Using HHT method to extract EEG signals feature, these
signals are synthesized into the sum of the basic signals before putting
into the neural network, which is a new proposal of the thesis as well
as helps the classification of signal samples to be fast and accurate rate
92.4%.
Reducing the number of channels that help reduce processing
time, reducing the number of signal channels base on the
characteristics of each electrode position in the scalp as well as
experiments.
The combination of camera to detect the eye direction helps
the system run stably and helps the trainees quickly become more
proficient with wheelchair control.
Using EEG signals in the field of control is also a new
proposal in the thesis because at present, serveral publics in the
country only research the theory of EEG signals, filter noise, and use
blink in identifying and detecting.
Building an EEG signals acquisition software and experiment
on wheelchair control model.
The author has built a wheelchair control system through EEG
signals processing that control the wheelchair physical model as
Figure 5.1. Combined on controlling the wheelchair model and
processing of continuous signals.
33 trang |
Chia sẻ: honganh20 | Ngày: 22/02/2022 | Lượt xem: 378 | Lượt tải: 0
Bạn đang xem trước 20 trang tài liệu Neural network in the wheelchair control system for severe disabilied people using eeg signal and camera, để xem tài liệu hoàn chỉnh bạn click vào nút DOWNLOAD ở trên
ata before putting in multilayer neural network to classify
these partterns.
Combining EEG signals and camera processing aim to identified
and classified process more easily and effectively.
1.5 Object and scope of the research
The main object in thesis is to use multilayer neural network to
classify 5 EEG signal partterns into control commands corresponding
to 5 commands to control the wheelchair as: Forward, Turn right, Turn
left, Reverse and Stop. In thesis also mention to image processing to
detect the eye direction aim to help the system work more easily and
effectivly. However, the thesis don’t focus more on image processing,
but on EEG signal processing.
1.6 The contribution of thesis
1.6.1 The contribution about theory
Find out the observation board which easy to use to collect data,
combine scientifically between feature extraction algorithm and
cluster data before putting into neural network to classify data
partterns.
1.6.2 Practical contribution
The experimental result of thesis show that the EEG signal
partterns classification through eye observation (with differnce image
partterns), for people who has mind and eyes as normal people could
absolutely performance.
CHAPTER 2 - THEORETICAL BASIS
2.1 EEG signal and its characterizations
Delta wave (0 – 3 Hz), the highest amplitude as figure 2.1, it often
appear at the child up to 1 year old and adult when sleeping, well sleep.
It represents the grey matter of the brain. This wave usually appears
everywhere on the scalp.
Trang 5
Figure 2.1 The Delta wave
The Theta wave (3 – 7 Hz), it appear when the eye are closed and
the mind is in a relax state as Figure 2.2, it apperas in adult or when
awake in the elderly and often appears in the temples.
Figure 2.2 The Theta wave
The Alpha wave (7 – 13 Hz), it appear often in elder as Figure
2.2, Alpha wave usually appear on both sides of scalp but having an
uneven amplitude, this waves appear when the eye are closed (a state
of relaxation) and often disappear when the eyes are opened or under
stress.
Figure 2.3 The Alpha wave
The Beta waves (13 – 30 Hz), with low amplitute as Figure 2.4,
This waves usually appears in patients who are often in a state of
alertness, prevention, and anxiety this waves are distributed
symmetrically on both sides and most clearly at the front, it usually
appear in front and at the top of the cerebral cortex, the amplitude is
less than 30uV.
Figure 2.4 The Beta wave
Trang 6
The Gamma wave (30 – 45 Hz), it often referred to as fast Beta
wave. This wave usually has low amplitude and rarely appears, but the
discovery of this wave plays a important role in identifying
neurological diseases that occur in the center of the cerebral cortex.
Figure 2.5 The Gamma wave
2.2 The electrode positions on the scalp
The brain is one of the largest and most complex organs in the
human body, it is made up of over 100 billion nerves, communicating
with 1,000 billion synapses. The electrode positions are mounted on
the scalp according to international standard 10/20 as Figure 2.6.
Figure 2.6 Electrode positions according to international
standard 10/20
CHAPTER 3
CONTROL MODEL CONTRUCTION
To begin the research process, author used the database provided
on the prestigious University of San Diego (UCSD) website of the
USA, ranked 38 in the world in 2018, this data was obtained from
participants when looked at 5 different image objects (human, city,
landscape, flower and animal), with 8 color bits and size (256 pixels
Trang 7
wide and 384 pixels high), the total number of samples is 21,000.
