Polarization Parameter Network (PPN)

• PPN is a CNN module that constructs task-specific polarization-parameter maps from four-angle polarimetric images.
• It employs stacked 1×1 convolutions with batch normalization and ReLU to fuse features without spatial mixing.
• Empirical evaluations show significant improvements in object detection mAP, highlighting its integration benefits with networks like Faster R-CNN.

The Polarization Parameter Network (PPN), more precisely termed the Polarization-Parameter-Constructing Network (PPCN), is a convolutional neural network (CNN) architectural module designed to enable end-to-end, task-driven extraction and fusion of features from polarimetric imaging data. The architecture addresses the absence of trainable, pixel-wise operations that can synthesize and exploit polarization-derived information in downstream vision tasks. Positioned as a front-end module between sensor-level polarimetric images and broader computer vision networks (e.g., Faster R-CNN for object detection), the PPCN learns to construct optimal, task-specific polarization-parameter images, generalizing and often surpassing classical representations such as Stokes parameters, degree of linear polarization (DoLP), and angle of linear polarization (AoLP) (Wang et al., 2020).

1. Architectural Definition and Placement

The PPCN is deployed as a pixel-wise, channel-fusion sub-network that processes the four raw, spatially aligned polarimetric intensity images—typically captured at 0°, 45°, 90°, and 135° (labeled $I_0, I_{45}, I_{90}, I_{135}$). It operates strictly along the channel dimension, stacking multiple 1×1 convolutional layers interleaved with batch normalization and ReLU nonlinearities, to produce $M$ output feature maps. The standard structural specification can be denoted as “4–$C_1$–$C_2$–…–$C_n$–$M$,” where the input is four channels, each subsequent fusion unit applies a 1×1 (no spatial mixing) convolution with output channels $C_i$, and $M$ is the final number of polarization-parameter maps provided to the subsequent vision task network.

Example Architecture

Layer	Input Channels	Output Channels	Operation
Input	4	$C_1$	1×1 Conv + BN + ReLU
Fusion 1	$C_1$	$C_2$	1×1 Conv + BN + ReLU
...	...	...	...
Output	$C_n$	$M$	1×1 Conv

This configuration enables the PPCN to learn, for each pixel location, parameterizations that are optimally fused for the subsequent vision objective.

2. Mathematical Framework

For each pixel $(x,y)$, the PPCN learns $M$ task-driven, differentiable functions of the four raw channel intensities:

$$P_j(x,y) = f_j\left(I_0(x,y), I_{45}(x,y), I_{90}(x,y), I_{135}(x,y)\right), \quad j=1\ldots M,$$

implemented via stacked 1×1 convolutions, batch normalization, and ReLU activations.

The 1×1 convolutional fusion at each step is

$Z_j(x, y) = \sum_k w_{jk}\,X_k(x, y) + b_j,$

where $X_k$ are the input channels and $w_{jk}, b_j$ are learned parameters, followed by

$$Y_j(x, y) = \mathrm{ReLU}(\mathrm{BatchNorm}(Z_j(x, y))).$$

By contrast, classical Stokes-based parameter construction is fixed:

• $S_0 = I_0 + I_{90}$
• $S_1 = I_0 - I_{90}$
• $S_2 = I_{45} - I_{135}$
• $\mathrm{DoLP} = \sqrt{S_1^2 + S_2^2}/S_0$
• $\mathrm{AoLP} = \frac{1}{2}\arctan2(S_2, S_1)$

The PPCN subsumes these by learning parameterizations without hard-coded forms.

3. Training Protocols and Loss Strategies

Two training regimes are employed:

• Stokes-fitting pre-training: The PPCN is supervised to approximate classical parameter maps via an $\ell_2$ fitting loss:

$$L_{\rm fit} = \sum_{n,x,y} \bigl\|\hat S_0 - S_0^{(n)}\bigr\|^2 + \|\widehat{\mathrm{DoLP}} - \mathrm{DoLP}^{(n)}\|^2 + \|\widehat{\mathrm{AoLP}} - \mathrm{AoLP}^{(n)}\|^2,$$

where hats denote network outputs and $(n)$ indexes normalized ground truth.

