Car Classification with Neural Networks

1
Car Classification with Neural Networks
Artificial Neural Systems
4/27/99 presentation

• Koichi Sato Sangho Park

2
Projects Goal

• input image containing cars
• output classified car model
• detected car

Ford Aerostar !
3
Process flow
Edge Detector
Image
Preprocessing
NN
Output

• Emphasizing H/V Lines
• X- y- projection
• Smoothing

x-/ y- projection
raw image
edge image
4
Major Characteristics

• Car has strong edge
• Much of Car Model features are in front face
• Many horizontal / Vertical edges

5
Approach

• Edge Detection
• Cropping
• X- / Y- Projection
• (reduction of feature dimensions)
• (orthogonal projection)

6
Details

• to solve this problem
• picture ( 320x240 grayscale image )
• edge detection ( Canny method )
• input data x / y- projection of edge
• neural network
• MLP supervised
• SOM unsupervised
• analysis of characteristics in input data
• experiments with different algorithms

7
Edge Detection

• reasons
• color-invariant
• brightness-invariant
• method
• Canny edge
• Gaussian smoothing
• gradient filtering
• zero Crossing

8
PreProcessing (1)

• H / V Line emphasis

Delete Horizontal Line Delete Vertical
Line
9
PreProcessing (2)

• X Y Projection

Smoothing
10
Neural Network Structure

• MLP spec. (supervised learning)
• 1 hidden layer
• 5 hidden units
• 2 output
• 1st output Ford Aerostar
• 2nd output Toyota Camry
• Input dimension depends on algorithms.
• ( we took several algorithms )

11

• SOM spec. (unsupervised learning)
• 5 x 5 grid weight nodes.
• output topology of interest
• Ford Aerostar / Toyota Camry / Other models
• Input dimension depends on algorithms.
• ( we took several algorithms )
• (1,000 iterations for each algorithm)

12
Experiment

• data size (unit images)

13
Algorithms for experiments

• Notations (
  y yes, n no )
• smoothing reduction of high frequency in
  projection
• -H delete Horizontal edge from x-projection
• -V delete Vertical edge from y-projection
• subsampling decimated image
• x-projection use only x-projection as input
• pseudo 2D individual projections of 16-tiles in
  image

14
MLP results
Error rate
(unit percent)

• Each error value was computed based on 30 runs
• with103 images,
  respectively.
• Each algorithm has used the identical data set.

15
What the Network sees ?

• (weight matching with image in algorithm-1)

Node 4
Node 1
Node 2
Node 5
Node 3

• White lines show strong
• weight values.

16

• (weight matching with image in algorithm-4)

Node 4
Node 1
Node 2
Node 5
Node 3

• White lines show strong
• weight values.

17
SOM results (1,000 iterations each)
Algorithm 1
Aerostar well clustered Camry more
diverse others large variations in
weights
(Refer the next slide)
18
Comparison of algorithms
Algorithm 4
Algorithm 2
Algorithm 3
Algorithm 6
Algorithm 5
Algorithm 7

• Each algorithm has used the identical data set.

19
Consideration on performance

• Effects of projection smoothing
• reduction of high frequency in input
• tolerance on edge locations

before smoothing
after smoothing
20

• Effects of viewing direction
• vertical vs. horizontal variation in input
• edge translation vs. edge orientation

Vertical variation
Horizontal variation
21
Application of NN to detection

• Detection of Aerostar in a scene using NN

NN output
Max(output)
Max(output)
Mapping
22

• Network Output and max(output)

23
Practical Consideration

• Hardware requirements
• memory (MLP 18K byte)
• speed of MLP (2.3K multiplications and 2.3K
  additions per image)
• preprocessing device for realtime

24
Problems to be solved

• Size dependency
• image normalization problem
• Distortion by viewpoint
• edge translation by vertical variation of view
  (minor problem)
• edge rotation by horizontal variation of view
  (major problem)
• Noise reduction
• smoothing / other methods

25
Future works

• Non-normalized pictures
• ( unfixed size picture )
• Multiple viewpoints
• (front view, profile view, semi-profile view)
• More car-models
• (car type passenger car, truck, bus etc.)
• (car model Camry, Aerostar, Civic etc.)
• Sensitivity improvement in detection

26
Conclusion

• Merits of neural network
• robust to the image distortion
• versatile diverse approaches
• What we learned from this project
• Properly reduced representation of input features
  works well.
• Generalization of the system needs more factors
  to be considered.
• Characteristics of input data needs
  investigation.