High-speed biomimetic intelligent imaging device and imaging method based on event camera

Abstract

This invention discloses a high-speed biomimetic intelligent imaging device and method based on an event camera. The device includes an event camera module for continuous imaging of a cable surface during high-speed movement, capturing brightness change events caused by normal and defective areas of the cable, and generating unstructured event stream data; an FPGA data preprocessing module connected to the event camera module for converting the unstructured event stream into a three-dimensional event tensor; a GPU imaging processing module connected to the FPGA data preprocessing module for generating cable biomimetic imaging data; a high-speed data bus and memory module connected to the event camera module, FPGA data preprocessing module, and GPU imaging processing module, respectively; and an auxiliary interface module connected to the GPU imaging processing module for providing visualization, remote data transmission, and device adaptation interfaces. This invention enables high-quality, low-latency, and intelligent biomimetic imaging of minute defects on the cable surface during high-speed movement.

Description

Technical Field

[0001] This invention relates to the field of cable imaging and event camera application technology, specifically to a high-speed bionic intelligent imaging device and imaging method based on an event camera. Background Technology

[0002] As a core infrastructure, the surface condition of cables directly determines the stability of system operation. Micro-cracks, wear, corrosion, and other minute defects can easily lead to safety hazards, and high-quality imaging in high-speed inspection scenarios is a prerequisite for accurate defect identification. Traditional cable imaging solutions mostly rely on frame exposure cameras, which have obvious technical limitations: limited by a fixed frame rate, they are prone to motion blur in scenarios such as online inspection on production lines, making it impossible to capture micron-level minute defects; the dynamic range is narrow, resulting in poor imaging quality in complex environments such as strong light, shadow, and nighttime, with blurred boundaries between defects and the background; and the generated frame image data is huge, increasing the pressure on transmission and storage, making it difficult to meet the real-time requirements of high-speed inspection.

[0003] Event cameras, as a novel biomimetic vision sensor, possess advantages such as microsecond-level temporal resolution, high dynamic range, low latency, and low data volume, effectively overcoming the bottlenecks of traditional frame-exposure cameras and solving imaging challenges caused by motion blur and changes in illumination. However, event cameras output unstructured event stream data, which is difficult to directly use for subsequent processing of cable defect imaging. Furthermore, existing imaging solutions lack specific optimizations for cable scenarios, resulting in problems such as insufficient extraction of event stream features, weak anti-interference capabilities in complex environments, and poor adaptability to biomimetic imaging, making it difficult to meet the imaging requirements of high-speed cable inspection.

[0004] Therefore, there is an urgent need to propose an innovative imaging device and method that can fully leverage the technical advantages of event cameras and perform in-depth optimization for the special needs of cable defect imaging, so as to achieve high-quality, low-latency, and intelligent biomimetic imaging of minute defects on the cable surface under high-speed movement. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a high-speed bionic intelligent imaging device and imaging method based on an event camera, which can realize high-quality, low-latency, and intelligent bionic imaging of minute defects on the surface of cables under high-speed movement.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows.

[0007] A high-speed biomimetic intelligent imaging device based on an event camera includes: The event camera module is used to continuously image the cable surface during high-speed movement, capture brightness change events caused by normal and defective areas of the cable, and generate unstructured event stream data. An FPGA data preprocessing module, connected to the event camera module, is used to process the event stream data in real time, performing adaptive noise reduction, spatial coordinate mapping, polarity channel separation and time dimension aggregation operations, and converting the unstructured event stream into a three-dimensional event tensor. The GPU imaging processing module is connected to the FPGA data preprocessing module and is used to perform a biomimetic feature enhancement algorithm on the three-dimensional event tensor. Through multi-scale feature extraction, dynamic receptive field adjustment and attention feature focusing, it generates cable biomimetic imaging data. A high-speed data bus and memory module includes a high-speed data bus and a memory; the high-speed data bus is connected to the event camera module, the FPGA data preprocessing module, the GPU imaging processing module, and the memory respectively, for realizing high-speed data transmission between the modules; the memory is used to provide persistent storage of data and real-time computing cache. An auxiliary interface module, connected to the GPU imaging processing module and the memory, is used to provide visualization display, remote data transmission and device adaptation interfaces.

[0008] Preferably, the event camera module includes at least three event cameras evenly arranged along the circumference of the cable. Each event camera has a pixel resolution of 1280×720, a minimum time window of 50 microseconds, a microsecond-level time resolution, and a dynamic range of ≥120dB. The event trigger threshold can be adaptively adjusted according to the ambient light intensity to capture brightness change events caused by minute defects with a feature size ≥1.2mm on the cable surface without motion blur at a detection speed of 30-100km / h.

[0009] Preferably, the FPGA data preprocessing module is built on a high-performance FPGA chip; the GPU imaging processing module uses a GPU that supports half-precision computing mode and has a built-in 3-layer bionic vision perception layer.