In this chapter, author present to built the model from simple to
complex step by step, and then evaluate the experimental results on
database provided by the University of San Diego (UCSD), and from
80 students from HITU to make clarify the contribution and scientific
meaning of the thesis.
3.1 Single-Layer neural network model
At first, author built a single neural network to separate two
parttern of signals (animals and not animals), the purpose of this study
is to evaluate whether the neural network meets the classification
requirements, thesis used Matlab software for this experimental
process.
The features of EEG database is extracted by Wavelet transform
(Mexico hat) and used a single neural network to identify. The system
model is shown in Figure 3.1, this model included of 2 stages: stage 1:
preprocessing raw data signals and then synthesize into 5 basic EEG
signals Delta, Theta, Alpha, Beta and Gamma. Stage 2: builting a
single neural network with 5 inputs corresponding to 5 basic EEG
signals: Delta, Theta, Alpha, Beta, Gamma and one output to
determine the clasified results.
Figure 3.1 Single-Layer neural network model
The training process is performed on the database with the
following parameters:
Learning rate: 0.7.
Raw data
Classified result
Pre-processing
Single-Layer Neural
Network
Trang 8
Initial random weight in the range from -0.5 to 0.5
The error threshold is 1x10-5 based on MSE (Mean Square
Error).
The maximun number loop: 5.000.
Experimental results of identification on the database are shown
in Table 3.1.
Table 3.1 Experimental results on database
Image types Animal/Landscape
Identification
Rate
France Landscape 99,13%
Wild sheep Animal 98,67%
Wild cats Animal 99,28%
Bali, Indonesia Landscape 62,44%
Wild animals Animal 99,64%
California Coasts Landscape 56,89%
Wolves Animal 98,64%
Mushrooms Landscape 95,16%
Kenya Animal 99,76%
The big Apple Landscape 98,79%
Snakes, lizards... Animal 98,32%
Caves Animal 67,18%
Polar bears Animal 99,03%
Exotic Hong Kong Landscape 98,72%
Images of France Landscape 99,37%
Fabulous fruit Landscape 98,25%
Wild animals Animal 93,97%
Sand & solitude Animal 98,42%
Lions Animal 62,78%
Trang 9
Image types Animal/Landscape
Identification
Rate
Great Silk Road Landscape 98,47%
From the experimental results in Table 3.1, author found that the
average accuracy rate of the identification results on the database was
91.15%.
3.2 Multi-Layer neural network model
3.2.1 System Model
Based on the results achieved from the single neural network
model, author continued to develop a multi neural network model with
the results of classifying 5 EEG signal partterns corresponding to 5
control signals with accurate rate 93.57%. Table 3.2 describes the
result of 05 control commands corresponding to the equivatent image
types.
This model uses Wavelet transform to noise signals and extract
features, then using K-mean algorithm to cluster the characteristics of
the signal partterns and then put into the multi-layer neural network to
classify, in this model, author chooses 10 channels to reduce
processing time and enhance performane. System model is shown in
Figure 3.2.
Figure 3.2 Multi-Layer neural network system model
EEG Signal
Selecting Channel
Wavelet Transform
Clustering
Muli-Layer Neural Network
Classified Result
Trang 10
3-layer neural network model is shown in Figure 3.3.
The first layer contains five nodes which are Delta, Theta, Alpha,
Beta and Gamma, these classes is called the input layer.
The second layer is the hidden layer, the number of nodes in the
hidden layer is set to 5, 10, 15, 20, 25, 30, 35, 40, 45 and 50.
The output layer contains a node, the result of this node is used
to classify the EEG signal. The activation function used in this model
is hyperbolic tangent, the value of the output is in the range [-1, 1].
Figure 3.3 Multi-Layer neural network model
Before using the model, the neural network needs to training
stage. The training algorithm is shown in figure 3.4, the model used
backpropagation algorithm.
Trang 11
Figure 3.4 Neural network training algorithm
3.2.2 Experimental results with multi-layer neural network
model
The dataset consists of 21,000 samples divided into subsets for
training (70%), validation (15%) and testing (15%). The system uses
Matlab and EEGLab tools for the testing process, the neural network
is divided into two test stages.
