Modular subsea compute camera with integration control and on-board learned scene representation generation

Abstract

An embedded processing unit 140 disposed within a pressure-rated subsea housing 110 and thermally coupled thereto for heat rejection to surrounding seawater, cause the embedded processing unit to: (a)

Description

FIELD OF THE INVENTION The present invention relates to subsea imaging systems for underwater vehicles and fixed subsea stations. More particularly, the invention relates to modular single-camera compute cameras and optional multi-sensor variants incorporating a disciplined time base, embedded high-performance computing for on-board generation of learned scene representations including three-dimensional and time-varying scene models, integration control managing the interdependent constraints of timing quality, computational thermal load, and transmission bandwidth, and thermal management architectures enabling sustained GPU-class processing within pressure-rated dry housings at operational depths. BACKGROUND OF THE INVENTION Subsea inspection operations for offshore energy infrastructure, carbon capture and storage monitoring, pipeline survey, and marine construction rely on remotely operated vehicles (ROVs) and autonomous underwater vehicles (AUVs) equipped with imaging systems. Current subsea camera systems typically capture two-dimensional video that is transmitted to the surface for real-time operator viewing and post-mission analysis. The subsea inspection industry faces several interconnected challenges that existing camera systems do not adequately address. First, bandwidth constraints are fundamental. The communication link between a subsea vehicle and its surface control station is bandwidth-limited, whether via fibre-optic tether, acoustic modem, or other links. Transmitting high-resolution raw imagery in real-time is impractical at the resolutions required for metrology-grade capture. Current systems address this by compressing video using conventional codecs, which can degrade geometric fidelity and does not provide a scene-level representation suitable for immediate measurement and model delivery. Second, three-dimensional reconstruction from subsea imagery is currently performed topside, post-mission. This introduces delays of hours to days between data capture and actionable 3D deliverables. For applications requiring real-time spatial awareness such as approach guidance to subsea assets, collision avoidance near complex structures, or immediate defect quantification post-mission processing is inadequate. Third, timing integrity in subsea environments presents unique challenges. Temperature gradients at depth cause clock drift in electronic oscillators, and the subsea communication medium introduces variable latency with asymmetric path delays. Even in a single-camera system, timestamp uncertainty relative to a reference clock degrades the integrity of time-indexed scene models and the alignment of imagery with navigation data; in multi-sensor variants, inter-sensor exposure alignment is additionally required. Fourth, thermal management of high-performance embedded computing within pressurerated dry housings is an unsolved integration challenge. GPU-accelerated scene reconstruction generates sustained thermal loads of 30 watts or more, which is substantially higher than the low-power electronics historically housed in subsea modules. Without adequate thermal coupling to the surrounding seawater junction temperatures exceed safe operating limits within minutes, causing throttling or shutdown. Fifth, there remains a need for systems that jointly manage synchronisation quality, computational thermal state, and output bandwidth as interdependent constraints. In current systems, these subsystems operate independently. A thermal overload may cause uncontrolled throttling that degrades synchronisation accuracy, producing geometrically warped 3D reconstructions without any indication of degraded quality. There is no mechanism to adaptively manage reconstruction parameters in response to thermal constraints while simultaneously ensuring that the output data carries a metrological confidence indicator tied to actual synchronisation performance. Sixth, subsea camera systems are typically hardwired to the vehicle and require dry-dock intervention for replacement or upgrade. As reconstruction algorithms and sensor technologies evolve rapidly, the inability to field-swap camera modules creates operational inflexibility and asset obsolescence. SUMMARY OF THE INVENTION The present invention provides a modular subsea imaging and processing system that addresses the above limitations through an integrated architecture combining five principal innovations. First, a disciplined timing subsystem operating within a pressure-rated dry housing maintains a time base for scheduled exposure and timestamp integrity despite environmental disturbances including thermal clock drift and tether latency variation with path-asymmetry compensation and optionally provides synchronised exposure across multiple image sensors in multi-sensor variants. Second, a thermal coupling architecture provides a conductive heat path from an embedded GPU-class processing unit to the pressure-rated housing wall, enabling passive dissipation to the surrounding seawater at rates sufficient to sustain continuous on-board scene processing without active cooling. Third, the embedded processing unit generates a learned scene representation comprising a three-dimensional scene model and optionally time-varying scene updates (for example, time-indexed model deltas) from captured imagery, enabling rapid delivery of 3D and 4D-style outputs during a mission. Fourth, a dual-path data output architecture simultaneously archives high-resolution raw image data locally and transmits a compressed representation of the learned scene representation and / or intermediate reconstruction data over a bandwidth-constrained communication link. In a preferred embodiment, an encoded high-resolution video stream is additionally transmitted over a second communication channel, which may be a separate fibre-optic link within the umbilical or a dedicated second umbilical, for real-time operator viewing. Fifth, an integration controller jointly manages timing quality, thermal state, and output bandwidth as interdependent constraints. The integration controller modifies reconstruction parameters of the embedded processing unit in response to thermal state measurements and annotates the learned scene representation with validity indicators derived from timing quality metrics, thereby providing quantitative confidence bounds on metrological quality and ensuring graceful degradation under thermal or link stress. For the purposes of this specification and the claims that follow, the following terms shall have the meanings set out below unless the context requires otherwise. As used herein, a “pressure-rated housing” means a housing configured to maintain an internal environment at approximately one atmosphere (dry, gas-filled) while structurally resisting external hydrostatic pressure at operational depth. The housing is sealed against water ingress and is not oil-filled or pressure-compensated by flooding with a dielectric fluid. Internal components operate in a dry gaseous environment. The housing walls, end-caps, seals, optical windows, and penetrators are each rated to withstand the maximum external hydrostatic pressure corresponding to the rated operational depth. As used herein, a “bandwidth-constrained communication link” means a communication link whose aggregate data throughput capacity is less than the aggregate raw output data rate of all image sensors in the imaging module operating simultaneously at full resolution and frame rate. This includes, without limitation, fibre-optic tethers, acoustic modems, copper Ethernet, and satellite relay links. As used herein, a “communication link” means one or more physical or logical channels between a subsea module and a surface control station. A communication link may comprise a first bandwidth-constrained channel for transmission of a compressed scene representation and a second channel for transmission of an encoded high-resolution video stream. The first and second channels may share a single physical umbilical, occupy