[0010] Preferably, the high-speed data bus adopts the PCIe 4.0 transmission protocol; the memory includes a solid-state drive for persistent full data storage and DDR memory for real-time computing caching; the auxiliary interface module includes an HDMI visualization interface for local display, a 4G / 5G remote transmission interface for wireless data transmission, and a standardized device interface for connecting external devices.

[0011] A high-speed biomimetic intelligent imaging method based on an event camera, applied to a high-speed biomimetic intelligent imaging device based on an event camera, includes the following steps: S1. Imaging parameter initialization configuration: Based on the speed and lighting conditions of the cable detection scene, preset and optimize the core imaging parameters; S2. High-speed acquisition of cable event stream data: The event camera modules in an array layout synchronously acquire brightness change events on the cable surface and generate unstructured event stream data; S3. Event Stream Preprocessing and Structured Transformation: The acquired event stream data is adaptively denoised using the FPGA data preprocessing module and converted into a three-dimensional event tensor with fixed spatial dimensions and polarity channels; S4. Biomimetic Feature Enhancement Imaging: Multi-scale feature extraction of the three-dimensional event tensor is performed through the GPU imaging processing module, and the processing focus is dynamically adjusted in combination with the event activity. The imaging features of the defect area are enhanced through the attention mechanism to generate cable biomimetic imaging data. S5. Imaging data standardization output: Convert the enhanced cable simulacrum imaging data into a standardized format and output it through multiple paths.

[0012] Preferably, step S1 specifically includes: Based on a preset detection speed range, a time window Δt and a cumulative number of frames are set for event aggregation processing; wherein, when the detection speed is greater than 80 km / h, the time window Δt is set to 50 microseconds; when the detection speed is less than or equal to 80 km / h, the time window Δt is set to 100 microseconds; the cumulative number of frames is fixed at 100 frames. Based on real-time monitored ambient light intensity, the event trigger threshold of the event camera is adaptively adjusted; specifically, in strong light conditions, the threshold is increased by 10%-20%; in nighttime conditions, the threshold is decreased by 15%-25%. Enable half-precision computing mode in the GPU imaging processing module, and preset the convolution kernel size and receptive field adjustment range of the bionic vision perception layer.

[0013] Preferably, the unstructured event stream data generated in step S2 is the original event stream E, represented as: E={(x, y, p, t)} Where (x, y) represents the pixel coordinates where the event occurs, x is the horizontal coordinate and y is the vertical coordinate; p is the brightness polarity, p∈ {+1, -1}, +1 represents a pixel brightness increase event, i.e., a positive event, and -1 represents a pixel brightness decrease event, i.e., a negative event; t is the timestamp of the event, with a precision in microseconds.

[0014] Preferably, step S3 specifically includes: S31. Adaptive Noise Reduction: Based on the spatiotemporal distribution characteristics of events, isolated noise events and high-frequency interference events are removed from the original event stream E to obtain a valid event set E_valid related to the cable surface condition, where... The adaptive noise reduction process ensures an effective event retention rate of ≥95%. S32. Spatial coordinate mapping: Map the pixel coordinates (x, y) of each event in the valid event set E_valid to a pixel grid of size H×W, where H is the image height, W is the image width, and H = 1280, W = 720; S33. Polarity-based channel splitting: The valid event set E_valid is split into a positive event set E+_valid and a negative event set E-_valid according to the brightness polarity p, where: E+_valid = {(x, y, t) | (x, y, p, t) ∈ E_valid, p = +1} E-_valid = {(x, y, t) | (x, y, p, t) ∈ E_valid, p = -1}; S34. Time Dimension Aggregation: After mapping in step S32, events that appear at the same pixel coordinate (y, x) within the preset time window Δt in step S1 are aggregated to generate a three-dimensional event tensor T; For a positive event channel, its tensor value T(y, x, 1) = Agg(E+_valid(y, x, t)); For a negative event channel, its tensor value is T(y, x, 2) = Agg(E-_valid(y, x, t)); Where Agg(·) is a time aggregation function that uses event frequency statistics, that is, it counts the number of times an event occurs on the pixel within a time window Δt. The final result is a three-dimensional event tensor T with dimensions (H, W, 2), where the two channels of the third dimension correspond to positive and negative events, respectively.

