Bidirectional Scanning

• Bidirectional scanning is a technique that alternates forward and backward traversals to enhance coverage and cancel environmental noise.
• It is widely applied in fields like fiber-laser systems, hyperspectral imaging, and state-space neural models to achieve higher precision and uniformity.
• By leveraging opposed scanning directions, this method improves efficiency and robustness, reducing computational complexity and enhancing performance metrics.

Bidirectional scanning refers to the process, architecture, or algorithmic mechanism in which the underlying scan, beam, feature, or sequence traversals are performed in both forward and backward directions. This principle spans physical instrumentation, optical architectures, and neural network designs across scientific and engineering fields. Bidirectional scanning is employed to improve coverage, model dependencies efficiently, achieve self-correction, or enhance context aggregation. Approaches and implementations differ widely by domain—including fiber-laser frequency combs, hyperspectral image denoising, antenna array control, and spatiotemporal neural modeling—yet key unifying features are reliance on opposed or alternating scan traversals and the resulting symmetry or robustness to environmental variations.

1. Core Principles and Theoretical Foundations

Key to bidirectional scanning is the application of traversal or interaction in opposing directions along a data structure, physical domain, or parameter axis. In the physical sciences, this may mean alternately driving scan elements forward and backward (e.g., optical axes, scanner motions), while in neural computation, it typically denotes sequence processing from start-to-end and end-to-start. Unified theoretical benefits include:

• Common-mode noise rejection: In fiber-laser frequency combs, bidirectional propagation ensures that environmental perturbations (temperature drift, vibration) are imparted identically on both pathways, enabling subtraction and first-order cancellation of timing jitter (Liu et al., 2024).
• Contextual symmetry: In state-space neural models, scanning forward and backward along sequences or graphs ensures each position or token can receive relevant context from both past and future, or from both extremities of the spatial sweep (Liu et al., 2024, Cao et al., 28 Aug 2025, Dong et al., 2024).
• Maximal domain coverage: Optical, mechanical, and electromagnetic scanning benefit from bidirectional approaches to guarantee coverage uniformity, minimize dead zones, and ensure robustness in the face of system asymmetries (Dabiri et al., 13 Dec 2025, Gharibyan et al., 2012, Qin et al., 2024).

2. Physical Instrumentation: Optical and Mechanical Domains

Physical bidirectional scanning is foundational to a number of metrological and communication systems:

• Fiber-laser frequency combs: Liu et al. demonstrate a bidirectional polarization-multiplexed fiber ring laser in which clockwise (CW) and counterclockwise (CCW) pulses share gain and optics. The repetition rate of the CCW comb is rapidly swept using a motorized delay line, while the CW path is fixed. The shared path ensures that any environmental fluctuation is experienced identically by both pulse trains, making their difference frequency immune to first-order noise and permitting reference-free operation with sub-Hz stability over 24 h (Liu et al., 2024).
• Twisting wire scanners: In charged-particle beam diagnostics, bidirectional scanning is achieved using a key-bit fork with two orthogonally mounted wires, attached to a single linear-rotary motor. The scan proceeds along one axis via translation, then perpendicularly via rotation—allowing both horizontal and vertical profiles to be measured in sequence but with bidirectional sweep capabilities and sub-micrometer positional repeatability (Gharibyan et al., 2012).
• Optical vehicular communication: In line-laser vehicle positioning, two orthogonal lasers scan the road longitudinally and transversely. Each laser’s fan beam is steered bidirectionally to ensure entire lane coverage. Analytical frameworks model the effects of scan geometry, beam divergence, dwell time allocation, and optimize for uniformity and link reliability (Dabiri et al., 13 Dec 2025).