separate fibres within a single umbilical, or use entirely separate umbilicals. As used herein, a “three-dimensional scene model” means a digital representation of a physical scene in three spatial dimensions, generated from two-dimensional imagery through a computational reconstruction process. The scene model encodes sufficient geometric and appearance information to enable re-rendering of the scene from arbitrary viewpoints and / or dimensional measurement of scene features. The term encompasses, without limitation, Gaussian splatting representations, neural radiance fields, compressed point clouds with attribute data, textured polygon meshes, and future three-dimensional representation types. As used herein, a “compressed scene representation” means a data-reduced encoding of a three-dimensional scene model that preserves sufficient geometric and appearance fidelity to enable remote re-rendering and / or dimensional measurement, while occupying substantially less data bandwidth than the raw image data from which the scene model was generated. The compression ratio is at least 10:1 relative to the aggregate raw sensor data rate. As used herein, an “integration controller” means a control subsystem that receives inputs from two or more of the timing and synchronisation subsystem, the thermal management subsystem, and the data output subsystem, and generates control outputs that modify operation of the embedded processing unit and / or annotate the data output, thereby managing synchronisation quality, thermal state, and output bandwidth as interdependent constraints. As used herein, a “validity indicator” means a metadata annotation applied to a learned scene representation or compressed representation thereof, derived from one or more of synchronisation quality metric, thermal state measurement, and sensor health status, indicating the metrological confidence level of the annotated data. As used herein, a “synchronisation quality metric” means a quantitative measure of timing integrity derived from precision time protocol offset measurements, oscillator stability variance, holdover duration, or equivalent timing quality indicators. In single-sensor configurations, the metric characterises time-base accuracy relative to a reference clock. In multi-sensor configurations, the metric additionally characterises the calculated temporal variance across the plurality of image sensors. As used herein, a “learned scene representation” means a three-dimensional scene model generated through an optimisation or machine-learning inference process from two-dimensional imagery, including but not limited to neural radiance fields, Gaussian splatting representations, point clouds derived from learned depth estimation, and multi-view stereo reconstructions enhanced by learned priors. As used herein, a “time-varying learned scene representation” means a learned scene representation that is updated over time and comprises a sequence of model states and / or incremental model updates associated with capture timestamps, thereby representing scene change over time. BRIEF DESCRIPTION OF THE DRAWINGS FIG. lisa schematic block diagram showing the overall system architecture, comprising the imaging module (100), host vehicle interface (200), surface control station (300), and communication link (400). FIG. 1 illustrates the integration controller (180) and its crosssubsystem feedback loops: receiving a thermal state measurement from the thermal management subsystem (150) and a synchronisation quality metric from the timing subsystem (130); outputting parameter control signals (resolution, frame rate, complexity, learned representation update rate) to the embedded processing unit (140); and outputting validity indicator annotations to the transmission path (164) of the dual-path data output architecture (160). FIG. 2 is a cross-sectional view of the imaging module (100) showing the arrangement of an image sensor (120), embedded processing unit (140), thermal management subsystem (150), integration controller (180), and local storage (166) within the pressure-rated housing (110). FIG. 3 is a detail view of the thermal management subsystem (150) showing the thermal interface plate (152), compliant thermal interface material (154), and machined interior housing surface for heat transfer to the housing wall. FIG. 4 is a schematic diagram of the timing and synchronisation subsystem (130) showing the local master oscillator (132), synchronisation bus (134), temperature sensor (136), drift correction feedback path, path-asymmetry correction module, and output of the synchronisation quality metric to the integration controller (180). FIG. 5 is a data flow diagram showing the dual-path data output architecture (160) with the local archive path (162) and transmission path (164), including the validity indicator annotation point where the integration controller (180) embeds metrological confidence metadata into the transmitted compressed learned scene representation, and optionally a second transmission channel (165) for an encoded high-resolution video stream. FIG. 6 is a flowchart showing the integration controller’s thermal throttling cascade: fullpower reconstruction mode, reduced-power mode (with reduced resolution, frame rate, learned representation update rate, or model complexity), and pass-through recording mode, with transition thresholds and hysteresis. FIG. 7 is a schematic showing the modular connection architecture with wet-mate connectors (170) and quick-release mounting mechanism (172). FIG. 8 is a flowchart showing the auto-configuration sequence (174) executed upon module connection to the host vehicle interface (200). FIG. 9 is a schematic showing the integration controller’s synchronisation quality monitoring and validity annotation process, including the critical threshold below which learned scene representation generation is suspended. DETAILED DESCRIPTION OF THE INVENTION 1. System Architecture Overview Referring to FIG. 1, the system comprises an imaging module (100), a host vehicle interface (200), and a surface control station (300) connected via a communication link (400). The imaging module (100) is contained within a pressure-rated housing (110) configured to maintain an internal environment at approximately one atmosphere (dry, gas-filled) while structurally resisting external hydrostatic pressure at operational depth. The depth rating is determined by the housing wall thickness, material selection, seal ratings, and penetrator specifications for the intended deployment environment. The housing (110) is fabricated from a corrosion-resistant material, preferably 6000-series anodised aluminium, 316L stainless steel, or titanium alloy, selected based on depth rating and weight requirements. The housing walls, end-caps, and all penetrators are sealed against water ingress using face seals, radial O-ring seals, or equivalent pressure-rated sealing arrangements. Internal components operate in a dry gaseous environment, typically dry air or dry nitrogen. 2. Image Sensor The imaging module (100) houses at least one image sensor (120), preferably a glob al-shutter CMOS sensor with a resolution sufficient for metrology-grade data capture and three-dimensional scene reconstruction at the required measurement precision. Global shutter sensors are preferred over rolling shutter sensors to reduce motion artefacts during vehicle movement, improving geometric reconstruction quality. In a single-sensor embodiment, the image sensor captures sequential frames as the host vehicle traverses the scene, providing the multi-view image data required for three-dimensional scene reconstruction. In optional multisensor embodiments, the module houses a plurality of image sensors arranged as an array to provide overlapping fields of view and increased reconstruction robustness; multi-sensor selfcalibration and alignment through visual and radio-frequency means is addressed in copending patent application PAT-005. The present invention is not limited to any particular number of sensors or sensor resolution. The image sensor (120) is mounted behind an optical port (122) in the housing wall. The optical port comprises a pressure-rated optical window rated to withstand the maximum external hydrostatic pressure at the rated operational depth. In embodiments employing