[0015] Preferably, S4 specifically includes: S41. Multi-scale feature extraction: The three-dimensional event tensor T is convolved using multiple two-dimensional convolution kernels of different sizes to extract defect features at different scales; this process is represented as:

[0016] in, This represents a two-dimensional convolution operation; K_s represents the convolution kernel, with size s ∈ {3, 5, 7}, corresponding to 3×3, 5×5, and 7×7 convolution kernels respectively; F_s is the feature map of the corresponding scale obtained after convolution; S42. Dynamic Receptive Field Adjustment: Calculate the event activity D(x, y) for each pixel, which serves as the basis for adjusting the convolutional receptive field; the event activity is obtained by adding the feature values ​​of the positive and negative event channels. D(x, y) = T(y, x, 1)+T(y, x, 2) Where T(y, x, 1) and T(y, x, 2) are the values ​​of the positive and negative event channels of the three-dimensional event tensor T at position (y, x), respectively; the receptive field range of the convolution kernel is adaptively adjusted according to the value of D(x, y): the receptive field is expanded for defect regions with high D(x, y) values ​​to enhance features, and the receptive field is shrunk for normal regions with low D(x, y) values ​​to reduce interference; S43. Attention Feature Focusing: Based on the event activity D calculated in step S42, an attention weight map A is generated using the Sigmoid activation function, where the weight A(x, y) of each pixel is proportional to σ(D(x, y)), and σ is the Sigmoid function; the multi-scale feature map F_s is multiplied element-wise with the attention weight map A to achieve feature focusing. F_enhanced = F_s⊙A Where ⊙ represents element-wise product; F_enhanced is the feature map enhanced by the attention mechanism, that is, the enhanced cable simulation imaging data.

[0017] Preferably, in step S5, the multi-path output specifically includes: Standardized cable simulation imaging data is output through the high-speed data bus, wherein the entire process delay from event acquisition to imaging data output is ≤3ms; The original event stream data, the preprocessed 3D event tensor T, and the final generated cable simulac image data are synchronously stored in the solid-state drive in the memory. The HDMI interface in the auxiliary interface module displays the cable bionic enhanced imaging image and defect area markings in real time, and the imaging data is uploaded via the 4G / 5G remote transmission interface.

[0018] Due to the adoption of the above technical solutions, the technical progress achieved by this invention is as follows.

[0019] This invention completely eliminates motion blur and is suitable for high-speed detection: by utilizing the microsecond-level response of the event camera and the processing architecture of this invention, clear defect images can still be obtained at high speeds of 30-100 km / h, effectively overcoming the fundamental problem of traditional frame cameras.

[0020] This invention has excellent environmental adaptability: the high dynamic range (≥120dB) of the event camera itself, combined with adaptive threshold adjustment and the noise reduction and feature enhancement mechanism in the algorithm of this invention, enables the system to work stably under complex lighting conditions such as strong light, weak light, and backlight, and significantly improves the signal-to-noise ratio (≥35dB).

[0021] This invention offers high data efficiency and strong real-time performance: processing only changing brightness information, it reduces data volume by more than 90% compared to traditional video streams under the same information capture conditions, greatly alleviating transmission and storage pressure. Through FPGA real-time preprocessing and GPU-accelerated biomimetic algorithms, it achieves an extremely low end-to-end latency of ≤3ms from acquisition to image output, meeting the real-time requirements of the highest speed detection.

[0022] This invention offers high imaging quality and strong defect detection capabilities: the innovative biomimetic feature enhancement algorithm (multi-scale convolution, dynamic receptive field, attention mechanism) can fully exploit the spatiotemporal and polarity features in the event stream, effectively enhancing the contrast between micro-defects (≥1.2mm) and the background, and significantly improving the imaging clarity and recognizability of defects such as micro-cracks, corrosion, and bulges.

[0023] The system of this invention is flexible and easy to deploy: its modular design and rich auxiliary interfaces enable it to be easily integrated into various platforms such as fixed testing stations, making it suitable for application scenarios in multiple fields such as power, transportation, and industry. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the device principle of the present invention; Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0025] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0026] A high-speed biomimetic intelligent imaging device based on an event camera, employing a modular architecture design, combined with... Figure 1 As shown, it includes an event camera module, an FPGA data preprocessing module, a GPU imaging processing module, a high-speed data bus and memory module, and an auxiliary interface module.

[0027] The event camera module is used for continuous imaging of the cable surface during high-speed movement, capturing brightness change events caused by normal and defective areas of the cable, and generating unstructured event stream data. Specifically, the event camera module includes at least three DVSlume event cameras evenly arranged along the circumference of the cable. Each event camera has a pixel resolution of 1280×720, adapting to the installation requirements of various inspection equipment such as production line inspection stations, and can achieve full-surface, blind-spot-free imaging of the cable. The event camera module has microsecond-level temporal resolution, with a minimum time window of 50 microseconds, and a high dynamic range of ≥120dB. It can capture brightness change events caused by minute defects (micro-cracks, wear, corrosion, bulges, etc.) with feature sizes ≥1.2mm on the cable surface without motion blur at inspection speeds of 30-100km / h. The event camera module supports adaptive adjustment of the event trigger threshold according to the ambient light intensity, resisting the influence of complex environments such as strong light, nighttime, and dust interference, avoiding motion blur and invalid imaging, and ensuring the integrity and accuracy of defect event capture.