The training stage is performed on the training database, using
structures with different hidden nodes in the hidden layer of neural
network, with the following parameters as follows:
Learning rate: 0.7.
Initial random weight in the range from 0 to 1
The error threshold is 1x10-5 based on RMSE (Root Mean
Initial random weight
Get value Delta, Theta, Alpha,
Beta, Gamma
Calculate the value of nodes in hidden layer
Calculate the ouput value of nodes in the hidden layer
Calculate the input value of the output nodes
Calculate the output value of the output nodes
Calculate the error of the output layer
Calculate the error of the hidden layer
Calculate the error of the
system
Error system <= threshold value?
End
Update weight
Begin
True
False
Trang 12
Square Error).
The maximun number loop: 5.000.
The accuracy of classification is measured by the ratio of result of
wrong classification to the total number of samples follow formula
(3.1).
=
−
(3.1)
Where n is the total number of samples, ntrue is the number of
samples with correct classification results.
Table 3.2 Experimental results
Number Nodes in
hidden layer
Average
Error
Minimum
Error
Accurate
Rate
5 25.21% 21.98% 78.02%
10 23.77% 20.04% 79.96%
15 20.44% 17.13% 82.87%
20 17.76% 14.21% 85.79%
25 15.43% 11.88% 88.12%
30 12.98% 10.06% 89.94%
35 9.87% 7.74% 92.26%
40 7.74% 6.43% 93.57%
45 9.56% 7.92% 92.08%
50 10.24% 8.63% 91.37%
Observing in Table 3.2 and see that the classification results of
the database are constantly increasing until the best possible value (40
Nodes in the hidden layer and the accurate rate is 93.57%), After that,
the result values begin to decrease as the number of nodes in the
hidden layer increases, This is called overfitting. A confusion matrix
contains about actual classifications and predictions made by system.
Table 3.3 describes the performance of the system evaluated by the
database in the matrix confused with 40 nodes in the hidden layer.
Table 3.3 Matrix confusion of classification results
Trang 13
Actual
Pred
ictio
n
Anima
l
Lands
cape
City
Huma
n
Flower
Animal 93.8% 1.4% 1.5% 2.0% 1.6%
Landscape 1.1% 93.6% 3.1% 1.5% 1.2%
City 1.6% 1.3% 93.5% 1.6% 1.8%
Human 1.9% 1.8% 1.3% 93.2% 1.7%
Flower 1.7% 1.9% 0.8% 1.8% 93.8%
To provide a more intuitive and easy-to-understand about result
of predictive quality, the following formulas are used for effective
quality test.
Accuracy (AC) is an accurate prediction rate. It is determined
using the formula (3.2).
=
+
+ + +
(3.2)
Precision (P) is the ratio of correct times that are predicted to be
accurated, calculated using the formula (3.3).
=
+
(3.3)
Where, True Possitive (TP) refers to the correct database that are
correctly classified to be true. True Negative (TN) refers to incorrect
database that are incorrectly classified to be false. False Possitive (FP)
refers to incorrect database that are incorrectly classified to be true.
False Negative (FN) refers to incorrectly categorized database to be
false.
The rate of identification with 40 hidden nodes in the hidden layer
is given in table 3.4
Table 3.4 Experimental results
TP TN FP FN AC P
Animal 93,8% 93,6% 6,4% 6,2% 93,7% 93,6%
Landscape 93,6% 93,2% 6,8% 6,4% 93,4% 93,2%
City 93,5% 93,8% 6,3% 6,6% 93,6% 93,7%
Trang 14
Human 93,2% 93,4% 6,6% 6,8% 93,3% 93,4%
Flower 93,8% 93,9% 6,1% 6,2% 93,9% 93,9%
These results are also compared with previous studies such as
determining EEG based on winking with 15,360 samples and reaching
90.85%, with decision tree reach the rate 85%, based on eye
movement by 2 experiments with 3,600 samples and 8,320 samples
reaching acuracy rate 85%.
3.3 Design of synthetic model for signal processing
This model was developed from a multi-layer neural network
model in section 3.2. Beside of identifying the EEG signal, it also
combined the user's eye direction signal through the camera, this
model focuses on the items as below:
Converting EEG signals using Hilbert Huang Transform
(HHT) method to reduce noise signal because HHT conforms to EEG
signal and for better results than other transform methods.