a single sensor, the optical port may be a flat window or a dome window sized for the field of view. In embodiments employing multiple sensors, respective windows may be provided for each sensor, or a shared dome port may be used for a co-located array, and window angles may be selected to provide overlapping fields of view. 3. Precision Timing and Synchronisation Subsystem A timing and synchronisation subsystem (130) implements a precision time protocol conforming to IEEE 1588-2019 or successor standards. A local master oscillator (132) within the imaging module serves as the time base. The time base is used to schedule exposure of the at least one image sensor (120) and to assign timestamps to captured imagery for alignment with navigation and other sensor streams. In embodiments comprising multiple image sensors, each sensor is connected to the time base via a synchronisation bus (134) that triggers exposure at scheduled times within a specified tolerance. In a first aspect, the timing subsystem compensates for thermal clock drift by monitoring the housing internal temperature via a temperature sensor (136) and applying a temperaturedependent drift correction model that adjusts the master oscillator frequency as a function of measured temperature. The drift correction model may be factory-calibrated or field-updated. This enables low-jitter timestamp integrity relative to a reference clock, and in multi-sensor embodiments maintains sub-100 microsecond exposure alignment across sensors over the operating temperature range of the subsea environment (typically 2 to 25 degrees Celsius). In a second aspect, the timing subsystem compensates for tether latency variation by operating the local master oscillator within the imaging module rather than deriving time from a surface clock. The local master oscillator is periodically disciplined by a surface PTP grandmaster when communication latency permits but operates autonomously in holdover mode between discipline events. The timing subsystem further measures round-trip communication delay to the surface reference and applies a path-asymmetry correction to account for asymmetric propagation delays in fibre-optic or electrical tethers. The timing subsystem continuously computes a synchronisation quality metric, such as the PTP offset from the reference clock, the oscillator stability variance, the holdover duration, or in multi-sensor embodiments the calculated temporal variance across the plurality of image sensors, or a combination thereof. This metric is provided to the integration controller (180) as described below. 4. Embedded Processing Unit and Thermal Management An embedded processing unit (140) comprising a GPU-accelerated computing module is housed within the imaging module (100). In a preferred embodiment, the processing unit is an NVIDIA Jetson-class system-on-module, though the invention is not limited to any particular computing platform. The processing unit is configured to execute a learned scene representation pipeline that generates a three-dimensional scene model and optionally a time-indexed sequence of scene updates from captured image data. In embodiments, the learned scene representation may comprise Gaussian splatting primitives (3DGS) and, where time-indexed updates are produced, a time-varying representation (4D-style) suitable for incremental streaming. A thermal management subsystem (150) provides a conductive thermal path from the processing unit (140) to the housing wall (112), which in turn rejects heat to the surrounding seawater by conduction and natural convection. In a preferred embodiment, this comprises a copper or aluminium thermal interface plate (152) mechanically pressed against both the processing unit heatsink and an interior machined flat on the housing wall. A compliant thermal interface material (154), such as a phase-change compound or thermally conductive gap pad, accommodates manufacturing tolerances and thermal expansion differentials. In an alternative embodiment, the thermal management subsystem comprises one or more heat pipes (156) connecting the processing unit to a dedicated thermal dissipation zone on the housing where the wall thickness is reduced or fins are machined on the exterior to maximise convective heat transfer to the surrounding seawater. The thermal management subsystem includes temperature sensing at the processing unit junction and / or along the conductive thermal path. The temperature measurements are provided to the integration controller (180). 5. Integration Controller An integration controller (180), shown in FIG. 1, acts as a unified control subsystem managing the interdependent constraints of computational thermal load, sensor timing accuracy, and transmission bandwidth. The integration controller (180) is the central coordinating element that distinguishes the present invention from prior art sealed camera systems in which thermal management, synchronisation, and data output operate as independent subsystems. The integration controller (180) receives at least two of the following inputs: (i) a thermal state measurement from the thermal management subsystem (150) or directly from the embedded processing unit (140), such as a junction temperature reading; (ii) a timing quality metric from the timing and synchronisation subsystem (130), such as the PTP offset from a reference clock, a jitter statistic, a holdover duration, and, in multi-sensor embodiments, a calculated temporal variance across sensors; and (iii) a bandwidth utilisation measurement from the transmission path (164). Based on these inputs, the integration controller (180) generates control outputs that modify operation of the embedded processing unit (140) and / or annotate the data output of the transmission path (164). In particular, the integration controller may jointly trade off reconstruction fidelity, sensor capture rate, learned representation update rate, and transmitted representation quality based on a combined evaluation of thermal headroom, synchronisation quality, and available link capacity. 5.1 Thermal Throttling Cascade In a preferred embodiment, the integration controller (180) is configured to execute a thermal throttling cascade comprising at least three operational modes, as illustrated in FIG. 6. Full-Power Mode: While the thermal state measurement remains below a first predetermined threshold (for example, 75 degrees Celsius junction temperature), the integration controller operates the embedded processing unit (140) in a full-power mode. In this mode, the system performs high-resolution, high-frequency three-dimensional scene reconstruction. For example, the processing unit may generate a learned scene representation from full-resolution images captured at 30 frames per second. Reduced-Power Mode: If the thermal state measurement exceeds the first predetermined threshold, the integration controller (180) automatically transitions to a reduced-power mode. In this mode, the integration controller actively modifies one or more operating parameters of the scene reconstruction pipeline to reduce the GPU compute load, thereby lowering thermal dissipation. This modification comprises at least one of: reducing the input image resolution (for example, pixel binning to a lower resolution), reducing the frame capture rate (for example, from 30 fps to 15 fps), reducing the learned representation update rate, or reducing the scene model complexity (for example, capping the maximum number of Gaussian primitives or limiting the number of inference iterations in the reconstruction pipeline). Pass-Through Mode: If the thermal state measurement continues to rise and exceeds a second, higher predetermined threshold (for example, 85 degrees Celsius junction temperature), the integration controller (180) transitions to a pass-through mode. In this mode, on-board learned scene representation generation is suspended entirely. The system reverts to archiving raw full-resolution image data to the local archive (162) while transmitting status metadata and / or a compressed video stream over the bandwidth-constrained communication link (400). In embodiments providing a second communication channel (165), the module may additionally continue transmitting an encoded high-resolution video stream subject to available link capacity and thermal headroom. Once the thermal state measurement drops back below the operational thresholds, the integration controller steps back up the cascade, restoring reconstruction capability progressively. 