[0028] The FPGA data preprocessing module, connected to the event camera module, performs real-time processing of the event stream data. It executes adaptive noise reduction, spatial coordinate mapping, polarity channel separation, and temporal dimension aggregation operations, converting the unstructured event stream into a (1280×720×2) three-dimensional event tensor, ensuring an effective event retention rate of ≥95%, thus providing suitable data for subsequent simulation imaging. Specifically, the FPGA data preprocessing module is built on a high-performance FPGA chip.

[0029] The GPU imaging processing module, connected to the FPGA data preprocessing module, performs biomimetic feature enhancement algorithms on the 3D event tensor. Through multi-scale feature extraction, dynamic receptive field adjustment, and attention feature focusing, it generates biomimetic imaging data of the cable. Specifically, the GPU imaging processing module uses an NVIDIA GeForce RTX 3090 GPU with a built-in three-layer biomimetic vision perception layer. It simulates the multi-scale perception mechanism of biological vision, employing 3×3, 5×5, and 7×7 multi-scale convolutional kernels and a dynamic receptive field adjustment algorithm to extract and enhance deep features from the event tensor. Targeting the event density characteristics of cable defect areas, it focuses on high-event-density regions, enhancing the imaging contrast between defects and the background, accurately adapting to the imaging needs of defects of different sizes, such as micron-level cracks and millimeter-level bulges. Simultaneously, it enables a half-precision calculation mode, significantly reducing computational latency while ensuring image quality, ensuring a total latency of ≤3ms for the entire biomimetic imaging process, suitable for high-speed inspection scenarios.

[0030] The high-speed data bus and memory module includes a high-speed data bus and memory. The high-speed data bus connects to the event camera module, FPGA data preprocessing module, GPU imaging processing module, and memory. The high-speed data bus uses the PCIe 4.0 transmission protocol to achieve high-speed data transmission between modules, and the transmission rate meets the high-speed real-time flow requirements of event stream data, preprocessed data, and enhanced imaging data, avoiding data transmission bottlenecks that could affect the efficiency of emulation imaging. The memory provides persistent data storage and real-time computation caching. Specifically, the memory includes a 1TB SSD solid-state drive and 32GB DDR5 DDR memory. The solid-state drive is used for persistent full data storage; the DDR memory is used for real-time computation caching.

[0031] The auxiliary interface module, connected to the GPU imaging processing module and memory, provides interfaces for visualization, remote data transmission, and device adaptation. Specifically, the auxiliary interface module includes an HDMI visualization interface, a 4G / 5G remote transmission interface, and a standardized device interface. The HDMI visualization interface is used for local display; the 4G / 5G remote transmission interface is used for wireless data transmission; and the standardized device interface is used to connect external devices, enabling rapid adaptation to different testing equipment such as production line testing stations, thus improving hardware compatibility and multi-scenario adaptability.

[0032] A high-speed biomimetic intelligent imaging method based on an event camera, such as Figure 2 As shown, it includes the following steps: S1. Imaging Parameter Initialization Configuration: Based on the speed and lighting conditions of the cable detection scene, preset and optimize the core imaging parameters. Details are as follows: Based on the preset detection speed range, the time window Δt and the cumulative number of frames used for event aggregation processing are set; specifically, in high-speed detection scenarios (detection speed greater than 80 km / h), the time window Δt is set to 50 microseconds; in medium-low speed scenarios (detection speed 30-80 km / h), the time window Δt is set to 100 microseconds. The cumulative number of frames is fixed at 100 frames.

[0033] Based on the real-time detected ambient light intensity, the event trigger threshold of the event camera is adaptively adjusted; in strong light environments, the threshold is increased by 10%-20%; in nighttime environments, the threshold is decreased by 15%-25%.

[0034] Enable half-precision computing mode in the GPU imaging processing module, and preset the convolution kernel size and receptive field adjustment range of the bionic vision perception layer.

[0035] S2. High-speed acquisition of cable event stream data: Start the detection equipment and imaging device, and synchronously acquire the brightness change events on the cable surface through the array-arranged event camera modules to generate unstructured event stream data.

[0036] In this step, continuous imaging of the cable surface is performed simultaneously to capture brightness change events caused by normal and defective areas (micro-cracks, wear, corrosion, bulges, broken strands, sheath peeling, etc.), generating unstructured event stream data containing timestamps, pixel coordinates, and brightness polarity (rising / falling). Specifically, the generated unstructured event stream data is the original event stream E, represented as: E={(x, y, p, t)} Where (x, y) represents the pixel coordinates where the event occurs, x is the horizontal coordinate and y is the vertical coordinate, with values ​​corresponding to the camera resolution (0≤x≤W) and (0≤y≤H), W = 720 and H = 1280, which can be adjusted to the target size through ROI cropping; p is the brightness polarity, p ∈ {+1, -1}, +1 represents a pixel brightness increase event, i.e., a positive event, and -1 represents a pixel brightness decrease event, i.e., a negative event; t is the timestamp of the event, with a precision of microseconds, which can accurately characterize the temporal features of brightness changes.

