Dual-Branch Inverse Kernel Prediction (DIKP)

• DIKP is a frequency-domain approach that leverages separate amplitude and phase branches to predict a compact bank of spatial inverse blur kernels.
• The method fuses amplitude-guided phase attention via complex exponentiation and inverse FFT, bridging physics-based blur modeling with learned restoration.
• By providing explicit inverse kernel predictions for downstream adaptive deconvolution and recurrent fusion, DIKP enhances kernel estimation accuracy in defocus deblurring.

Searching arXiv for the specified DIKP-related papers to ground the article in the cited literature. Dual-Branch Inverse Kernel Prediction (DIKP) is a component of the Frequency-Driven Inverse Kernel Prediction (FDIKP) framework for single image defocus deblurring. In that framework, DIKP is the first part of the Frequency Inverse Kernel Predictor (FIKP), and its function is to estimate a small set of spatial inverse blur kernels from a defocused input image by exploiting frequency-domain cues rather than relying solely on spatial features. The method is designed around the observation that defocus blur is a frequency-domain filter, and it converts this physical characterization into a learned prediction of inverse filters that can be consumed by downstream adaptive deconvolution and recurrent fusion modules (Zhang et al., 18 Aug 2025).

1. Placement within the FDIKP framework

Within each stage of FDIKP, the architecture is organized into two sub-modules: FIKP and DSRM. FIKP itself contains DIKP and PAC, while DSRM denotes the Dual-Domain Scale Recurrent Module. DIKP therefore precedes both the position-adaptive deconvolution step and the subsequent cross-scale fusion stage.

Operationally, DIKP receives a defocused image and explicitly estimates a small bank of spatial inverse blur kernels. The detailed description specifies that these are, for example, five $5 \times 5$ filters. These predicted inverse kernels $k^*$ are then passed to Position Adaptive Convolution (PAC), which uses them to generate per-kernel deconvolution feature maps. DSRM later fuses those feature maps using learned coefficient maps and refines the result in a coarse-to-fine, scale-recurrent manner (Zhang et al., 18 Aug 2025).

This placement clarifies the role of DIKP in the larger pipeline. It is not itself the complete deblurring model; rather, it is the kernel-prediction mechanism that supplies a physics-guided intermediate representation for downstream reconstruction. A plausible implication is that the method is intended to bridge explicit blur modeling and learned restoration, rather than to replace either one completely.

2. Dual-branch construction in the frequency domain

DIKP uses two branches operating on frequency-domain representations of the blurred image $X$: an amplitude branch and a phase branch. The amplitude branch takes as input the amplitude spectrum

$A = |\mathcal{F}(X)|,$

where $\mathcal{F}(\cdot)$ is the 2D FFT. Its predictor consists of two stacked units, each containing $3 \times 3$ Conv $\rightarrow$ ReLU $\rightarrow$ AdaptiveAvgPool to collapse the spatial size to $1 \times 1$, followed by a final $1 \times 1$ convolution expanding to $k^*$0 channels and a Softmax along the $k^*$1 dimension. The branch outputs a set of $k^*$2 amplitude maps in frequency space, denoted $k^*$3. The accompanying rationale is that the amplitude spectrum of a blur encodes its global filter shape, such as disk-shaped attenuation, making structural estimation more reliable (Zhang et al., 18 Aug 2025).

The phase branch takes as input the phase spectrum

$k^*$4

Its base predictor uses the same architecture as the amplitude branch and produces an initial tensor $k^*$5. DIKP then applies amplitude-guided attention. Specifically, it computes a residual from $k^*$6, for example $k^*$7 minus its spatial average, passes that residual through a small Conv $k^*$8 GAP $k^*$9 Softmax stack to obtain an attention vector $X$0, and broadcasts $X$1 over the $X$2 grid so that

$X$3

The resulting $X$4 is interpreted as the phase values for each predicted kernel in frequency space. The stated rationale is that phase carries fine-scale cues such as edges and local shifts, but is noisy without amplitude guidance; the attention mechanism therefore focuses the phase predictor on kernel-relevant directions (Zhang et al., 18 Aug 2025).