5.2 Synchronisation Quality and Validity Annotation According to a further aspect of the present invention, the integration controller (180) tightly couples the precision timing subsystem (130) with the scene reconstruction pipeline to ensure metrological integrity of the output data. Subsea environmental disturbances particularly rapid thermal gradients caused by thermocline crossings, proximity to hot subsea equipment, or rapid depth changes can induce transient clock drift in the local master oscillator (132) that exceeds the temperature-dependent drift correction model’s steady-state assumptions. The integration controller (180) continuously monitors the synchronisation quality metric provided by the timing subsystem (130). As the embedded processing unit (140) generates the learned scene representation, the integration controller (180) dynamically annotates the representation or the compressed representation thereof with a validity indicator derived directly from the synchronisation quality metric. Metrological Confidence Annotation: The validity indicator serves as a real-time confidence bound embedded in the output data. If the timing quality metric indicates a low offset and jitter relative to a reference clock, the validity indicator flags the resulting learned scene representation as metrology-grade suitable for precise dimensional measurement such as corrosion pit depth quantification, weld profile measurement, or anode depletion assessment, with quantified uncertainty bounds. If the timing quality metric degrades (for example, increased offset during thermal shock or extended holdover), the validity indicator flags the resulting output as degraded suitable for remote visualisation and piloting but possessing a higher uncertainty bound for quantitative measurement. In multi-sensor embodiments, the validity indicator may additionally reflect inter-sensor exposure alignment. The validity indicator may further encode a timing uncertainty value, a thermal confidence state, or both, enabling downstream systems to make automated quality-gated decisions. Reconstruction Suspension on Sync Degradation: In severe cases where the synchronisation quality metric falls below a critical operational threshold indicating a loss of acceptable timing precision the integration controller (180) is configured to immediately suspend generation of the learned scene representation to prevent the creation of geometrically warped data. The system reverts to archiving raw image data to the local archive (162), which can be reconstructed post-mission once timing alignment is verified. This suspension is independent of the thermal throttling cascade and may occur at any thermal state. 6. Dual-Path Data Output Architecture The imaging module implements a dual-path data output architecture (160) comprising a local archive path (162) and a transmission path (164). The local archive path (162) stores raw image data from the at least one image sensor at full resolution and bit depth to a local storage medium (166) within the imaging module. The storage medium is preferably a solid-state drive rated for the operational vibration and temperature environment. The raw data archive preserves the full information content of the captured imagery for post-mission processing, including re-running reconstruction algorithms as they improve over time. The transmission path (164) outputs a compressed scene representation generated by the embedded processing unit (140). This compressed representation may take one of several forms. In a first embodiment, it comprises a serialised learned scene model, such as a set of Gaussian splatting primitives encoded with position, covariance, colour coefficients, and opacity, suitable for real-time re-rendering at the surface control station (300). In a second embodiment, it comprises a compressed point cloud with per-point colour, normal vectors, and confidence attributes. In a third embodiment, it comprises a textured polygon mesh with a compressed texture atlas. In all embodiments, the compressed scene representation occupies substantially less bandwidth than the raw imagery at a compression ratio of at least 10:1 relative to the aggregate raw sensor data rate enabling real-time or near-real-time delivery over the bandwidth-constrained communication link (400). The transmission path also carries intermediate reconstruction data, such as keyframes, camera poses, or partial model blocks, enabling completion or refinement of the learned scene representation at the surface control station. In a preferred embodiment, the module further comprises a video encoder configured to encode a high-resolution video stream from the raw image data. The encoded video stream is transmitted over a second communication channel (165), which may be a separate fibre-optic link within the same umbilical, a dedicated second umbilical, or a multiplexed channel on the same physical link. In a preferred embodiment, the video encoder implements H.265 / HEVC or successor codec and produces an encoded video stream at the full native resolution of the image sensor. The encoded video stream and the compressed learned scene representation are co-equal first-class outputs of the imaging module. 7. Modular Physical Architecture The imaging module (100) connects to the host vehicle interface (200) via one or more wetmate connectors (170). The wet-mate connectors provide power, communication, and optionally auxiliary sensor data connections. The connectors are of a type suitable for subsea mating and de-mating, including but not limited to SubConn Circular, Teledyne ODI Nautilus, or equivalent types. The imaging module is physically secured to the host vehicle via a quick-release mounting mechanism (172) permitting removal and replacement without specialist tools or with simple ROV-operable latches. This enables field replacement of the complete imaging and processing module at the quayside, on the vessel deck, or in certain embodiments subsea using an intervention ROV. Upon connection of a replacement module, an auto-configuration subsystem (174) detects the module type and capabilities, loads appropriate processing parameters, and initiates a sensor health check and synchronisation verification sequence. This auto-configuration eliminates the need for manual software configuration following module replacement. 8. Housing Deformation Monitoring In an optional embodiment, the imaging module further comprises a housing deformation monitoring subsystem. In a first form, one or more strain sensors (for example, strain gauges or fibre Bragg grating sensors) are bonded to the interior surface of the housing wall at locations where finite-element analysis predicts maximum deformation under hydrostatic loading. In a second form, fiducial markers (precision-machined reference features) are mounted at known positions within the housing interior and are visible to at least one image sensor. In a third form, a pressure-depth lookup table correlates external hydrostatic pressure with predicted housing deformation from factory calibration data. When the deformation monitoring subsystem detects deformation exceeding a predetermined geometric tolerance, the integration controller (180) flags or invalidates learned scene representation data accordingly. 9. Integration with SDAV Framework The imaging module is designed to interoperate with the applicant’s Subsea Data Acquisition and Verification (SDAV) framework, as described in existing patent applications GB2602238.4 and GB2602720.1. The three-dimensional scene representations generated by the imaging module may be provided to a subsea asset verification system as described in copending patent application PAT-006 for temporal comparison and degradation quantification. The imaging module may also receive navigation data from a cross-modal navigation system for geo-referencing of captured scene data. Multi-sensor self-calibration and alignment through visual and radio-frequency means is the subject of co-pending patent application PAT-005. These integrations are optional and the imaging module operates independently as a standalone system.