[0037] S3. Event Stream Preprocessing and Structured Transformation: The acquired event stream data is adaptively denoised using the FPGA data preprocessing module and converted into a three-dimensional event tensor with fixed spatial dimensions and polarity channels. Details are as follows: S31. Adaptive Noise Reduction: Based on the spatiotemporal distribution characteristics of events, isolated noise events and high-frequency interference events are removed from the original event stream E to obtain a valid event set E_valid related to the cable surface condition, where... The adaptive noise reduction process ensures that the effective event retention rate is ≥95%, and the effective events must meet the conditions that adjacent pixels have synchronous responses and their timestamps are continuous.

[0038] S32. Spatial Coordinate Mapping: Map the pixel coordinates (x, y) of each event in the valid event set E_valid to a pixel grid of size H×W, where H is the image height, W is the image width, and H = 1280, W = 720. In this step, ROI cropping can be performed to optimize the data volume, ultimately matching the event camera resolution.

[0039] S33. Polarity-based channel splitting: The valid event set E_valid is split into a positive event set E+_valid and a negative event set E-_valid according to the brightness polarity p, where: E+_valid = {(x, y, t) | (x, y, p, t) ∈ E_valid, p = +1} E-_valid = {(x, y, t) | (x, y, p, t) ∈ E_valid, p = -1} These correspond to the two channels of the tensor, respectively.

[0040] S34. Time dimension aggregation: After mapping in step S32, events that appear at the same pixel coordinate (y, x) within the preset time window Δt in step S1 (adjusted to 50μs or 100μs depending on the detection speed) are aggregated to generate a three-dimensional event tensor T; For a positive event channel, its tensor value T(y, x, 1) = Agg(E+_valid(y, x, t)); For a negative event channel, its tensor value is T(y, x, 2) = Agg(E-_valid(y, x, t)); Where Agg(·) is a time aggregation function that uses event frequency statistics, that is, it counts the number of times an event occurs on the pixel within a time window Δt. The final result is a three-dimensional event tensor T with dimensions (H, W, 2), where the two channels of the third dimension correspond to positive and negative events, respectively, i.e., T∈R. H×W×2 It strictly matches the format requirements while fully preserving the core information (x,y,t,p), thus fully preserving the spatiotemporal dynamics and polarity characteristics of cable defects and adapting to subsequent simulation imaging operations; at the same time, it performs adaptive noise reduction, eliminating isolated events and high-frequency noise events based on the spatiotemporal distribution characteristics of events, ensuring an effective event retention rate of ≥95% and improving the purity of imaging data.

[0041] S4. Bionic Feature Enhancement Imaging: This method uses a GPU imaging processing module to extract multi-scale features from the 3D event tensor and dynamically adjusts the processing focus based on event activity. An attention mechanism is then used to enhance the imaging features of defective areas, generating biomimetic cable imaging data. Details are as follows: S41. Multi-scale feature extraction: Multiple two-dimensional convolutional kernels of different sizes are used to convolve the three-dimensional event tensor T to extract defect features at different scales; this process is represented as:

[0042] in, This represents a two-dimensional convolution operation; K_s represents the convolution kernel, with size s ∈ {3, 5, 7}, corresponding to 3×3, 5×5, and 7×7 convolution kernels respectively; F_s is the feature map of the corresponding scale obtained after convolution, F_s ∈ R. H×W×Cs C STo accommodate the number of feature channels at the corresponding scale, it adapts to the imaging needs of defects of different sizes, such as micron-level cracks and millimeter-level bulges, and accurately captures the density variation characteristics of positive and negative events in the defect area.

[0043] S42. Dynamic Receptive Field Adjustment: Calculate the event activity D(x, y) for each pixel, which serves as the basis for adjusting the convolutional receptive field; the event activity is obtained by adding the feature values ​​of the positive and negative event channels. D(x, y) = T(y, x, 1)+T(y, x, 2) Where T(y, x, 1) and T(y, x, 2) are the values ​​of the positive and negative event channels of the three-dimensional event tensor T at position (y, x), respectively; the receptive field range of the convolution kernel is adaptively adjusted according to the value of D(x, y): the receptive field is expanded for defect regions with high D(x, y) values ​​to enhance the imaging intensity of defect features; the receptive field is reduced for normal regions with low D(x, y) values ​​to reduce interference from invalid information and improve imaging contrast.

[0044] S43. Attention Feature Focusing: Based on the event activity D calculated in step S42, an attention weight map A is generated using the Sigmoid activation function, where the weight A(x, y) of each pixel is proportional to σ(D(x, y)), where σ is the Sigmoid function with a value range of (0,1); the multi-scale feature map F_s is multiplied element-wise with the attention weight map A to achieve feature focusing. F_enhanced = F_s⊙A Where ⊙ represents element-wise product, focusing on the high event density features of the defect area, further enhancing the differential imaging effect between the defect and the background, making it easier to distinguish subtle defects such as microcracks and epidermal peeling in the imaging data; F_enhanced is the feature map enhanced by the attention mechanism, that is, the enhanced cable simulacral imaging data.