Eye direction recognition based on the user's face image
combined with EEG signal recognition to improve the effectiveness
of the identification.
Design a system that includes hardware and software for
testing in realtime.
Experimental data was collected from 80 students of HITU
who volunteered to participate.
The system consists of 3 blocks as shown in Figure 3.5.
Trang 15
Figure 3.5 Synthetic system model
The first block is the EEG signal recognition block to extract
5 features.
The second block is the eye direction recognition block by
recognizing the eyes and eyebrows from the user's face image to
extract 4 features.
The third block is a multi-layer neural network with 9 input
nodes (4 for camera and 5 for EEG), 5 output nodes are classified
equivalent 5 control commands as “FORWARD”, “TURN RIGHT”,
“TURN LEFT”, “REVERSE” and “STOP”.
3.3.1 EEG signal recognition block
This block behaves like the model in section 3.2, at first, selecting
useful information channels and remove channels with redundant
information, next, using HHT to extract feature and eliminate noise
signals, and then, use the K-Means algorithm to cluster data.
3.3.2 Identifing eye direction signal
Face images are received from the camera and are cropped in the
area of eye information to reduce processing time, after that, the face
image is converted into a binary image that satisfies the requirements
Face image
Converting to
binary image
Detecting eye and
eyebrow
Rating between
eye and eyebrow
Extracting 4
features
EEG signals
Selecting channels
Hibert Huang
Transform
Clustering
Extracting 5
features
MULTILAYER NEURAL
NETWORK
Trang 16
for showing both eyebrows and eyes as shown in Figure 3.6. Threshold
values are calculated by isodata algorithm.
Figure 3.6 Detect eyes and eyebrows
Using image segmentation algorithm to detect the center of the
eye, the segment of eye and eyebrows as Figure 3.7.
Figure 3.7 Pupil, segmented eyes and eyebrows
Calculate the ratio of left eye and left eyebrow according to
formula (3.4) as Figure 3.8
=
(3.4)
Figure 3.8 The ratio of left eye and left eyebrow
Calculate the ratio of right eye and eyebrow according to formula
(3.5) as Figure 3.9.
=
(3.5)
Trang 17
Figure 3.9 Right eye ratio and right eyebrow
Calculate the percentage of pupil center of the left eye and the
length of the left eye by the formula (3.6) as Figure 3.10.
=
(3.6)
Figure 3.10 Rate of pupil and eye length
Calculate the proportion of the pupil center of the right eye and
the length of the right eye according to formula (3.7) as Figure 3.10.
=
(3.7)
3.3.3 Multilayer neural network model backforwards
The backpropagation multi-layer neural network model consists
of 3 layers as Figure 3.11.
Trang 18
Figure 3.11 Multi-layer neural network model
The first layer contains 9 nodes like Delta, Theta, Alpha, Beta,
Gamma, d1, d2, d3 and d4. This class is called the input class.
The second layer is the hidden layer, the number of nodes in
the hidden layer is 11 nodes.
The output layer contains 5 nodes, the result of this node is
used to classify EEG signals. Because the activation function is
hyperbolic tangent, the value of the output node is between [-1, 1].
Since the output has 5 nodes, which one has the largest value, it will
be selected and that is the control signal.
3.3.4 Select the data set and experimental results
Experimental data was collected from 80 volunteering students
of Ho Chi Minh City College of Industry and Trade (HTIU). Students
wear the Emotiv EEG device and sit 120 cm away from the
observation board. Experimental data is divided into 3 data sets as
follows:
Training data set was collected from 70% data of 60 students.
1
2
3
4
5
6
7
8
9
1
2
3
4
5
6
7
8
9
10
11
1
2
3
4
5
Input Layer
Hidden Layer
Output Layer
Trang 19
The first test data set was collected from 30% remaining data
of 60 student.
The second set of test data was collected from the remaining
20 students.
After training the neural network from the training database,
experimental results of the first test data set are shown in the confusion
matrix, shown in Table 3.5 and identification rate in Table 3.6.