Claims

1. A modular subsea imaging and processing module comprising: (a) a pressure-rated housing (110) configured for subsea operation, the housing maintaining approximately one atmosphere internal pressure and configured to resist external hydrostatic pressure at operational depth; (b) at least one image sensor (120) arranged within the housing to capture images through an optical window; (c) a timing and synchronisation subsystem (130) implementing a precision time protocol to maintain a time base for the at least one image sensor (120) and to initiate exposure of the at least one image sensor at scheduled times, the timing and synchronisation subsystem comprising temperature-dependent drift correction based on internal temperature measurement; (d) an embedded processing unit (140) disposed within the housing and thermally coupled to the housing via a conductive thermal path (150) configured to reject heat to surrounding seawater, the embedded processing unit configured to generate a learned scene representation comprising a three-dimensional scene model and optionally a time-varying sequence of scene model updates from image data captured by the at least one image sensor; (e) a dual-path data output (160) comprising: (i) a local archive storing raw image data at full resolution; and (ii) a transmission path (164) configured to transmit, over a communication link (400) whose data capacity is lower than the data rate required to transmit the raw image data, a compressed representation of the learned scene representation and / or intermediate reconstruction data enabling completion or refinement of the learned scene representation at a surface control station (300); and (f) an integration controller (180) configured to: (f-i) receive a synchronisation quality metric from the timing and synchronisation subsystem (130); (f-ii) receive a thermal state measurement from the conductive thermal path (150) or the embedded processing unit (140); (f-iii) modify one or more operating parameters of the embedded processing unit in response to the thermal state measurement exceeding a first threshold, wherein the modification comprises at least one of reducing image resolution, reducing frame capture rate, reducing a learned representation update rate, or reducing scene model complexity; and (f-iv) annotate the learned scene representation or the compressed representation thereof with a validity indicator based on at least the synchronisation quality metric.

2. The module of claim 1, wherein the at least one image sensor (120) comprises a single image sensor such that the module is a single-camera subsea compute camera.

3. The module of claim 1, wherein the conductive thermal path comprises a thermal interface plate (152) mechanically pressed between the embedded processing unit (140) and amachined interior surface of the housing wall (112), with a compliant thermal interface material (154) accommodating thermal expansion differentials.

4. The module of claim 1, wherein the conductive thermal path comprises one or more heat pipes connecting the embedded processing unit (140) to a thermal dissipation zone on the housing where exterior surface features are provided to maximise convective heat transfer to surrounding seawater.

5. The module of claim 1, wherein the integration controller (180) is configured to operate the embedded processing unit (140) in at least three thermal modes comprising: a full-power mode in which learned scene representation generation operates at a first fidelity; a reduced-power mode activated when the thermal state measurement exceeds the first threshold; and a pass-through mode activated when the thermal state measurement exceeds a second threshold higher than the first threshold, in which on-board learned scene representation generation is suspended and the module archives raw image data while transmitting only status metadata and / or a compressed video stream.

6. The module of claim 1, wherein the timing and synchronisation subsystem (130) comprises a local master oscillator (132) within the module, the local master oscillator being periodically disciplined by an external precision time protocol grandmaster and operating autonomously in holdover mode between discipline events.

7. The module of claim 6, wherein the timing and synchronisation subsystem further compensates for tether latency variation by measuring round-trip communication delay to a surface reference and applying a path-asymmetry correction to account for asymmetric propagation delays.

8. The module of claim 1, further comprising a housing deformation monitoring subsystem configured to detect deformation of the pressure-rated housing (110) under hydrostatic loading, the deformation monitoring subsystem comprising at least one of: (a) strain sensors bonded to the housing interior; (b) fiducial markers at known positions within the housing visible to the at least one image sensor (120); or (c) a pressure-depth lookup table correlating external hydrostatic pressure with predicted housing deformation from factory calibration data; wherein the integration controller (180) flags or invalidates learned scene representation data when detected or predicted deformation exceeds a predetermined geometric tolerance.