Taken together, the two-branch design separates global structural estimation from detail-sensitive correction. This suggests that DIKP treats amplitude and phase as complementary but not equally stable signals for inverse-kernel modeling.

3. Fusion into spatial inverse kernels

After the two branches have produced $X$5 and $X$6, DIKP combines them into a complex frequency-domain representation for each predicted kernel. Let $X$7 denote the amplitude value at frequency bin $X$8 for the $X$9-th kernel, and let $A = |\mathcal{F}(X)|,$0 denote the corresponding phase value. The fusion is defined as

$A = |\mathcal{F}(X)|,$1

An inverse FFT is then applied kernel-wise: $A = |\mathcal{F}(X)|,$2 and the collection $A = |\mathcal{F}(X)|,$3 is stacked into a tensor

$A = |\mathcal{F}(X)|,$4

The description notes that the Softmax in the amplitude branch guarantees

$A = |\mathcal{F}(X)|,$5

which is described as a typical normalization for blur filters (Zhang et al., 18 Aug 2025).

This fusion mechanism is central to the identity of DIKP. Rather than predicting spatial kernels directly, the module predicts amplitude and phase components in frequency space and reconstructs spatial-domain inverse kernels by complex exponentiation followed by inverse FFT. A plausible implication is that this design constrains kernel prediction through a frequency-aware parameterization that is more closely aligned with the physics of defocus blur.

4. Downstream interaction with PAC and DSRM

DIKP does not directly output the final all-in-focus image. Its outputs are first consumed by Position Adaptive Convolution (PAC). PAC computes a per-pixel dilation map $A = |\mathcal{F}(X)|,$6 from the blurred image $A = |\mathcal{F}(X)|,$7 using a small Conv + Sigmoid head. For each pixel location $A = |\mathcal{F}(X)|,$8 and kernel index $A = |\mathcal{F}(X)|,$9, PAC forms a deconvolution feature map

$\mathcal{F}(\cdot)$0

where $\mathcal{F}(\cdot)$1 are the offsets of the $\mathcal{F}(\cdot)$2 sampling grid. Each $\mathcal{F}(\cdot)$3 is then refined with one additional $\mathcal{F}(\cdot)$4 convolution. The resulting FIKP output is

$\mathcal{F}(\cdot)$5

with $\mathcal{F}(\cdot)$6 channel per kernel (Zhang et al., 18 Aug 2025).

These feature maps are then passed to the Dual-Domain Scale Recurrent Module (DSRM). DSRM encodes them via ResBlocks, spatial self-attention, and FFT blocks, predicts coefficient maps $\mathcal{F}(\cdot)$7, and forms the fused feature

$\mathcal{F}(\cdot)$8

A final $\mathcal{F}(\cdot)$9 convolution on $3 \times 3$0 produces the stage output $3 \times 3$1, which is upsampled and passed to the next finer scale (Zhang et al., 18 Aug 2025).

The significance of this data flow is that DIKP supplies a small, explicit bank of inverse filters, while PAC and DSRM provide the spatial adaptivity and multiscale aggregation needed to apply them effectively. This suggests that the predicted kernels are intentionally limited in number and are expected to be combined downstream rather than used as a complete per-pixel blur model on their own.

5. Mathematical specification and implementation settings

The inverse-kernel prediction path is stated explicitly as

$3 \times 3$2

$3 \times 3$3

$3 \times 3$4

$3 \times 3$5

$3 \times 3$6

and for each $3 \times 3$7,

$3 \times 3$8

At each scale $3 \times 3$9, the network produces $\rightarrow$0, and the overall loss is

$\rightarrow$1

with

$\rightarrow$2

where $\rightarrow$3 and $\rightarrow$4 (Zhang et al., 18 Aug 2025).