Table 3.5 The confusion matrix of the result from the first database
classification
Actual classification
Predic
tive
Classi
ficatio
n
Huma
n
Anim
al
Flowe
r
City
Lands
cape
Human 91,2% 1,6% 3,1% 0,9% 2,7%
Animal 1,9% 91,1% 1,2% 3,2% 2,0%
Flower 2,8% 2,5% 92,8% 1,7% 1,5%
City 2,4% 2,1% 0,7% 92,1% 1,9%
Landscape 1,7% 2,7% 2,2% 2,1% 91,9%
Table 3.6 Experimental results on the first database
TP TN FP FN AC P
Human 91.2% 91.3% 8.7% 8.8% 91.3% 91.3%
Animal 91.1% 91.1% 8.9% 8.9% 91.1% 91.1%
Flower 92.8% 92.6% 7.4% 7.2% 92.7% 92.6%
City 92.1% 92.2% 7.8% 7.9% 92.2% 92.2%
Landscape 91.9% 92.1% 7.9% 8.1% 92.0% 92.1%
The experimental results of the second test database are shown in
the confusion matrix in Table 3.7 and the identification rate in Table
3.8.
Trang 20
Table 3.7 Confusion matrix of classification result for the second
database
Actual classification
Predic
tive
Classi
ficatio
n
Human
Anim
al
Flowe
r
City
Lands
cape
Human 90.7% 1.9% 2.7% 1.1% 3.1%
Animal 1.4% 90.8% 2.1% 2.6% 2.4%
Flower 2.3% 2.3% 92.3% 3.2% 1.7%
City 3.1% 2.6% 1.2% 91.6% 1.3%
Landscape 2.5% 2.4% 1.7% 1.5% 91.5%
Table 3.8 Experimental results on the second test dataset
TP TN FP FN AC P
Human 91.2% 91.3% 8.7% 8.8% 91.3% 91.3%
Animal 91.1% 91.1% 8.9% 8.9% 91.1% 91.1%
Flower 92.8% 92.6% 7.4% 7.2% 92.7% 92.6%
City 92.1% 92.2% 7.8% 7.9% 92.2% 92.2%
Landscape 91.9% 92.1% 7.9% 8.1% 92.0% 92.1%
Experimental results for eye direction signals are shown in Table
3.9 and for EEG signals shown in Table 3.10. Table 3.11 compares 3
experimental results. The graph in Figure 3.12 shows the chart of two
signals when identifying them separately.
Table 3.9 Experimental results on eye direction signals
TP TN FP FN AC P
Human 85.1% 85.7% 14.3% 14.9% 85.4% 85.6%
Animal 84.5% 84.1% 15.9% 15.5% 84.3% 84.2%
Flower 87.3% 86.3% 13.7% 12.7% 86.8% 86.4%
City 83.6% 84.0% 16.0% 16.4% 83.8% 83.9%
Landscape 84.2% 83.2% 16.8% 15.8% 83.7% 83.4%
Trang 21
Table 3.10 Experimental results on EEG signals
TP TN FP FN AC P
Human 90.2% 89.9% 10.1% 9.8% 90.1% 89.9%
Animal 90.3% 90.0% 10.0% 9.7% 90.2% 90.0%
Flower 92.3% 91.8% 8.2% 7.7% 92.1% 91.8%
City 90.7% 90.4% 9.6% 9.3% 90.6% 90.4%
Landscape 90.4% 90.5% 9.5% 9.6% 90.5% 90.5%
Table 3.11 Experimental results of 3 identification methods
EEG and Camera EEG Camera
Human 90,8% 90,1% 85,4%
Animal 90,9% 90,0% 84,3%
Flower 92,2% 91,8% 86,8%
City 91,7% 90,4% 83,8%
Landscape 91,7% 90,5% 83,7%
Figure 3.12 Experimental results on 3 separate methods
Trang 22
CHAPTER 4
CONSTRUCTING SOFTWARE AND HARDWARE TO
CONTROL THE WHEELCHAIR MODEL
In this chapter, author built the wheelchair hardware and software
system based on the models presented in Chapter 3. Then evaluate the
experimental results to clarify the practical contribution of the thesis.
4.1 Wheelchair control software system
Wheelchair control software is designed on Visual Studio C#
2015, the software interface includes 4 functions as follows:
4.1.1 Login to the system
In this section, it is mandatory for anyone who uses software to
have an account to log in to the system, personal account used to
manage the database of brain signals, training time, accurate rate
during the control, the interface of the login part in Figure 4.1.