[0045] S5. Imaging data standardization output: Convert the enhanced cable simulacrum imaging data into a standardized format and output it through multiple paths.

[0046] In this step, the GPU imaging processing module converts the enhanced cable simulacrum imaging data into a standardized format.

[0047] Multi-path output adapts to subsequent defect detection and management needs, specifically including: First, it provides real-time imaging output. The standardized cable-based imaging data is transmitted in real time via a high-speed data bus. The entire process from event acquisition to imaging data output has a latency of ≤3ms, meeting the real-time requirements of high-speed detection at 100km / h.

[0048] Secondly, data storage involves synchronously storing the original event stream data, the preprocessed 3D event tensor T, and the final generated cable simulacrum imaging data to a solid-state drive in the memory for subsequent defect verification, data tracing, and imaging parameter optimization.

[0049] Thirdly, visualization and remote transmission are achieved through the HDMI interface in the auxiliary interface module, which displays the cable bionic enhanced imaging image and defect area markings in real time. At the same time, the imaging data is remotely uploaded to the back-end management system through the 4G / 5G remote transmission interface, realizing dual protection of on-site monitoring and remote management. Example 1

[0050] (1) Equipment deployment and parameter configuration The imaging device is mounted on an inspection platform to perform inspection imaging of industrial control cables. The device configuration is as follows: the event camera module uses three 1280×720 pixel DVSlume event cameras to cover the full width of the cable; the FPGA preprocessing module enables a 50μs time window and 100 frames of cumulative frames, with the adaptive noise reduction threshold set to the default setting; the GPU imaging processing module uses an NVIDIA GeForce RTX 3090, enables half-precision computing, and presets the bionic vision perception layer convolution kernel size to 3×3, 5×5, and 7×7; the memory uses a 1TB SSD + 32GB DDR5 to ensure data storage and caching requirements; the auxiliary interface module enables HDMI visualization and 5G remote transmission functions.

[0051] (2) Implementation of the imaging process The detection scenarios are set to two lighting environments: strong light and nighttime. The imaging process is as follows: Parameter initialization: In high-speed scenes (80km / h), a 50μs time window is configured, and the event trigger threshold is increased by 15% in strong light environments; in medium-speed scenes (50km / h), a 100μs time window is configured, and the threshold is reduced by 20% in nighttime environments. The GPU imaging processing module enables half-precision calculation to complete the preset of the simulacral imaging parameters.

[0052] Event stream acquisition: Three event cameras simultaneously acquire event stream data on the cable surface, capturing brightness change events caused by defects such as microcracks, corrosion, and bulges, and transmit the data to the FPGA data preprocessing module in real time.

[0053] Preprocessing and Transformation: The FPGA data preprocessing module performs noise reduction on the event stream, removing environmental dust interference events to obtain a valid event set E_valid, with an effective event retention rate of 96.2%. Subsequently, spatial coordinate mapping is performed, mapping the event pixel coordinates (x, y) to a (1280×720) pixel grid, and ROI clipping is performed as needed. Then, based on polarity, channels are separated to obtain a positive event set E+_valid and a negative event set E-_valid. Temporal aggregation is performed, counting the occurrence frequency of positive and negative events of the same pixel within a 50μs / 100μs time window to obtain feature values. Finally, a three-dimensional event tensor T∈R is generated. 1280 ×720×2 The third dimension has two channels that correspond to positive and negative events, respectively, completing the conversion of unstructured data into a standardized two-dimensional representation, providing adapted data for subsequent GPU biomimetic enhancement imaging.

[0054] Bionic Feature Enhancement Imaging: GPU Imaging Processing Module for 3D Event Tensors T∈R 1280×720×2 The positive and negative event dual-channel features are extracted through a biomimetic visual perception layer to extract multi-scale features. Combined with dynamic receptive field adjustment, the positive and negative event comprehensive features of the focusing defect area are enhanced to strengthen the contrast between the defect and the background and generate optimized enhanced imaging data.

[0055] Standardized output: The GPU imaging processing module converts the biomimetic enhancement imaging data into a standardized format and outputs it in real time. At the same time, the imaging screen is displayed through the HDMI interface, uploaded to the back-end management system through the 5G network, and the full amount of data is stored synchronously on the SSD solid-state drive. The latency of the entire biomimetic imaging process is stable at 2.7-2.9ms.

