Orientation Canonicalization: Rotation-Invariant Processing Through Gradient Normalization

Mechanism

The canonicalization mechanism operates as a deterministic pipeline applied to content before any identity computation is performed. The pipeline accepts the content in its native presentation, extracts the geometric features that determine its orientation, normalizes the orientation against a canonical reference, and emits a transformed representation whose canonical orientation is invariant under the bounded class of input transforms.

The first stage of the pipeline computes a dense gradient field over the content. The gradient field measures the local rate of change of the content's intensity in each spatial direction. The dominant orientation of the content is then estimated by aggregating the gradient field into an orientation histogram and selecting the histogram peak, weighted by gradient magnitude. The peak corresponds to the direction along which the content's structural variation is most strongly aligned, and that direction is taken as the content's intrinsic orientation.

The second stage rotates the content so that its intrinsic orientation aligns with a canonical reference axis. The rotation is performed using a continuous interpolation kernel that preserves the gradient structure across the resampling. The third stage translates the content so that its centroid, computed as the gradient-weighted center of mass, sits at the origin of the canonical coordinate system. The fourth stage rescales the content so that a configured structural radius, typically the gradient-weighted second moment, matches the canonical reference scale.

The output of the pipeline is the canonical form. Two inputs that differ only by rotation, translation, and scale within the configured bounds map to byte-equivalent canonical forms, and downstream identity computation produces identical results. Inputs that differ by transforms outside the bounds, such as non-rigid deformations, content edits, or compositional changes, produce distinguishable canonical forms.

The pipeline is deterministic in the strict sense: given the same input bytes and the same pipeline parameters, every node in the content anchoring system produces the same output bytes. Determinism is enforced by specifying the gradient operator, the histogram quantization, the peak-selection rule, the interpolation kernel, and the rescaling rule precisely, and by requiring all participants to apply identical specifications. There is no learned model in the pipeline; every stage is a closed-form function of its inputs.

The canonical form is committed to the lineage of the content as a structural anchor. The commitment binds the canonical bytes to the cryptographic identity that the rest of the content anchoring system uses, so any subsequent operation that consults the identity is consulting a value derived from the canonical form rather than from a particular presentation of the content.

Operating Parameters

The mechanism exposes a structured set of operating parameters. The first parameter is the gradient operator. The pipeline declares the convolution kernel that produces the gradient field. Standard choices include centered finite-difference kernels, Sobel kernels, and Scharr kernels. The choice is fixed at protocol level so that all participants apply the same operator.

The second parameter is the histogram quantization. The orientation histogram is quantized into a fixed number of bins covering the angular range. The quantization determines the resolution at which the dominant orientation can be detected. Finer quantization improves orientation precision at the cost of histogram noise; coarser quantization reduces noise at the cost of precision.

The third parameter is the peak-selection rule. When the histogram has multiple comparable peaks, the rule specifies which peak is taken as the dominant orientation. The rule must be deterministic and independent of presentation order. Standard rules include the highest-magnitude peak with a tie-breaking convention based on bin index, and the peak with the highest spatial coherence in the underlying gradient field.

The fourth parameter is the interpolation kernel for the rotation, translation, and scaling stages. The kernel determines how the resampled values are computed at output positions that do not coincide with input grid points. Standard choices include bicubic kernels, Lanczos kernels with a configured support, and B-spline kernels. The kernel must be specified precisely enough that all participants produce identical outputs.

The fifth parameter is the canonical reference geometry. The reference axis, the reference origin, and the reference scale are declared at protocol level. Two participants that disagree on the reference geometry will produce non-matching canonical forms even on identical inputs, so the reference must be globally fixed.

The sixth parameter is the bounded transform class. The mechanism guarantees invariance only within a configured bound on the rotation angle, the translation distance, and the scale factor. Inputs that exceed the bound may map to distinguishable canonical forms because the canonicalization stages clip or fail rather than produce arbitrary outputs. The bounds are declared per scope and represent the maximum transform that the scope considers a presentation of the same underlying content.

The seventh parameter is the substrate portability profile. The pipeline declares the floating-point precision, the rounding mode, and the byte order of its intermediate computations. The profile is what allows the canonical form to be byte-identical across different processors, operating systems, and runtime environments. A participant that cannot meet the profile cannot produce conforming canonical forms.

Alternative Embodiments

In a first embodiment, the content is two-dimensional imagery and the gradient field is computed by direct convolution with a two-dimensional kernel. In a second embodiment, the content is three-dimensional volumetric data and the gradient field is computed in three dimensions, with the dominant orientation determined by a three-dimensional histogram and the canonicalization stages extended to three rotational axes. In a third embodiment, the content is a one-dimensional waveform and the gradient field reduces to a derivative, with orientation expressed as the sign of a configured polarity feature.

In a fourth embodiment, the histogram is replaced by a continuous orientation estimator based on the structure tensor. The structure tensor's principal eigenvector identifies the dominant orientation directly without quantization. The structure-tensor embodiment trades implementation complexity for finer orientation precision and is suitable for content where the gradient field is dense and well-conditioned.