Figure 4.1 System login interface
4.1.2 Wheelchair control training
For a person who has never participated in control, this step must
be done, as well as someone who has never drive, they had to learn to
drive. In order to control the wheelchair, the participants have to
control 5 commands with the acuary rate of more 90% for each
command. The purpose of this training is to help participants become
familiar with wheelchair control and concentrate in control. The
training software interface in Figure 4.2.
Trang 23
Figure 4.2 Training software interface
4.1.3 View the EEG signal via graph
The software also has the function of reviewing the graph of EEG
signals for every channel, depending on the purpose of the research,
just click on the electrode channel position as shown in Figure 4.3.
Figure 4.3 Viewing EEG signals via graph
4.1.4 Extracting the feature of EEG signals
To extract the feature of EEG signals, author used the HHT
algorithm as presented in the theoretical basis, total number of
analyzied channels is 10, each channel is analyzed into 12 IMF
(intrinsic functions), so we have all 120 IMFs for each processing. The
program extracted 1 signal channel into IMFs, and then from these
IMFs extracted in 5 basic waves as Figure 4.4.
Trang 24
Figure 4.4 An EEG signal channel is transformed into the basic
wave
4.2 Hardware system
The hardware system includes devices such as observation board,
wheelchair model, Emotiv equipment and computers with control
software, the hardware system is shown as Figure 4.5.
Figure 4.5 Sitting posture and hardware devices
4.2.1 Observation board
The observation board is 46x42 cm in size, can adjust the tilting
direction like a laptop screen to suit each participants, on the
observation board there are 5 images of 8x12 cm each placed at even
intervals at a distance of 8 cm, a camera is placed between the image
of the person and the flower to record the direction of the eye as Figure
4.6.
Trang 25
Figure 4.6 Observation board
Camera is mounted on the observation board to detect eye
direction corresponding to 05 image types on the board, camera of
Logitech Co. used with model C615, the specifications as follows:
resolution HD 1080, 30fps, field of view 78o, connect to computer via
USB, the camera is shown in Figure 4.7.
Figure 4.7 Camera Logitech C615
4.2.2 Wheelchair model
Wheelchairs used to simulate the process of commands from
computer, the wheelchair with compact size can run forward,
backward, right turn, left turn and stop in accordance with the
commans form computer, the schematic diagram of the circuit is
shown in Figure 4.9, specifications of wheelchair is shown in Table
4.1.
Trang 26
Figure 4.8 Model of Wheelchair
Table 4.1 Specifications of model wheelchairs
No Dicriptions Specifications
1 size (length – width – high) 35x30x35 cm
2 Motor 200 rpm, 9 VDC
3 Number of motor 02
4 Battery 9VDC 2000 mA
5 Weight 0.8 kg
Wheelchair control system principle schematic includes the
arduino UNO3 processor, driver LM298 for two motors, and bluetooth
module HC-05 to receive control commands from computer.
Figure 4.9 Principle schematic of control board
4.2.3 Emotiv device
Trang 27
An indispensable device in experimental control is the Emotiv
EPOC + device (EPOC+ head), this device record EEG signals and
send it to the computer via bluetooth, EPOC + device is shown in
Figure 4.10.
Figure 4.10 EPOC+ device
EPOC + has the specifications as follows:
Number of channels: 14 channels + 2 reference channels.
Sampling frequency: 128 SPS / 256 SPS.
Data resolution 14 bits, LSB = 0.51 uV.
Connect by bluetooth, 2.4GHz band.
Battery using time 12 hours
The computer connect to 2 devices, EPOC + head and wheelchair
model as Figure 4.11.
Figure 4.11 The computer connect to 2 devices via bluetooth
4.3 Selecting a group to participate in the system evaluation
After finishing the above steps, author conducted the final
experimental step to re-evaluate the entire research results. During
this period, author selected 3 groups of participant as follows:
Group 1: Selecting 20 from 60 students in the first phase of the
best experimental results.
Group 2: Choosing 20 from 60 remaining student.
Group 3: Choosing 20 student never participated in the
experiment test.
Trang 28
In groups 1 and 2, student do not have to go through training
steps, but group 3, student must be trained on the
Các file đính kèm theo tài liệu này:
- neural_network_in_the_wheelchair_control_system_for_severe_d.pdf