[0056] (3) Verification of imaging effect This embodiment compares the imaging scheme of the traditional frame exposure camera with the scheme of the present invention to verify the imaging effect. The results are as follows: The scheme of the present invention has no motion blur in high-speed detection at 80km / h, and the imaging of 1.2mm microcracks is clearly distinguishable; the imaging signal-to-noise ratio is ≥35dB in strong light and nighttime environments, and the anti-interference ability is excellent; under the same information capture conditions, the amount of imaging data is reduced by 90% compared with the traditional scheme, the transmission and storage pressure is significantly reduced, and the quality and efficiency of the biomimetic imaging are outstanding.

[0057] Ablation experiments verified the effectiveness of the core modules: after removing the biomimetic feature enhancement layer, the contrast of defect imaging decreased by 40%, and the identification of small defects was significantly reduced; after turning off the adaptive noise reduction function, the effective event retention rate dropped to 78%, and the imaging data interference was obvious, proving the key supporting role of each module in imaging quality.

Claims

1. A high-speed biomimetic intelligent imaging device based on an event camera, characterized in that: include: The event camera module is used to continuously image the cable surface during high-speed movement, capture brightness change events caused by normal and defective areas of the cable, and generate unstructured event stream data. An FPGA data preprocessing module, connected to the event camera module, is used to process the event stream data in real time, performing adaptive noise reduction, spatial coordinate mapping, polarity channel separation and time dimension aggregation operations, and converting the unstructured event stream into a three-dimensional event tensor. The GPU imaging processing module is connected to the FPGA data preprocessing module and is used to perform a biomimetic feature enhancement algorithm on the three-dimensional event tensor. Through multi-scale feature extraction, dynamic receptive field adjustment and attention feature focusing, it generates cable biomimetic imaging data. A high-speed data bus and memory module includes a high-speed data bus and a memory; the high-speed data bus is connected to the event camera module, the FPGA data preprocessing module, the GPU imaging processing module, and the memory respectively, for realizing high-speed data transmission between the modules; the memory is used to provide persistent storage of data and real-time computing cache. An auxiliary interface module, connected to the GPU imaging processing module and the memory, is used to provide visualization display, remote data transmission and device adaptation interfaces.

2. The high-speed bionic intelligent imaging device based on an event camera according to claim 1, characterized in that: The event camera module includes at least three event cameras evenly arranged along the circumference of the cable. Each event camera has a pixel resolution of 1280×720, a minimum time window of 50 microseconds, microsecond-level time resolution, and a dynamic range of ≥120dB. The event trigger threshold can be adaptively adjusted according to the ambient light intensity to capture brightness change events caused by minute defects with feature size ≥1.2mm on the cable surface without motion blur at a detection speed of 30-100km / h.

3. The high-speed bionic intelligent imaging device based on an event camera according to claim 1, characterized in that: The FPGA data preprocessing module is built on a high-performance FPGA chip; the GPU imaging processing module uses a GPU that supports half-precision computing mode and has a built-in 3-layer bionic vision perception layer.

4. The high-speed bionic intelligent imaging device based on an event camera according to claim 1, characterized in that: The high-speed data bus adopts the PCIe 4.0 transmission protocol; the memory includes a solid-state drive for persistent full data storage and DDR memory for real-time computing caching; the auxiliary interface module includes an HDMI visualization interface for local display, a 4G / 5G remote transmission interface for wireless data transmission, and a standardized device interface for connecting external devices.

5. A high-speed biomimetic intelligent imaging method based on an event camera, characterized in that: The high-speed biomimetic intelligent imaging device based on an event camera, as described in any one of claims 1-4, comprises the following steps: S1. Imaging parameter initialization configuration: Based on the speed and lighting conditions of the cable detection scene, preset and optimize the core imaging parameters; S2. High-speed acquisition of cable event stream data: The event camera modules in an array layout synchronously acquire brightness change events on the cable surface and generate unstructured event stream data; S3. Event Stream Preprocessing and Structured Transformation: The acquired event stream data is adaptively denoised using the FPGA data preprocessing module and converted into a three-dimensional event tensor with fixed spatial dimensions and polarity channels; S4. Biomimetic Feature Enhancement Imaging: Multi-scale feature extraction of the three-dimensional event tensor is performed through the GPU imaging processing module, and the processing focus is dynamically adjusted in combination with the event activity. The imaging features of the defect area are enhanced through the attention mechanism to generate cable biomimetic imaging data. S5. Imaging data standardization output: Convert the enhanced cable simulacrum imaging data into a standardized format and output it through multiple paths.

6. The high-speed biomimetic intelligent imaging method based on an event camera according to claim 5, characterized in that: Step S1 specifically includes: Based on a preset detection speed range, a time window Δt and a cumulative number of frames are set for event aggregation processing; wherein, when the detection speed is greater than 80 km / h, the time window Δt is set to 50 microseconds; when the detection speed is less than or equal to 80 km / h, the time window Δt is set to 100 microseconds; the cumulative number of frames is fixed at 100 frames. Based on the real-time detected ambient light intensity, the event trigger threshold of the event camera is adaptively adjusted; specifically, in strong light environments, the threshold is increased by 10%-20%; in nighttime environments, the threshold is decreased by 15%-25%. Enable half-precision computing mode in the GPU imaging processing module, and preset the convolution kernel size and receptive field adjustment range of the bionic vision perception layer.