9. The module of claim 1, wherein the compressed representation comprises a serialised learned scene representation.

10. The module of claim 9, wherein the learned scene representation comprises a set of Gaussian primitives each defined by a position, covariance parameters, and radiometric parameters including at least colour and opacity.

11. The module of claim 10, wherein the learned scene representation further comprises a time-indexed sequence of updates to the Gaussian primitives, the updates being associated with capture timestamps to represent scene change over time.

12. The module of claim 1, further comprising: (i) a video encoder configured to encode a high-resolution video stream from the raw image data; and (ii) a second communication channel (165) configured to transmit the encoded high-resolution video stream to the surface control station (300) independently of transmission of the compressed representation of the learned scene representation.

13. The module of claim 12, wherein the high-resolution video stream is encoded at the full native resolution of the image sensor (120) using an inter-frame video codec.

14. The module of claim 12, wherein the second communication channel (165) comprises a dedicated fibre-optic link.

15. The module of claim 1, further comprising one or more wet-mate connectors (170) providing power and communication connections, the connectors being of a type suitable for subsea mating and de-mating.

16. The module of claim 15, further comprising an auto-configuration subsystem (174) that, upon connection to a host vehicle, detects module type and capabilities, loads processing parameters, and initiates a sensor health check and timing verification sequence.

17. The module of claim 1, wherein the validity indicator encodes at least one of: a synchronisation uncertainty value representing time base uncertainty relative to a reference clock; and a thermal confidence state indicating a current thermal operating mode of the integration controller (180).

18. A method of subsea imaging and processing, the method comprising: (a) maintaining a time base for at least one image sensor (120) within a pressure-rated subsea housing (110) using a precision time protocol with temperature-dependent drift correction based on internal temperature measurement; (b) capturing image data from the at least one image sensor according to scheduled exposures based on the time base; (c) generating a learned scene representation comprising a three-dimensional scene model and optionally a time-varying sequence of scene model updates from the image data using an embedded processing unit (140) disposed within the housing and thermally coupled to the housing via a conductive thermal path (150) for heat rejection to surrounding seawater; (d) archiving full-resolutionraw image data to a local storage medium (166) within the housing; (e) transmitting a compressed representation of the learned scene representation and / or intermediate reconstruction data over a bandwidth-constrained communication link (400); (f) monitoring a thermal state of the embedded processing unit; (g) modifying one or more operating parameters including at least one of image resolution, frame capture rate, learned representation update rate, and scene model complexity in response to the thermal state exceeding a first threshold; and (h) annotating the learned scene representation or the compressed representation thereof with a validity indicator based on a synchronisation quality metric derived from the precision time protocol.

19. The method of claim 18, further comprising compensating for tether latency variation by operating a local master oscillator (132) within the housing, periodically disciplining the local master oscillator against an external time reference, and applying a path-asymmetry correction to account for asymmetric propagation delays.

20. The method of claim 18, further comprising encoding a high-resolution video stream from the raw image data on-board and transmitting the high-resolution video stream over a second communication channel (165) distinct from transmission of the compressed representation.

21. The method of claim 20, wherein the high-resolution video stream is encoded at the full native resolution of the image sensor (120) and transmitted over a dedicated fibre-optic link.

22. The method of claim 18, further comprising embedding navigation uncertainty metadata in the compressed representation for geo-referencing of the learned scene representation.

23. The method of claim 18, wherein the compressed representation occupies less than onetenth of the bandwidth that would be required to transmit the raw image data.

24. The method of claim 18, further comprising monitoring housing deformation and flagging learned scene representation data captured when detected deformation exceeds a predetermined geometric tolerance.

25. The method of claim 18, further comprising suspending generation of the learned scene representation when the synchronisation quality metric falls below a critical operational threshold.

26. The method of claim 18, wherein the learned scene representation comprises a set of Gaussian primitives produced through an optimisation or inference process from the image data.

27. A subsea imaging system comprising: (a) a modular imaging and processing module as claimed in claim 1; (b) a host vehicle interface (200) providing power and communication tothe modular imaging and processing module via wet-mate connectors (170); (c) a first communication link (400) between the modular imaging and processing module and a surface control station (300), the first communication link being bandwidth-constrained relative to raw image data; and (d) a surface control station (300) configured to receive and render the compressed representation of the learned scene representation for in-mission visualisation and dimensional measurement.

28. The system of claim 27, further comprising a second communication link (165) configured to receive an encoded high-resolution video stream from the modular imaging and processing module.

29. The system of claim 28, wherein the second communication link comprises a dedicated fibre-optic link.

30. The system of claim 27, wherein the host vehicle interface (200) further receives navigation data from a navigation subsystem of the host vehicle and provides said navigation data to the modular imaging and processing module for geo-referencing of the learned scene representation.

31. The system of claim 27, wherein the surface control station (300) is further configured to perform temporal differencing between learned scene representations captured at different times for subsea asset integrity assessment.

32. The system of claim 27, wherein the modular imaging and processing module is field-replaceable subsea by an intervention vehicle without shutdown of the host vehicle.

33. The system of claim 27, wherein the host vehicle is selected from a remotely operated vehicle, an autonomous underwater vehicle, a crawling inspection platform, and a fixed subsea monitoring station.

34. The system of claim 27, wherein the modular imaging and processing module is configured to maintain autonomous timing and local data archiving during degradation or loss of the first communication link (400).

35. A non-transitory computer-readable medium storing instructions that, when executed by an embedded processing unit (140) disposed within a pressure-rated subsea housing (110) and thermally coupled thereto for heat rejection to surrounding seawater, cause the embedded processing unit to: (a) receive image data from at least one image sensor (120) timed according to a precision time protocol with temperature-dependent drift correction; (b) generate a learned scene representation comprising a three-dimensional scene model and optionally a time-varying sequence of scene model updates from the image data; (c) store raw image data at full resolution to a local archive within the housing; (d) transmit a compressedrepresentation of the learned scene representation and / or intermediate reconstruction data over a bandwidth-constrained communication link (400); (e) modify one or more operating parameters including at least one of image resolution, frame capture rate, learned representation update rate, and scene model complexity in response to a thermal state measurement exceeding a first threshold; and (f) annotate the learned scene representation or the compressed representation thereof with a validity indicator based on a synchronisation quality metric.