In a fifth embodiment, the rotation, translation, and scaling stages are merged into a single affine resampling step parameterized by a 2-by-3 matrix. The affine merge reduces the number of resampling passes and therefore the cumulative interpolation error, at the cost of a more complex specification.

In a sixth embodiment, the canonical form is augmented with a reflection canonicalization stage that selects a canonical chirality. The chirality stage is appropriate for content classes where mirror reflections should map to the same identity; it is omitted for content classes where chirality is structurally significant.

In a seventh embodiment, the bounded transform class is parameterized per scope, so different scopes can declare different invariance bounds depending on their semantic requirements. A scope that anchors documents may declare tight bounds because document orientation is meaningful; a scope that anchors object scans may declare loose bounds because the scan direction is incidental.

In an eighth embodiment, the canonical form is produced redundantly by multiple independent implementations of the pipeline, and the canonical form is accepted only when all implementations agree byte-for-byte. The redundant embodiment defends against silent implementation drift and is appropriate for high-assurance deployments where the canonical form is consumed by downstream parties that cannot tolerate divergence.

In a ninth embodiment, the lineage commitment of the canonical form is accompanied by a witness package that includes the orientation histogram, the centroid coordinates, and the structural radius. The witness package allows third parties to verify that the canonicalization was applied correctly without needing to re-run the full pipeline.

Composition

The orientation canonicalization mechanism composes with the broader content anchoring system along several axes. It composes with structural identity computation: the identity is computed over the canonical form rather than the raw input, so the identity is invariant under the bounded transform class by construction. Two presentations that differ only by such transforms produce identical identities, and the system can recognize them as the same content without consulting any external metadata.

The mechanism composes with the lineage layer. The canonical form is committed to the lineage as a structural anchor, and any operation that subsequently mutates, derives, or republishes the content references the canonical form rather than the presentation. The lineage therefore carries a single canonical thread through the content's history regardless of how many presentations of the content exist.

The mechanism composes with substrate portability. The canonical form is independent of the device, format, or pipeline that produced the content, which means the same identity can be reproduced on any substrate that implements the canonicalization specification. Cross-substrate portability is structural rather than procedural: there is no migration step required when the same content moves between substrates because the canonical form is the same on both.

The mechanism composes with derivative-content tracking. A derivative that preserves the canonical form within the bounded transform class is recognized as a presentation of the original content. A derivative that exceeds the bound, such as a substantive edit or a compositional change, produces a distinguishable canonical form and is recognized as a distinct content with its own lineage. The threshold between presentation and derivative is therefore determined by the bound parameter rather than by an external classifier.

Prior-Art Distinctions

Prior systems for content identity fall into several categories. Cryptographic hash functions applied directly to content bytes produce identities that are sensitive to every bit of the input, including bits that change under presentation transforms such as rotation. The mechanism described here produces identities that are invariant under such transforms within configured bounds, which is structurally distinct from byte-level hashing.

Perceptual hash functions produce identities that are robust to certain transforms but generally rely on learned models or heuristics whose outputs can drift across implementations and across versions of the model. The mechanism described here is fully deterministic and specified at the protocol level, so two participants that implement the specification correctly produce byte-identical canonical forms. There is no learned component whose drift could cause identity divergence.

Watermarking and metadata schemes attach external identifiers to content. The identifiers can be stripped, altered, or detached during normal handling, and their continued presence depends on every intermediate party preserving them. The mechanism described here computes identity from the content's intrinsic structure, so no external attachment is required and no intermediate party can detach the identity by stripping a tag.

Feature-descriptor schemes such as those used in image retrieval extract local descriptors and match them across instances. These schemes are designed for similarity search rather than for identity, and they do not produce a single canonical form whose downstream hash is reproducible. The mechanism described here is structurally distinct in that it produces a single canonical form with byte-level reproducibility, suitable for use as a primary identity rather than as a similarity score.

Disclosure Scope

The disclosure encompasses any embodiment of the mechanism that satisfies the structural properties described above: deterministic preprocessing of content into a canonical form; invariance of the canonical form under rotation, translation, and scale within configured bounds; portability of the canonical form across substrates; and commitment of the canonical form to the lineage of the content as a structural anchor.

The disclosure encompasses two-, three-, and one-dimensional content embodiments; histogram-based and structure-tensor-based orientation estimators; sequential and merged resampling pipelines; with-chirality and without-chirality canonicalization profiles; per-scope parameterization of the bounded transform class; and redundant multi-implementation production of the canonical form. It encompasses lineage commitments accompanied by witness packages for third-party verification, and lineage commitments without such packages for minimal deployments.

Embodiments that omit the bounded-transform invariance, that produce non-deterministic canonical forms, or that depend on external metadata for identity continuity fall outside the disclosure. The structural properties are not optional features; they are the properties that distinguish the mechanism from prior art and that justify its use as a primary identity layer in content anchoring systems.