Spectrum-Matched Complex Gabor Initialization for Complex-Valued Radar Networks

Complex-valued neural networks are networks whose weights, activations and arithmetic are complex numbers. They are the natural tool for radar, MRI and radio data, where each pixel carries both an amplitude and a phase. These networks still start training from pure noise: every kernel gets a random phase and a random Rayleigh-distributed magnitude. That randomness costs the most where these networks are used most, in radar classification with only a handful of labels. There, a plain complex network trained on 1% of the Flevoland labels reaches only 42.20% overall accuracy. In real-valued vision, the long-standing fix is to start the first layer from Gabor filters, which are localized oriented waves, instead of noise. A Gabor filter is natively a complex object: a cosine and a sine in a shared Gaussian window. Yet it has only ever been poured into real-valued weights.

This note proposes starting the first complex convolution of such a network from a complex Gabor bank. The important twist concerns complex inputs such as single-look radar images. Their frequency content is not mirror-symmetric, so a positive and a negative spatial frequency carry different information. In radar, these frequency bands correspond physically to different viewing angles and sub-bands of the synthetic aperture. The bank therefore has to tile the whole frequency plane, and it should be centered where the data's own spectrum actually has energy. Real-image Gabor recipes do neither. The claim is testable in two sentences. A complex Gabor bank whose centers are fitted to the unlabeled data spectrum, scaled to match standard variance-preserving initialization, should make low-label complex radar networks markedly better than random initialization. The advantage should come specifically from full-plane quadrature structure, which cannot be reproduced by real Gabor filters, half-plane banks, or random kernels with matched spectra.

Core Claim

Each first-layer kernel $k$ is initialized as

$$ g_k(\mathbf{x}) = a\,\exp\!\left(-\frac{\mathbf{x}^\top \Sigma_k^{-1}\mathbf{x}}{2}\right)\exp\!\left(i\,\boldsymbol{\omega}_k\cdot\mathbf{x} + i\phi_k\right), $$

In this formula, $\mathbf{x}$ is the pixel offset within the kernel, $\Sigma_k$ sets the size and shape of the Gaussian window, $\boldsymbol{\omega}_k$ is the center frequency, and $\phi_k$ is a phase offset. The scalar $a$ is chosen so that the average squared kernel magnitude equals the variance prescribed by the Rayleigh scheme of Deep Complex Networks. The real part of $g_k$ becomes the weight $A$ and the imaginary part becomes $B$ in the complex convolution $W = A + iB$.

The center frequencies $\boldsymbol{\omega}_k$ are not a fixed grid. They are cluster centers of the two-dimensional power spectrum of unlabeled training patches, computed over the full plane $[-\pi,\pi]^2$ with no conjugate-symmetry folding. For real images this construction reduces to the familiar half-plane bank, because the spectrum is symmetric. For complex radar images it places filters on asymmetric, physically meaningful bands.

The proposal is therefore not "Gabor init, but complex." It is a statement about which structural ingredients matter for complex data, and each ingredient has an ablation that isolates it.

Research Program

The work builds on three kinds of prior results.

The first kind motivates the proposal. Two studies show that first-layer Gabor initialization helps real-valued CNNs, and that the filters must be fine-tuned rather than frozen: Gabor Filters as Initializers for Convolutional Neural Networks and Enhancing CNNs Performance on Object Recognition Tasks with Gabor Initialization. Robust CNN-based speech recognition with Gabor filter kernels shows the same in spectro-temporal audio. Gabor Convolutional Networks goes further and hardwires fixed Gabor modulation into every layer. All of these use only real, even-symmetric parts, and none touches complex arithmetic. On the complex side, C²N² and KINN both show that complex radar networks need structural priors at low label counts. They add them as extra machinery, a parallel contourlet branch and an unfolded scattering-center solver respectively, and leave the backbone's own kernels randomly initialized. The question here is how much of that prior can live in the initialization alone, at no inference cost.

The second kind constrains the design. Real-Valued Backpropagation is Unsuitable for Complex-Valued Neural Networks proves something important about the common activation that applies ReLU separately to the real and imaginary parts (CReLU). In the infinite-width limit, a complex network with this activation trains like a real network. Only activations that treat the complex number as a whole, such as zReLU, retain distinct complex dynamics. If that holds at practical widths, a quadrature initialization could be washed out during training under CReLU. We therefore cross the initialization with the activation choice (CReLU, modReLU, zReLU). We also measure, after training, whether first-layer kernels remain approximate quadrature pairs, meaning that the real and imaginary parts stay close to Hilbert-transform partners. The kernel-size finding in the real-valued study (Gabor filters need more than $3\times3$ support) constrains us to a $7\times7$ first layer while keeping $3\times3$ kernels deeper. This contains the fourfold compute cost of complex convolution noted in Deep Complex Networks.

