LIPS-PC
An end-to-end learning framework for single-pixel imaging classification that parameterizes the measurement operator as a residual correction to a 2D-DCT basis and trains it jointly with a compact classifier, enabling efficient classification without image reconstruction.
Timeline (last 90 days)
LIPS-PC (Learning Illumination Patterns for Single-Pixel Classification) is a single-pixel imaging method that directly learns illumination patterns for classification, without the need to first reconstruct the image and then classify it.
LIPS-PC parameterizes the measurement operator as a residual correction to the 2D-DCT basis matrix, i.e., D = D_base + Δ, where Δ is trained end-to-end jointly with a compact classifier via a cross-entropy loss.
LIPS-PC was evaluated on the MNIST, Fashion-MNIST and RIAWELC datasets under measurement budgets M ∈ {8,16,32,64,128}.
At M=8, the operator learned by LIPS-PC outperforms the best fixed transform (such as 2D-DCT) by 9-15 percentage points on all three datasets.
The accuracy LIPS-PC achieves with only 8 measurements requires about 16-32 measurements for the best fixed transform to reach, i.e., a 2x to 4x measurement saving at comparable accuracy.
As the measurement budget M increases, the advantage of LIPS-PC over fixed transforms gradually shrinks, and at high budgets on RIAWELC a slight reversal even occurs.
LIPS-PC demonstrated its practical value in industrial X-ray weld defect screening, significantly reducing the number of measurements by directly targeting classification, thereby cutting acquisition time and radiation dose.
The LIPS-PC method has limitations: at high budgets its advantage weakens or even reverses, possibly because the learned patterns overfit to low-measurement conditions; and the experiments involve only linear measurements and simple classifiers.
In the LIPS-PC parameterized policy, the DCT basis provides an information-rich initialization starting point, and the residual term Δ only needs to learn small task-relevant adjustments, thereby accelerating convergence and improving performance at low budgets; learning from a random matrix or zero matrix would significantly increase optimization difficulty.
All Facts (9)
LIPS-PC (Learning Illumination Patterns for Single-Pixel Classification) is a single-pixel imaging method that directly learns illumination patterns for classification, without the need to first reconstruct the image and then classify it.
50%UnverifiedLIPS-PC parameterizes the measurement operator as a residual correction to the 2D-DCT basis matrix, i.e., D = D_base + Δ, where Δ is trained end-to-end jointly with a compact classifier via a cross-entropy loss.
50%UnverifiedLIPS-PC was evaluated on the MNIST, Fashion-MNIST and RIAWELC datasets under measurement budgets M ∈ {8,16,32,64,128}.
50%UnverifiedAt M=8, the operator learned by LIPS-PC outperforms the best fixed transform (such as 2D-DCT) by 9-15 percentage points on all three datasets.
50%UnverifiedThe accuracy LIPS-PC achieves with only 8 measurements requires about 16-32 measurements for the best fixed transform to reach, i.e., a 2x to 4x measurement saving at comparable accuracy.
50%UnverifiedAs the measurement budget M increases, the advantage of LIPS-PC over fixed transforms gradually shrinks, and at high budgets on RIAWELC a slight reversal even occurs.
50%UnverifiedLIPS-PC demonstrated its practical value in industrial X-ray weld defect screening, significantly reducing the number of measurements by directly targeting classification, thereby cutting acquisition time and radiation dose.
50%UnverifiedThe LIPS-PC method has limitations: at high budgets its advantage weakens or even reverses, possibly because the learned patterns overfit to low-measurement conditions; and the experiments involve only linear measurements and simple classifiers.
50%UnverifiedIn the LIPS-PC parameterized policy, the DCT basis provides an information-rich initialization starting point, and the residual term Δ only needs to learn small task-relevant adjustments, thereby accelerating convergence and improving performance at low budgets; learning from a random matrix or zero matrix would significantly increase optimization difficulty.
50%