36. The medium of claim 35, wherein the instructions further cause the embedded processing unit to suspend generation of the learned scene representation when the synchronisation quality metric falls below a critical operational threshold.

37. The medium of claim 35, wherein the instructions further cause the embedded processing unit to embed navigation uncertainty metadata in the compressed representation for georeferencing.

38. The medium of claim 35, wherein the instructions further cause the embedded processing unit to encode a high-resolution video stream from the raw image data and transmit the encoded high-resolution video stream over a second communication channel (165) distinct from transmission of the compressed representation.Amendments to the Claims have been filed as followsAMENDED CLAIMS 1 TO 3803 08 261. A modular subsea imaging and processing module comprising: (a) a pressure-rated housing (110) configured for subsea operation, the housing maintaining approximately one atmosphere internal pressure and configured to resist external hydrostatic pressure at operational depth; (b) at least one image sensor (120) arranged within the housing to capture images through an optical window; (c) a timing and synchronisation subsystem (130) implementing a precision time protocol to maintain a time base for the at least one image sensor (120) and to initiate exposure of the at least one image sensor at scheduled times, the timing and synchronisation subsystem comprising temperature-dependent drift correction based on internal temperature measurement; (d) an embedded processing unit (140) disposed within the housing and thermally coupled to the housing via a conductive thermal path (150) configured to reject heat to surrounding seawater, the embedded processing unit configured to generate a learned scene representation comprising a three-dimensional scene model from image data captured by the at least one image sensor; (e) a dual-path data output (160) comprising: (i) a local archive storing raw image data at full resolution; and (ii) a transmission path (164) configured to transmit, over a communication link (400) whose data capacity is lower than the data rate required to transmit the raw image data, a compressed representation of the learned scene representation and / or intermediate reconstruction data enabling completion or refinement of the learned scene representation at a surface control station (300); and (f) an integration controller (180) configured to: (f-i) receive a synchronisation quality metric from the timing and synchronisation subsystem (130); (f-ii) receive a thermal state measurement from the conductive thermal path (150) or the embedded processing unit (140); (f-iii) modify one or more operating parameters of the embedded processing unit in response to the thermal state measurement exceeding a first threshold, wherein the modification comprises at least one of reducing image resolution, reducing frame capture rate, reducing a learned representation update rate, or reducing scene model complexity; and (f-iv) annotate the learned scene representation or the compressed03 08 26representation thereof with a validity indicator based on at least the synchronisation quality metric.

2. The module of claim 1, wherein the at least one image sensor (120) comprises a single image sensor such that the module is a single-camera subsea compute camera.

3. The module of claim 1, wherein the conductive thermal path comprises a thermal interface plate (152) mechanically pressed between the embedded processing unit (140) and a machined interior surface of the housing wall (112), with a compliant thermal interface material (154) accommodating thermal expansion differentials.

4. The module of claim 1, wherein the conductive thermal path comprises one or more heat pipes connecting the embedded processing unit (140) to a thermal dissipation zone on the housing where exterior surface features are provided to maximise convective heat transfer to surrounding seawater.

5. The module of claim 1, wherein the integration controller (180) is configured to operate the embedded processing unit (140) in at least three thermal modes comprising: a full-power mode in which learned scene representation generation operates at a first fidelity; a reduced-power mode activated when the thermal state measurement exceeds the first threshold; and a pass-through mode activated when the thermal state measurement exceeds a second threshold higher than the first threshold, in which on-board learned scene representation generation is suspended and the module archives raw image data while transmitting only status metadata and / or a compressed video stream.

6. The module of claim 1, wherein the timing and synchronisation subsystem (130) comprises a local master oscillator (132) within the module, the local master oscillator being periodically disciplined by an external precision time protocol grandmaster and operating autonomously in holdover mode between discipline events.

7. The module of claim 6, wherein the timing and synchronisation subsystem further compensates for tether latency variation by measuring round-trip communication delay to a surface reference and applying a path-asymmetry correction to account for asymmetric propagation delays.

8. The module of claim 1, further comprising a housing deformation monitoring subsystem configured to detect deformation of the pressure-rated housing (110) under hydrostatic loading, the deformation monitoring subsystem comprising at least one of: (a) strain sensors03 08 26bonded to the housing interior; (b) fiducial markers at known positions within the housing visible to the at least one image sensor (120); or (c) a pressure-depth lookup table correlating external hydrostatic pressure with predicted housing deformation from factory calibration data; wherein the integration controller (180) flags or invalidates learned scene representation data when detected or predicted deformation exceeds a predetermined geometric tolerance.

9. The module of claim 1, wherein the compressed representation comprises a serialised learned scene representation.

10. The module of claim 9, wherein the learned scene representation comprises a set of Gaussian primitives each defined by a position, covariance parameters, and radiometric parameters including at least colour and opacity.

11. The module of claim 10, wherein the learned scene representation further comprises a time-indexed sequence of updates to the Gaussian primitives, the updates being associated with capture timestamps to represent scene change over time.

12. The module of claim 1, further comprising: (i) a video encoder configured to encode a high-resolution video stream from the raw image data; and (ii) a second communication channel (165) configured to transmit the encoded high-resolution video stream to the surface control station (300) independently of transmission of the compressed representation of the learned scene representation.

13. The module of claim 12, wherein the high-resolution video stream is encoded at the full native resolution of the image sensor (120) using an inter-frame video codec.

