1. A Block Based MAP Segmentation for Image Compression
By Waseem Khatri

2. Motivation Need for a Robust Segmentation method
Segmentation should yield larger homogeneous regions which are suitable for image compression
Object based image compression for better compression efficiency

3. Objective
Segmentation with Spatial Homogeneity – as it yields accurate segmentation results
Encoding and Decoding segmented objects independently – for better motion compensation

4. Drawback of Object Based Segmentation
Irregular shape of objects needs a significant number of bits for encoding
Many computation for object segmentation
Not suitable for real-time video coding applications
Solution ?
Block – based object segmentation

5. MAP using GRF To identify edges, monotone and textured blocks
Clique potentials for the GRF – guarantee the edge connectivity and the smoothness of the homogeneous block

6. Structure of Proposed Segmentation
Object based image compression encoder
Partition
Definition
Partition
Coding
Texture
Coding
Block
Classification/
Segmentation
Numbering to
Connected
Homogeneous
Blocks
Assignment of
Uncertainty
Blocks

7. Block
Classification/
Segmentation
Numbering to
Connected
Homogeneous
Blocks
Assignment of
Uncertainty
Blocks
Block Classification – Determines whether the given image block belongs to a homogeneous block( textured or monotone) or an edge block
Numbering Homogenous blocks - Region Number is assigned to connected homogenous blocks
Uncertainty blocks – Edge blocks and Unnumbered isolated homogeneous blocks
Region Growing Method – Assigns uncertainty blocks to one of its neighboring homogeneous regions

8. MAP Criterion
Let X denote a 2D random field representing a set of all possible characteristics
where
represents the set of all block indexes.
means that the grey level distribution in the block at t is close to an edge block.
implies a monotone block
indicates a textured block

9. MAP Criterion
Let Y denote a random field for the set of all observed gray levels in
The optimal block label realization that maximizes the a posteriori probability is given by
Where
i.e. the gray level configuration of all blocks are independent
Using GRF

10. Cliques

11. Numbering of Connected Homogeneous Blocks
Implemented using the following paper,
L.Vincent, “Watershed in digital spaces: an efficient algorithm based on immersion simulations,” IEEE Trans. Pattern Anal. Machine Intell., vol. 13, pp , June 1991.

12. Assignment of Uncertainty Blocks
Uncertainty blocks are assigned to one of their neighboring numbered homogeneous regions
Process :
Assign an uncertainty block to one of its neighboring regions if it yields the smallest difference between the block mean and that of the neighboring region to be assigned
Repeat for all uncertainty blocks until no block is left – i.e. we obtain a final block based image segmentation.

13. Expected Results

14. Expected Results

15. Conclusion More accurate segmentation by reducing the block size
Proposed method can separate the contour blocks from the homogeneous region blocks
Since the contour of an object is represented in terms of blocks rather than pixels, we can change the balance between reconstructed video quality and the coding cost by adjusting the block size.
Textured region can be identified by considering the statistical distribution of the gray levels in the block into the segmentation criterion.

16. Thank-you !