With the increasing amount of detection data, the traditional detection method performance will
become saturated. The detection performance will gradually improve, yet the improvement decreases
after a certain amount of data. However, the method of deep learning is different. While the data of the
scene distribution accumulates, the detection performance promote continuously.
In this article, a set of data of daily supplies is collected, and then different training object detection
models are applied on the data. And by comparing the direct training and parameter adjustment model
training, it will be proved that the convergence speed and accuracy of object detection are improved
by adjusting the parameters.

2. Literature Survey
In ILSVRC 2014, deep learning increases the average object detection rate to 43.933%.
R-CNN proposed by Ross Girshick introduced CNN method into target detection field for the first
time. It introduces CNN method into target detection field. Selective Search[5] window extraction

Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
SAMSE IOP Publishing
 
IOP Conf. Series: Materials Science and Engineering 322 (2018) 062024 doi:10.1088/1757-899X/322/6/062024
1234567890‘’“”

algorithm is proposed to replace the traditional sliding window extraction method. Girshick proposed
Fast R-CNN model, which integrates feature extraction and classification into a classification
framework. In the R-CNN model, the deep convolution network for feature extraction and support
vector machines for classification are trained separately. The training time of Fast R-CNN model is 9
times faster than that of R-CNN model. The region proposal extraction and part of Fast-RCNN are put
into a network model RPN (region proposal net) in Faster R-CNN. The detection stage is very
convenient and the accuracy is similar with Fast R-CNN. YOLO (You Only Look Once) [6] proposed
by Joseph Redmon et al., is a one-time convolutional neural network predict multiple candidate frame
position and classification, target detection and recognition can achieve end to end. It solves object
detection as a regression problem. Based on a single end-to-end network, the output from the original
image to the position and category of the object is completed. R-FCN [7] is an accurate and effective
method for object detection. Compared with the previous region detection, R-FCN is based on the
whole image convolution calculation. In order to achieve this goal, the model uses a position-sensitive
score maps to balance the shift invariance in image classification and the translational transformation
in object detection.

3. Proposed Method
Object detection algorithm usually contains three parts. the first is the design of features, the second is
the choice of detection window, and the third is the design of classifier. Feature design methods
include artificial feature design and neural network feature extraction. The selection of detection
window mainly includes: Exhaustive Search [4], Selective Search [5], and RPN method based on deep
learning. This article adopts deep convolutional neural network (CNN) image feature extraction, using
the most advanced RPN as the detection window selection method, the bounding box regression
analysis, using softmax classification processing, and output the detection result. The model structure
is shown with the help of blocks in Figure 1.

CNN Feature Extraction RPN RoI Pooling Regression and Classification

Fig.1 Object Detection Model

3.1. CNN Feature Extraction

At present, the image feature extraction using CNN mainly includes three steps:
convolution ,activation and pooling.
Convolution is the core of feature extraction, and the feature of the image is obtained by different
convolution kernels. Convolution kernel is equivalent to a filter, and different filters extract different
features. Convolution is used to cross the convolution of the image with the size of 3*3 and the
number of 2n. The activation function generally uses ReLU to perform non-linear operation on the
output of the stacked layer, and has the function of accelerating convergence. Formula is as follows:

where x represents the input vector, w represents the parameters of a convolution kernel, b
represents the bias term, f represents the activation function, and y represents the output.
A pooling layer is placed behind each roll layer to reduce dimension. Generally, the output matrix
size of the original convolution layer is changed to half of the original one, which is convenient for the
operation at the back. In addition, the pooling layer increases the robustness of the system, turns the
original accurate description into a rough description, and avoids overfitting to some extent.

3.2. Region Proposal Networks

Region Proposal Networks (RPN) take the feature map extracted from the upper CNN as the input of
this layer, maps the midpoint of the feature map back to the original image, and designs these different

2
SAMSE IOP Publishing
 
IOP Conf. Series: Materials Science and Engineering 322 (2018) 062024 doi:10.1088/1757-899X/322/6/062024
1234567890‘’“”

fixed scale windows in the original design. According to the window and the ground truth
Intersection-over-Union (IoU) value to its positive and negative labels, let it learn whether there are
objects inside, so training a Region Proposal Network.
Only the approximate place need to be found, because that the precise positioning position and size
can be accomplished by following works. As the consequence the anchors can be fixed in three
aspects: fixed scale changes (three scales), fixed length and width ratio changes (three ratio), fixed
sampling method, only in the eigenvalues of each point in the original map of the corresponding
Region of Interest(RoI) on the sampling, in the back of the work can be adjusted. This can reduce the
complexity of the task.
After extracting the proposal on the feature map, the convolution calculation can be shared in front
of the network. The result of this network is that each point of the convolution layer has an output
about the k anchor boxes, including whether it is an object or not, adjusting the corresponding position
of the box. The RPN's overall Loss function can be defined as:
(2)

where the i denotes the i-th anchor, = 1 when the anchor is positive, and = 0 when the
anchor is negative. represents a ground true box coordinate associated with the positive sample
anchor (each positive anchor may correspond to only one ground true box: a positive sample anchor
corresponds to a ground true box, then the anchor with the ground true box is either the largest IoU of
all anchor, or greater than 0.7).

3.3. RoI Pooling

For traditional CNN, when the network is trained, the input image size must be a fixed value, and the
network output is also a fixed size vector or matrix. If the input image size is uncertain, the problem
becomes more cumbersome. There are two kinds of solutions: 1. Cut from the image part of the
incoming network; 2. Zoom the image into the desired size and then into the network. But the clipping
will destroy the complete structure of the image, and the zoom will destroy the original shape
information of the image. So Faster R-CNN proposed RoI Pooling to solve this problem.
First, the region proposal is mapped back to the original feature graph scale, and then each proposal
level and vertical are divided into k copies, each of which is max pooling processing. After this
processing, even if the size of the proposal, the output is k * k size, to achieve a fixed-length output.