• End-to-end, task-driven training: The PPCN is trained jointly with a downstream vision network (e.g., Faster R-CNN), such that only the final detection loss propagates through the entire architecture:

$$\min_{\theta_{\rm PPCN}, \theta_{\rm RPN}} L_{\rm det}(\mathrm{PPCN}(I; \theta_{\rm PPCN}), \theta_{\rm RPN}),$$

without explicit regularization or parameter constraints on the PPCN.

No additional regularization, orthogonality, or sparsity penalties were imposed during object-detection experiments.

4. Integration with Vision Task Networks

To incorporate the PPCN into a standard object detection pipeline (e.g., Faster R-CNN with ResNet-50 backbone):

• The first convolutional layer’s input channels are replaced so as to accept $M$ polarization-parametric maps produced by the PPCN, instead of three RGB channels.
• All other layers of the backbone and detection heads remain as original.
• During a forward pass, input is processed as: raw polarimetric images $\rightarrow$ PPCN $\rightarrow$ $M$-channel feature maps $\rightarrow$ vision task network. Gradients derived from the task loss are backpropagated through the entire PPCN.

This structure is also directly extensible to other CNN-based tasks such as semantic segmentation or multimodal image fusion.

5. Empirical Evaluation and Analysis

Experiments were performed on a dataset consisting of 3,000 sets of polarimetric (four-angle) and RGB images annotated for cars and pedestrians, with conventional splits into train/validation/test.

Main quantitative findings:

• PPCN structural ablation: Increasing channel widths in the PPCN results in reduced pixel-wise Stokes fitting loss (example: “4–8–16–8–3” achieves $1.16 \times 10^{-2}$ fit-loss; “4–128–96–48–32–3” achieves $4.21 \times 10^{-3}$), with larger models incurring higher memory usage.
• Number of output maps ($M$): For detection of both cars and pedestrians, $M=9$ yields the highest mean Average Precision (mAP) at $82.7\%$, with larger values offering no improvement. For single-class (car) detection, $M=5$ suffices ($\sim$91.5% AP).
• Quantitative gains (Intersection over Union threshold 0.5):

Method	mAP (%)	AP_car (%)	AP_ped (%)
Baseline (raw pol, R-50)	72.6	83.7	61.4
PPCN (4–48–96–32–16–9)+R-50	82.7	92.7	72.7

This represents improvements of +10.1 mAP, +9.0 AP_car, +11.3 AP_ped. Increasing the backbone depth to ResNet-101, without a PPCN module, does not improve performance (mAP = 71.2%), indicating that PPCN contributions are not trivially recoverable by scaling conventional network depth.

Qualitative findings:

Learned polarization-parameter maps display diverse, target-focused activations: cars remain strongly emphasized across most maps, while background classes (roads, vegetation, buildings) are variably suppressed or isolated—demonstrating the network’s capacity to extract non-redundant, class-relevant polarization cues.

6. Generalization, Limitations, and Future Directions

The PPCN’s learned outputs often correlate, but are not strictly redundant with, canonical Stokes, DoLP, or AoLP maps. The network may discover richer, task-specialized parametric mixtures, exploiting the full potential of raw polarization information. This suggests applicability to any CNN-based vision task ingesting polarimetric data, beyond object detection, such as classification, segmentation, or multimodal imaging.

The PPCN framework’s simplicity—consisting of stacked 1×1 convolutions with batch normalization and ReLU—facilitates easy integration, reproducibility, and extensibility for future polarimetric vision research. Code repositories are available to support these efforts (Wang et al., 2020).

Editor’s term: While the source refers specifically to the Polarization-Parameter-Constructing Network (PPCN), “Polarization Parameter Network” is used here as a practical shorthand to align with the topic designation. All technical details trace to the PPCN architecture and results (Wang et al., 2020).