290 likes | 416 Views
IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 21, NO. 1, JANUARY 2012. Xiulian Peng , Jizheng Xu , Senior Member, IEEE, You Zhou, Member, IEEE , Feng Wu, Senior Member, IEEE. Highly Parallel Line-Based Image Coding for Many Cores. Guan-Yu liu. Introduction Proposed Method
E N D
IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 21, NO. 1, JANUARY 2012 • XiulianPeng, JizhengXu, Senior Member, IEEE, • You Zhou, Member, IEEE, • FengWu, Senior Member, IEEE Highly Parallel Line-Based Image Coding for Many Cores • Guan-Yu liu
Introduction • Proposed Method • Hierarchical Coding • Adaptive Line Prediction • Variable-Length Transform • Parallel Entropy Coding • Extension to Lossless Coding Outline #1
Parallelism Analysis • Experimental Result • Encode Parallelism • Decode Parallelism • Coding Performance • Conclusion Outline #2
Pipeline approaches [4] Slice partitioning approaches MB-reordering approaches Introduction #1
line-by-line image coding (LBLC) • Divided into small fixed-length segments • Different units • Three different line prediction methods • Parallel context-adaptive binary arithmetic coding Introduction #2
The current block needs the reconstructed pixels of its upper, left, upper-left, and upper-right blocks for prediction. Proposed Method #1
Proposed Method #2 Equal M N lines N
Hierarchical Coding • prediction method unit (PMU) • prediction parameter unit (PPU) • transform unit (TU) Proposed Method.A #1
Hierarchical Coding Proposed Method.A #2 Then how to make a decision?
Block-based • RDO • There are three implicit prediction methods • two local prediction methods • predefined filters (PDF) • direction-aided local training (DA-LT) • one nonlocal prediction • line–template matching (LTM) Proposed Method.B #1
Training window Proposed Method.B #2
predefined filters (PDF) Proposed Method.B #3
Minimize the prediction MSE for all pixels inside W • 123 Proposed Method.B #5
direction-aided local training (DA-LT) Proposed Method.B #6
Train a Wiener filter • 123 • This problem can simply be solved by the least-square approach. Proposed Method.B #7
line template matching Proposed Method.B #8
Variable-Length Transform • After the prediction, 1-D transforms are optionally performed on the predicted residues to further decorrelate them. Proposed Method.C #1
Parallel Entropy Coding • Context-based arithmetic coding method is used, which is similar to the CABAC. Proposed Method.D #1
Extension to Lossless Coding Proposed Method.E #1
Extension to Lossless Coding Proposed Method.E #2
The segments on the same line are assigned to different cores. All kth segments are assigned to the same core and coded sequentially. Proposed Method
Encode Experimental Result.A #1
Encode Experimental Result.A #2
Decode Experimental Result.B #1
Coding Performance Experimental Result.C #1
Coding Performance Experimental Result.C #2
It reduces the dependence among different cores and the independent coding of each segment within a line makes it easy to achieve a high parallelism. • The computational complexity is high at the decoder due to the derivation of prediction parameters. Conclusion