3. Bidirectional Scanning in State-Space Neural Models

Bidirectional scanning has achieved prominence in artificial neural architectures, particularly state-space models (SSMs) such as Mamba, which are designed for efficient long-range dependency modeling with lower computational cost than traditional attention:

• Hyperspectral Image Denoising: In HSIDMamba, the Bidirectional Continuous Scanning Mechanism (BCSM) applies SSMs along eight spatial-spectral directions, with forward and backward scans per direction, dramatically enlarging the receptive field over unidirectional or convolutional baselines. Fusing outputs from opposing scan directions results in outputs with improved local and long-range consistency at linear computational complexity. Ablation results show PSNR improvements of nearly +2 dB versus unidirectional scans, with higher structure preservation (Liu et al., 2024).
• Dense Prediction and Cross-task Interaction: The Bidirectional Interaction Scan (BI-Scan) used in multi-task dense prediction serializes task- and position-first views into forward and backward sequences. This structure achieves O(T) complexity (T = number of tasks) for cross-task information mixing versus O(T^2) for pairwise-attention methods, efficiently transferring critical information in both directions and boosting semantic/quantitative prediction metrics (Cao et al., 28 Aug 2025).
• Pose and Hand Estimation: Bidirectional global-local spatio-temporal SSMs (PoseMamba) and graph-guided bidirectional scans (Hamba) use opposed traversals along joint, spatial, and temporal dimensions. This enables fusion of body-wide, limb-specific, and time-symmetric cues, resulting in significant empirical improvements in pose accuracy and parameter efficiency relative to attention-based counterparts (Huang et al., 2024, Dong et al., 2024).

4. Electromagnetic Systems and Antenna Engineering

Bidirectional beam scanning is central to modern reconfigurable antenna design:

• Hybrid Transmitarrays and Polarization Control: In the electromagnetic domain, bidirectional scanning is implemented in hybrid transmitarrays that leverage polarization-rotating metasurfaces, with simultaneous or independent forward and backward beam control. A polarization-switchable multi-feed array positioned between two metasurface layers allows toggling between unidirectional (either +z or –z) and true bidirectional (+/–z) scanning by switching feed polarization among x, y, or 45° states. This results in multibeam operation, low loss, and low-profile apertures, with peak gain up to 23.6 dBi and bidirectional scan ranges of +/-40° (forward) and +/-22° (backward) (Qin et al., 2024).
• Reconfiguration and Efficiency: Using a single set of reconfigurable polarization elements, these antenna arrays achieve bidirectional scanning electronically, without mechanical steering or high power draw, and their architecture can be extended to more complex multibeam or integrated sensing/communication (ISAC) systems.

5. Image Acquisition and Spatiotemporal Alignment

Bidirectional scanning in biomedical imaging addresses both acquisition speed and spatial misalignment:

• Photoacoustic Microscopy: In optical-resolution photoacoustic microscopy (OR-PAM), bidirectional scanning alternates the sense of raster scanning on each line (odd/even), doubling volumetric frame rates. However, this introduces direction-dependent domain shifts and geometric distortions due to actuator dynamics and system lags. Domain-invariant registration cannot be achieved by standard methods that assume brightness constancy. The SAS-Net architecture separates scene and appearance codes, enabling cross-domain reconstructions and robust alignment. Empirical results show normalized cross-correlation of 0.961 and SSIM of 0.894 on in vivo datasets, outperforming all conventional baselines (Qin, 6 Feb 2026).

6. Practical Impact, Limitations, and Comparative Assessment

Bidirectional scanning ensures robustness, context, and coverage enhancements unattainable or less efficiently achieved via unidirectional methods. Distinctive advantages, as documented across instrumentation and neural computing domains, include:

• Noise resilience: Intrinsic self-correction and common-mode rejection in shared-path architectures (Liu et al., 2024).
• Efficiency and scalability: Linear complexity in neural interaction modeling, parameter and memory reduction, and near real-time processing (Liu et al., 2024, Cao et al., 28 Aug 2025, Dong et al., 2024, Huang et al., 2024).
• Maximal spatial or temporal coverage: Minimized coverage “holes,” improved high-SNR communication regions, and uniformity in spatial scanning for sensing and data transfer (Dabiri et al., 13 Dec 2025).
• Architectural parsimony: Compact mechanical and electromagnetic implementations, reduced need for additional reference or calibration elements, and power efficiency (Gharibyan et al., 2012, Qin et al., 2024).

Nonetheless, some limitations are inherent: Physical bidirectional mechanisms may be bounded by mechanical inertia or scanning speed (e.g., mechanical stage limits in fiber-laser scanning), and neural bidirectional passes are not always feasible in strictly causal or low-latency applications. Opportunities for further research include extending bidirectional scanning to higher-dimensional data, optimizing fusion strategies, and integrating with adaptive control or online feedback to further enhance robustness or contextual reception.