7. The high-speed biomimetic intelligent imaging method based on an event camera according to claim 5, characterized in that: The unstructured event stream data generated in step S2 is the original event stream E, represented as: E={(x, y, p, t)} Where (x, y) represents the pixel coordinates where the event occurs, x is the horizontal coordinate and y is the vertical coordinate; p is the brightness polarity, p ∈{+1, -1}, +1 represents a pixel brightness increase event, i.e., a positive event, and -1 represents a pixel brightness decrease event, i.e., a negative event; t is the timestamp of the event, with a precision in microseconds.

8. The high-speed biomimetic intelligent imaging method based on an event camera according to claim 7, characterized in that: Step S3 specifically includes: S31. Adaptive Noise Reduction: Based on the spatiotemporal distribution characteristics of events, isolated noise events and high-frequency interference events are removed from the original event stream E to obtain a valid event set E_valid related to the cable surface condition, where... The adaptive noise reduction process ensures an effective event retention rate of ≥95%. S32. Spatial coordinate mapping: Map the pixel coordinates (x, y) of each event in the valid event set E_valid to a pixel grid of size H×W, where H is the image height, W is the image width, and H = 1280, W = 720; S33. Polarity-based channel splitting: The valid event set E_valid is split into a positive event set E+_valid and a negative event set E-_valid according to the brightness polarity p, where: E+_valid = {(x, y, t) | (x, y, p, t) ∈ E_valid, p = +1} E-_valid = {(x, y, t) | (x, y, p, t) ∈ E_valid, p = -1}; S34. Time Dimension Aggregation: After mapping in step S32, events that appear at the same pixel coordinate (y, x) within the preset time window Δt in step S1 are aggregated to generate a three-dimensional event tensor T; For a positive event channel, its tensor value T(y, x, 1) = Agg(E+_valid(y, x, t)); For a negative event channel, its tensor value is T(y, x, 2) = Agg(E-_valid(y, x, t)); Where Agg(·) is a time aggregation function that uses event frequency statistics, that is, it counts the number of times an event occurs on the pixel within a time window Δt. The final result is a three-dimensional event tensor T with dimensions (H, W, 2), where the two channels of the third dimension correspond to positive and negative events, respectively.

9. The high-speed biomimetic intelligent imaging method based on an event camera according to claim 8, characterized in that: S4 specifically includes: S41. Multi-scale feature extraction: The three-dimensional event tensor T is convolved using multiple two-dimensional convolution kernels of different sizes to extract defect features at different scales; this process is represented as: in, This represents a two-dimensional convolution operation; K_s represents the convolution kernel, with size s ∈ {3, 5, 7}, corresponding to 3×3, 5×5, and 7×7 convolution kernels respectively; F_s is the feature map of the corresponding scale obtained after convolution; S42. Dynamic Receptive Field Adjustment: Calculate the event activity D(x, y) for each pixel, which serves as the basis for adjusting the convolutional receptive field; the event activity is obtained by adding the feature values ​​of the positive and negative event channels. D(x, y) = T(y, x, 1)+T(y, x, 2) Where T(y, x, 1) and T(y, x, 2) are the values ​​of the positive and negative event channels of the three-dimensional event tensor T at position (y, x), respectively; the receptive field range of the convolution kernel is adaptively adjusted according to the value of D(x, y): the receptive field is expanded for defect regions with high D(x, y) values ​​to enhance features, and the receptive field is shrunk for normal regions with low D(x, y) values ​​to reduce interference; S43. Attention Feature Focusing: Based on the event activity D calculated in step S42, an attention weight map A is generated using the Sigmoid activation function, where the weight A(x, y) of each pixel is proportional to σ(D(x, y)), and σ is the Sigmoid function; the multi-scale feature map F_s is multiplied element-wise with the attention weight map A to achieve feature focusing. F_enhanced = F_s⊙A Where ⊙ represents element-wise product; F_enhanced is the feature map enhanced by the attention mechanism, that is, the enhanced cable simulation imaging data.

10. The high-speed biomimetic intelligent imaging method based on an event camera according to claim 9, characterized in that: In S5, the multi-path output specifically includes: Standardized cable simulation imaging data is output through the high-speed data bus, wherein the entire process delay from event acquisition to imaging data output is ≤3ms; The original event stream data, the preprocessed 3D event tensor T, and the final generated cable simulac image data are synchronously stored in the solid-state drive in the memory. The HDMI interface in the auxiliary interface module displays the cable bionic enhanced imaging image and defect area markings in real time, and the imaging data is uploaded via the 4G / 5G remote transmission interface.

Images