Aline Martin alinemartinwisc

Aline Martin alinemartin@wisc

edu ECE738 Project – Spring 2005

Detecting blurring artifacts in jpeg2000 compressed images using Classification

• Problem statement • Previous work • Proposed approach:

classification • Results • Conclusion and future work

outline

2

Problem Statement

Jpeg2000 creates blurring artifacts

Blurred patches in highly textured regions

3

Previous work

Rajas A. Sambhare: “Detecting Artifacts and Textures in Wavelet Coded Images” ECE 783 Project – Spring 2003

Targets

Source

Original image

Texture detection

Blurring artifacts detection

4

Previous work

Algorithm:

1 – Detect Textured Regions 2 – Segmentation: k-mean algorithm 3 – Identification of Textured Segments 4 – Identification of segments adjacent to textured Segments For each Textured Segment For each adjacent segment if |mean Source – mean adjacent segment| < 0.2 then target segment end end

5

Previous work

Proposed Approach: Use Classification to automatically classify smooth regions as blurring artifacts or smooth regions

Drawbacks: Heuristic Threshold Does not perform well on other images

6

Proposed Approach

Class0: Source is a texture and Target is not a blurring artifact Class1: Source is a texture and Target is a blurring artifact

Algorithm:

1 – Detect Textured Regions 2 – Segmentation: k-mean algorithm 3 – Identification of Textured Segments 4 – Identification of segments adjacent to textured Segments For each Textured Segment For each adjacent segment compute feature vector classify it as Class0 or Class1 end end

7

Proposed Approach

Original Image

Identification of Textured areas

Segmentation

Segments edges

Textured Segments

8

Proposed Approach

Classification

1- Features extraction -> 6-d Feature vector - mean Source Segment - variance Source Segment - mean Target Segment - variance Target Segment - | mean Source Segment - mean Target Segment | - | mean Source Segment - mean Target Segment |

9

Proposed Approach

Classification

2- Training the classifier

Compute m0, m1 : 6 by 1 So, S1 : 6 by 6

10

Proposed Approach

Classification

3- Classifier

x: 6-d feature vector computed from a possible Target

P0(x) ~ N(m0,S0) P1(x) ~ N(m1,S1)

If P1(x) > P0(x) then Class 1 If P0(x) > P1(x) then Class 0

11

Results

12

Results

13

Results

Limitations due to segmentation

Refinement?

14

Conclusion and Future work

An algorithm to detect blurring artifacts in jpeg2000 compressed images was developed

Need to improve segmentation: Refinement?

Need a better segmentation algorithm for black and white images

Need to increase the images database

15