Claim construction and explanation of the principal claim
1) Claim type and transitional phrase
- Statutory class: The claim is directed to a method (process) for image denoising.
- Transitional term: The use of “comprising” renders the claim open-ended. Under the broadest reasonable interpretation (BRI) consistent with the specification, the presence of additional, unrecited steps or components in a practiced method does not avoid infringement, provided each recited limitation is performed.
2) Preamble
- “An image denoising method” functions as a statement of intended use and field. Unless the preamble is necessary to give life, meaning, or vitality to the claim, it is generally not limiting beyond identifying the context in which the steps are performed. Here, the body of the claim recites a complete process; the preamble primarily supplies context.
3) Element-by-element construction under BRI
For clarity, the limitations are enumerated as [A]–[E]:
[A] Acquiring a noisy target image
- Scope: “Acquiring” encompasses receiving, loading, capturing, or otherwise obtaining a digital image that contains noise. No limitation is imposed as to the source, sensor, file format, or noise model.
- Order: This step logically precedes subsequent processing steps; performing the later steps presupposes access to the noisy image.
[B] Extracting multi-scale features with convolution kernels
- Scope: This requires feature extraction at multiple scales using convolutional operations. Under BRI, “multi-scale” reasonably includes use of different kernel sizes, strides, dilations, pyramid features, or hierarchies that yield representations at more than one spatial scale. The claim does not limit the specific network topology, number of layers, or exact convolutional parameters, only that convolution kernels are used to obtain multi-scale features.
- Exclusions: Purely non-convolutional feature extraction (e.g., a pipeline consisting solely of transformer self-attention without any convolution, or only wavelet transforms) would not meet this limitation as written.
[C] Recalibrating channel weights via an attention module
- Scope: This requires an “attention module” that performs channel-wise recalibration, i.e., it computes and applies weights that modulate feature channels. Under BRI, this encompasses any learned mechanism that adjusts per-channel responses conditioned on the features (for example, channel-attention style operations), without limiting the specific architecture or mathematical form.
- Exclusions: An attention mechanism that operates exclusively in the spatial domain without adjusting channel weights would not satisfy this limitation. A mechanism that includes both spatial and channel attention would satisfy the clause insofar as it performs channel-weight recalibration.
[D] Reconstructing a denoised image through deconvolution based on the recalibrated features
- Scope: “Deconvolution” in the convolutional neural network context is reasonably construed to include transposed convolution (also known as fractionally strided convolution) or learned deconvolution operations used for upsampling/decoding. The reconstruction must be “based on the recalibrated features,” i.e., downstream of the channel-weight attention operation.
- Potential ambiguity: If “deconvolution” is intended in the classical signal-processing sense of inverse filtering, the specification should clarify. Under BRI in the CNN art, “deconvolution” commonly denotes transposed convolution. If reconstruction uses only non-deconvolutional upsampling (e.g., bilinear upsampling followed by standard convolution), that may fall outside this limitation.
[E] Wherein parameters of the attention module are obtained by training and remain fixed during inference
- Scope: This is a limiting “wherein” clause. It requires that the attention module’s parameters are learned during a separate training process and are not updated during the claimed inference-time denoising method. The clause does not preclude the attention module from producing different per-input attention outputs at inference; it requires only that the underlying learned parameters are fixed during inference.
- Extent: The clause speaks specifically to the parameters of the attention module; it does not expressly impose the same constraint on other modules, though in ordinary practice inference uses fixed parameters throughout. It also does not require that the training process itself be part of the claimed method.
4) Step order
- Although the claim does not expressly impose an order, the logical and grammatical structure implies the following sequence: acquire the noisy image [A]; extract multi-scale features [B]; apply channel-attention recalibration [C]; and reconstruct the denoised image via deconvolution [D]. Under established principles, an order of steps is required when the claim language, logic, or grammar so dictates. Here, later steps depend on the outputs of earlier steps, supporting an ordered performance.
5) Scope implications and boundaries
- Open-ended nature: Because the claim “comprises” the listed steps, additional operations (e.g., normalization, residual connections, skip connections, non-linearities, loss computations during training, pre- or post-processing such as clipping or color space conversion) may be present without departing from the claim.
- Required technical features:
- Convolution-based multi-scale feature extraction is mandatory.
- Channel-wise attention-based recalibration is mandatory.
- Deconvolution-based reconstruction is mandatory.
- The attention module’s parameters must be learned and fixed at inference.
- Potential noninfringing variations:
- Entirely non-convolutional feature extractors (e.g., pure transformer blocks) for the multi-scale features.
- Attention that does not recalibrate channel weights.
- Reconstruction that omits deconvolution in favor of other upsampling-only schemes.
- Methods that update the attention module’s parameters during inference (e.g., test-time training or online adaptation of those parameters).
6) Plain-language synthesis
- In substance, the claim covers an inference-time image-denoising pipeline in which: a noisy image is obtained; multi-scale features are produced using convolutional layers; these features are subjected to a channel-attention recalibration whose weights are governed by parameters learned offline and kept fixed during inference; and the denoised image is generated from the recalibrated features using a deconvolution-based decoder.