The implementation details given for DIKP and its immediate surroundings are as follows:

Setting	Value
Number of predicted kernels $\rightarrow$5	5
Kernel size $\rightarrow$6	$\rightarrow$7
Predictor architecture	$\rightarrow$8 Conv, ReLU, AdaptiveAvgPool $\rightarrow$9, repeated twice; final $\rightarrow$0 Conv; Softmax over $\rightarrow$1
Attention head	$\rightarrow$2 Conv on amplitude residual, Global AvgPool, Softmax over $\rightarrow$3
PAC dilation head	$\rightarrow$4 Conv, Sigmoid, $\rightarrow$5
Training	Adam $\rightarrow$6, initial $\rightarrow$7, MultiStepLR, 1500 epochs
Augmentation	random crop, horizontal/vertical flip, $\rightarrow$8 rotation
Hardware / framework	NVIDIA RTX 3090, PyTorch

These settings indicate that DIKP is a compact predictor rather than a large dense estimator. The use of only five predicted kernels and a $\rightarrow$9 kernel size is presented as an explicit design choice; the subsequent PAC and DSRM stages compensate for the limited cardinality of the kernel bank (Zhang et al., 18 Aug 2025).

6. Interpretation, terminology, and acronym ambiguity

The defining interpretation of DIKP in the FDIKP paper is that it converts the statement “defocus blur is a frequency-domain filter” into a learned prediction of inverse filters. The amplitude branch is described as stabilizing the global filter shape, while the phase branch corrects for local detail, and the two are fused through complex exponentiation and inverse FFT to obtain spatial kernels that PAC can apply adaptively and DSRM can aggregate (Zhang et al., 18 Aug 2025).

An important point of terminology is that the acronym “DIKP” has also been used in earlier literature for a different concept: “Deep Image and Kernel Priors,” a 2019 blind image deconvolution method in which two generator networks $1 \times 1$0 and $1 \times 1$1 parameterize the recovered image and blur kernel via

$1 \times 1$2

and optimize the data-fidelity objective

$1 \times 1$3

That earlier DIKP uses dual branches in the sense of a deep-image-prior branch and a deep-kernel-prior branch, both built with hourglass encoder-decoder networks and jointly optimized with Adam, but it is not the same method as Dual-Branch Inverse Kernel Prediction in FDIKP (Wang et al., 2019).

This acronym overlap can lead to confusion. In the 2025 FDIKP context, DIKP refers specifically to a frequency-domain inverse-kernel predictor with amplitude and phase branches. In the 2019 deconvolution context, DIKP refers to a learning-free blind deconvolution framework based on deep image prior and deep kernel prior parameterizations. The two share an interest in kernel modeling, but their formulations, architectures, and problem settings are distinct (Wang et al., 2019).

The distinction is methodologically consequential. The 2019 work frames deconvolution as direct optimization over network parameterizations of image and kernel, whereas the 2025 DIKP module is a subnetwork embedded inside a larger supervised defocus deblurring system. A plausible implication is that the later use of the acronym is more localized and component-specific, while the earlier one denotes a complete inverse-problem solver.

7. Reported role and significance in defocus deblurring

The motivating claim for DIKP is that most existing methods rely on spatial features for kernel estimation, and that their performance degrades in severely blurry regions where local high-frequency details are missing. The proposed response is to incorporate frequency-domain representations, exploiting what the paper calls the superior discriminative capability of the frequency domain for blur modeling. DIKP is the mechanism by which that frequency information is converted into inverse kernels suitable for deconvolution (Zhang et al., 18 Aug 2025).

The reported broader outcome is that extensive experiments demonstrate that the full method outperforms existing approaches, although the detailed block provided here does not enumerate benchmark names or numerical metrics. Within the architecture, DIKP is presented as the element that improves the accuracy of kernel estimation while maintaining stability, and PAC is introduced because the number of predicted inverse kernels is limited. DSRM then fuses the resulting deconvolution outputs and progressively improves deblurring quality from coarse to fine (Zhang et al., 18 Aug 2025).

In that sense, DIKP occupies a specific niche in contemporary defocus deblurring: it is a physically informed kernel-prediction module that uses separate amplitude and phase pathways to derive a compact bank of inverse filters. Its importance lies less in being a standalone restoration algorithm than in providing a structured intermediate representation for adaptive, multiscale deblurring.