The third kind supplies the method. Complex convolution, complex batch normalization and the variance scale all come from Deep Complex Networks. The symmetry fact from Fast 2-D Complex Gabor Filter with Kernel Decomposition states that orientations $\theta$ and $\pi-\theta$ give conjugate responses on real inputs. That fact is exactly what breaks on complex inputs, and it is the reason full-plane tiling is needed.

A falsifiable prediction follows from the physics. Single-look complex images such as S1SLC_CVDL and complex MSTAR chips have spatially coherent phase, so their spectra are asymmetric. Flevoland-style PolSAR inputs are usually entries of a coherency matrix, and these are local covariances whose spatial phase structure is weak. The full-plane, spectrum-fitted bank should therefore gain more over a half-plane bank on the single-look datasets than on the coherency-matrix datasets. If the gain is equal on both, the physical story is wrong, and the benefit is generic regularization.

Baselines

The do-nothing reference is the standard complex CNN with uniform-phase, Rayleigh-magnitude initialization from Deep Complex Networks. Its low-data weakness is documented: the baseline complex CNN in C²N² reaches 42.20% overall accuracy on Flevoland with 1% labels, against 55.03% for an SVM. We will re-run it in our own pipeline so that splits and augmentation match. The proposal must clearly exceed it across seeds.

The strongest existing method on the headline benchmark is C²N². It reports 71.60% overall accuracy on Flevoland with 1% labels, using a spatial complex CNN plus a nonsubsampled-contourlet branch and a statistical fusion module. We will run it with the authors' published configuration on identical splits. There are two ways to beat it. The first is to match or exceed it with a single-stream network, which would show that the transform prior can move into the initialization. The second is to add our initialization to C²N²'s own spatial branch and improve on 71.60%, which would show the two priors are complementary. For MSTAR at 10% labels, KINN reports an average 11% accuracy improvement over lightweight competitors, but the file gives no absolute accuracy, so we will reproduce KINN-CNN (0.7M parameters) and compare at matched parameter count. For S1SLC_CVDL, Shift-Equivariant Complex-Valued Convolutional Neural Networks reports its polyphase-sampling model as best in overall accuracy and F1, but no reported result found on Lacuna for the numeric accuracy on S1SLC_CVDL. We will train that architecture ourselves, with and without our initialization.

Four ablations isolate the key ingredient, and all share one backbone, optimizer, epoch budget, and an identical hyperparameter search of 20 trials per arm. The first sets the imaginary part to zero at initialization, testing whether quadrature matters. The second uses a half-plane bank, the standard real-image recipe, testing whether full-plane tiling matters. The third uses a fixed orientation-and-scale grid instead of spectrum-fitted centers, testing whether data adaptation matters. The fourth uses random complex kernels whose power spectra match the Gabor bank's, which tests whether spatial localization and coherent phase matter beyond spectral energy.

Why not just…?

Why not just use a real network with two input channels and even/odd real Gabor filters? This is the most serious alternative. Its first layer can represent the same quadrature pair, and nothing forces the complex product structure afterward. The file's evidence against it is indirect. Real networks underperform complex ones on S1SLC_CVDL, PolSF and ALOS-2 in Shift-Equivariant Complex-Valued Convolutional Neural Networks, and on MusicNet a complex network scores 72.9% average precision against a 69.6% real-valued prior best in Deep Complex Networks. None of those real networks, however, had Gabor initialization. We run this arm at matched parameters and FLOPs. If it ties our method under CReLU, that would fit Tan et al.'s collapse result and would narrow our claim to holistic activations only.

Why not just freeze a complex Gabor bank as a fixed front end? In real vision, freezing hurt final accuracy in both Gabor initialization studies. At 0.1% labels, though, a frozen bank may generalize better than a fine-tuned one, much as C²N²'s fixed contourlet branch helps. We include frozen and fine-tuned arms and report where the crossover falls as the label fraction grows.

Why not just initialize from principal components of unlabeled complex patches? This is a genuine threat, because it is data-adapted and has no free parameters. We include complex principal-component initialization at matched variance. If it matches our bank, the honest conclusion is that data-adapted spectral structure is what matters, not the Gabor shape. The spectrum-matched random ablation would then help tell whether localization adds anything.

Proposed Experiments

We will use Flevoland, Oberpfaffenhofen and Xi'an at 0.1% and 1% labels, S1SLC_CVDL, and complex MSTAR at 10% labels. We report overall accuracy and F1 over ten seeds, epochs needed to reach 95% of final validation accuracy, and, for S1SLC_CVDL, circular shift consistency.

Three results are decisive. First, the full method must beat Rayleigh initialization by a margin larger than the variation across seeds at 1% labels, and it must approach or beat C²N²'s 71.60% on Flevoland without the extra branch. Second, the gap between the full-plane and half-plane banks must be larger on the single-look datasets than on the coherency-matrix datasets, as the physical prediction requires. Third, the advantage over the two-channel real Gabor network should hold under zReLU. Under CReLU it should either hold or shrink. After training, the retained quadrature structure of the first-layer kernels will show whether the initial prior survived or was erased, as Tan et al.'s theory predicts for CReLU at large width.