14. The module of claim 12, wherein the second communication channel (165) comprises a dedicated fibre-optic link.

15. The module of claim 1, further comprising one or more wet-mate connectors (170) providing power and communication connections, the connectors being of a type suitable for subsea mating and de-mating.

16. The module of claim 15, further comprising an auto-configuration subsystem (174) that, upon connection to a host vehicle, detects module type and capabilities, loads processing parameters, and initiates a sensor health check and timing verification sequence.

17. The module of claim 1, wherein the validity indicator encodes at least one of a synchronisation uncertainty value representing time base uncertainty relative to a reference03 08 26clock; and a thermal confidence state indicating a current thermal operating mode of the integration controller (180).

18. A method of subsea imaging and processing, the method comprising: (a) maintaining a time base for at least one image sensor (120) within a pressure-rated subsea housing (110) using a precision time protocol with temperature-dependent drift correction based on internal temperature measurement; (b) capturing image data from the at least one image sensor according to scheduled exposures based on the time base; (c) generating a learned scene representation comprising a three-dimensional scene model from the image data using an embedded processing unit (140) disposed within the housing and thermally coupled to the housing via a conductive thermal path (150) for heat rejection to surrounding seawater; (d) archiving full-resolution raw image data to a local storage medium (166) within the housing; (e) transmitting a compressed representation of the learned scene representation and / or intermediate reconstruction data over a bandwidth-constrained communication link (400); (f) monitoring a thermal state of the embedded processing unit; (g) modifying one or more operating parameters including at least one of image resolution, frame capture rate, learned representation update rate, and scene model complexity in response to the thermal state exceeding a first threshold; and (h) annotating the learned scene representation or the compressed representation thereof with a validity indicator based on a synchronisation quality metric derived from the precision time protocol.

19. The method of claim 18, further comprising compensating for tether latency variation by operating a local master oscillator (132) within the housing, periodically disciplining the local master oscillator against an external time reference, and applying a path-asymmetry correction to account for asymmetric propagation delays.

20. The method of claim 18, further comprising encoding a high-resolution video stream from the raw image data on-board and transmitting the high-resolution video stream over a second communication channel (165) distinct from transmission of the compressed representation.

21. The method of claim 20, wherein the high-resolution video stream is encoded at the full native resolution of the image sensor (120) and transmitted over a dedicated fibre-optic link.

22. The method of claim 18, further comprising embedding navigation uncertainty metadata in the compressed representation for geo-referencing of the learned scene representation.03 08 2623. The method of claim 18, wherein the compressed representation occupies less than onetenth of the bandwidth that would be required to transmit the raw image data.

24. The method of claim 18, further comprising monitoring housing deformation and flagging learned scene representation data captured when detected deformation exceeds a predetermined geometric tolerance.

25. The method of claim 18, further comprising suspending generation of the learned scene representation when the synchronisation quality metric falls below a critical operational threshold.

26. The method of claim 18, wherein the learned scene representation comprises a set of Gaussian primitives produced through an optimisation or inference process from the image data.

27. A subsea imaging system comprising: (a) a modular imaging and processing module as claimed in claim 1; (b) a host vehicle interface (200) providing power and communication to the modular imaging and processing module via wet-mate connectors (170); (c) a first communication link (400) between the modular imaging and processing module and a surface control station (300), the first communication link being bandwidth-constrained relative to raw image data; and (d) a surface control station (300) configured to receive and render the compressed representation of the learned scene representation for in-mission visualisation and dimensional measurement.

28. The system of claim 27, further comprising a second communication link (165) configured to receive an encoded high-resolution video stream from the modular imaging and processing module.

29. The system of claim 28, wherein the second communication link comprises a dedicated fibre-optic link.

30. The system of claim 27, wherein the host vehicle interface (200) further receives navigation data from a navigation subsystem of the host vehicle and provides said navigation data to the modular imaging and processing module for geo-referencing of the learned scene representation.03 08 2631. The system of claim 27, wherein the surface control station (300) is further configured to perform temporal differencing between learned scene representations captured at different times for subsea asset integrity assessment.

32. The system of claim 27, wherein the modular imaging and processing module is field-replaceable subsea by an intervention vehicle without shutdown of the host vehicle.

33. The system of claim 27, wherein the host vehicle is selected from a remotely operated vehicle, an autonomous underwater vehicle, a crawling inspection platform, and a fixed subsea monitoring station.

34. The system of claim 27, wherein the modular imaging and processing module is configured to maintain autonomous timing and local data archiving during degradation or loss of the first communication link (400).

35. A non-transitory computer-readable medium storing instructions that, when executed by an embedded processing unit (140) disposed within a pressure-rated subsea housing (110) and thermally coupled thereto for heat rejection to surrounding seawater, cause the embedded processing unit to: (a) receive image data from at least one image sensor (120) timed according to a precision time protocol with temperature-dependent drift correction; (b) generate a learned scene representation comprising a three-dimensional scene model from the image data; (c) store raw image data at full resolution to a local archive within the housing; (d) transmit a compressed representation of the learned scene representation and / or intermediate reconstruction data over a bandwidth-constrained communication link (400); (e) modify one or more operating parameters including at least one of image resolution, frame capture rate, learned representation update rate, and scene model complexity in response to a thermal state measurement exceeding a first threshold; and (f) annotate the learned scene representation or the compressed representation thereof with a validity indicator based on a synchronisation quality metric.

36. The medium of claim 35, wherein the instructions further cause the embedded processing unit to suspend generation of the learned scene representation when the synchronisation quality metric falls below a critical operational threshold.

37. The medium of claim 35, wherein the instructions further cause the embedded processing unit to embed navigation uncertainty metadata in the compressed representation for georeferencing.

38. The medium of claim 35, wherein the instructions further cause the embedded processing unit to encode a high-resolution video stream from the raw image data and transmit the encoded high-resolution video stream over a second communication channel (165) distinct from transmission of the compressed representation.03 08 26T +44(0)30 0300 2000A

Images