3.4. Classification and regression

Using the region proposal feature map that has been obtained, the probability vector of each proposal
is calculated by the full connection layer and the softmax layer. At the same time, the bounding box
regression is used to obtain the position offset of each proposal again for the more accurate target
detection box.

4. Experimental Results

4.1. Data Set

There are many data sets in the field of object detection, such as the Oxford-IIIT Pets data set, the
PASCAL VOC data set, the Microsoft COCO data set. This article collects a small daily object dataset,
including the selection of 20 categories of items that are frequently touched in everyday life:
backpacks, handbags, umbrellas, beds, cups, chairs, sofas, potted plants, laptops, cell phones, books
Clocks, tables, toothbrushes, bottles, bananas, apples, bikes, cars and trucks.
we collect pictures of the 20 categories on the web, collect 100 pictures for each category, a total of
2000 images, and use the marking tool to mark. In 100 pictures, 80 pictures as a training set, 20
pictures as a validation set. This data set will be used in the next experiment.

3
SAMSE IOP Publishing
 
IOP Conf. Series: Materials Science and Engineering 322 (2018) 062024 doi:10.1088/1757-899X/322/6/062024
1234567890‘’“”

4.2. Experimental Environment

This experiment is based on the TensorFlow framework. TensorFlow is a deep learning library
supported by Google. The workflow is relatively easy, the API is stable and the compatibility is good.
One of its highlights is that it supports distributed computing of heterogeneous devices. It can be used
on various platforms Automatically run models, from mobile phones, single CPU / GPU to hundreds
of thousands of GPU cards consisting of distributed systems. This experiment runs on google cloud
machine learning platform, using 8 NVIDIA Tesla K80 GPU for parallel computing, greatly
accelerating the model convergence speed.

4.3. Experiment
In this article, the daily object detection data set is trained with different object detection models, and
the results of object detection under different models are obtained. The network structure is shown in
the following figure:

Fig.2 Faster_rcnn_resnet_101 Model

Fig.3 Faster_rcnn_resnet_152 Model

CONV RPN
Regression

Resnet_101 CONV Position_sensitive RoI

Classification

Fig.4 R-FCN_ resnet_101 Model

The network structure of each model is shown in Fig.2~Fig.4. Resnet_101[8] or Resnet_152[8]
represents the use of this convolutional network structure for feature extraction. RPN represents the
area to generate the network, CONV represents the convolution layer, FC represents the full
connection layer, and Position_sensitive RoI [7] represents the location sensitive pooling operation.
Table.1 Mean Average Precision Table
model mAP(%)
Faster_rcnn_resnet_101 69.5
Faster_rcnn_resnet_152 72.3
R-FCN_ resnet_101 69.8
Finetune Faster_rcnn_resnet_101 70.1
Table 1: Through the above experiment, we use the same data set to train different models, and get
the corresponding model of the mean average precision. It can be seen that the model trained by the
daily item identification data set is highly accurate in the test. But also can find the deep of the neural

4
SAMSE IOP Publishing
 
IOP Conf. Series: Materials Science and Engineering 322 (2018) 062024 doi:10.1088/1757-899X/322/6/062024
1234567890‘’“”

network, structure, parameter settings affect the accuracy of object detection. In the case of parameter
setting, the number of convolution kernel sizes, the learning rate setting, the regularization mode and
the gradient descent algorithm type play a key role in the final convergence rate and effect of the
model, and it is necessary to try to get good results, By repeated experiments, fine-tuning parameters
obtained by the model, compared with the original model accuracy is improved.

5. Conclusion
This article mainly established a small daily items detection data set. The data set is then trained on
different object detection models and has achieved good results in daily object detection. These well-
trained models can be used in the mobile platform, Nao robot platform or other intelligent devices, the
daily items to achieve accurate detection. In the future, we can get a better model by increasing the
capacity of the data set, the optimization of the model structure and the fine tuning of the parameters.

References
[1] Girshick R, Donahue J, Darrell T, et al. Rich Feature Hierarchies for Accurate Object Detection
and Semantic Segmentation[J]. 2013:580-587.
[2] Girshick R. Fast R-CNN[J]. Computer Science, 2015.
[3] Ren S, He K, Girshick R, et al. Faster R-CNN: Towards Real-Time Object Detection with Region
Proposal Networks.[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2015,
39(6):1137.
[4] P. F. Felzenszwalb, R. B. Girshick, D. McAllester, and D. Ra-manan. Object detection with
discriminatively trained part based models. TPAMI, 32:1627–1645, 2010.
[5] Sande KEAVD, Uijlings JRR, Gevers T, et al. Segmentation as selective search for object
recognition[C]// IEEE International Conference on Computer Vision. IEEE, 2012:1879-1886.
[6] Redmon J, Divvala S, Girshick R, et al. You only look once: Unified, real-time object
detection[C]. In: Proceedings of the IEEE Conference on Computer Vision and Pattern
Recognition. 2016: 779 -788.
[7] Dai J, Li Y, He K, et al. R-FCN: Object Detection via Region-based Fully Convolutional
Networks[J]. 2016.
[8] He K, Zhang X, Ren S, et al. Deep Residual Learning for Image Recognition[C]// Computer
Vision and Pattern Recognition. IEEE, 2016:770-778