Image acquisition method and device based on intelligent high-speed machine vision and storage medium

Abstract

The invention relates to the technical field of image processing, and discloses an image acquisition method and device based on intelligent high-speed machine vision, and a storage medium. The method comprises the following steps: performing cache allocation processing on trigger time sequence data to obtain memory pre-allocation data; performing region division processing on the trigger time sequence data and the memory pre-allocation data to obtain multi-region configuration data, and performing line acquisition processing on the multi-region configuration data to obtain original multi-region image data; performing row enhancement processing on the original multi-region image data to obtain enhanced image data, and performing spectrum fusion processing on the enhanced image data to obtain multi-spectrum composite image data; and performing feature extraction processing on the multispectral composite image data to obtain multi-level feature data, and performing hierarchical compression processing on the multi-level feature data to obtain compressed feature data and feature index data. According to the invention, the image acquisition efficiency based on intelligent high-speed machine vision is improved.

Description

Technical Field

[0001] The present application relates to the field of image processing, and in particular to an image acquisition method, device and storage medium based on intelligent high-speed machine vision. Background Art

[0002] With the rapid advancement of smart city construction, video surveillance technology has been widely used in urban security management, traffic monitoring, public place management and other fields. Traditional video surveillance systems usually adopt a single parameter configuration acquisition method, and often use preset exposure parameters and fixed frame rates during image acquisition. At the same time, the image processing methods in the prior art mostly adopt a serial processing method to analyze and process the collected video stream frame by frame, and usually use a single mode of image information for target detection and behavior analysis. In addition, in terms of data storage, the prior art mostly uses a unified compression method to compress and store video data, and fails to adopt differentiated compression strategies for different scenes and information of different importance.

[0003] However, this traditional video acquisition and processing method has some obvious shortcomings: first, the fixed parameter acquisition method cannot adapt to the complex and changeable lighting conditions in the urban environment (such as daytime, nighttime, rainy days, etc.), which easily leads to unstable image quality; second, the single-mode acquisition method is difficult to meet the needs of different monitoring scenarios (such as crowded areas, open roads, dark corners, etc.), affecting the information integrity of the image; third, the serial processing method is inefficient and difficult to meet the real-time processing requirements in large-scale video surveillance; finally, the unified compression storage strategy will lead to the loss of key monitoring information, affecting the subsequent analysis and judgment results. Summary of the invention

[0004] The present application provides an image acquisition method, device and storage medium based on intelligent high-speed machine vision, which are used to improve the efficiency of image acquisition based on intelligent high-speed machine vision.

[0005] In a first aspect, the present application provides an image acquisition method based on intelligent high-speed machine vision, the image acquisition method based on intelligent high-speed machine vision comprising: acquiring and processing the light intensity of multiple points in a detection area to obtain discrete light intensity data, and performing light mapping processing on the discrete light intensity data to obtain continuous light distribution data and initial exposure parameter data; performing multi-spectral analysis processing on the continuous light distribution data and the initial exposure parameter data to obtain basic lighting configuration data, and performing timing control processing on the basic lighting configuration data to obtain multi-spectral lighting sequence data; performing trigger timing processing on the multi-spectral lighting sequence data and the initial exposure parameter data to obtain trigger timing. The trigger timing data is processed, and the cache allocation processing is performed on the trigger timing data to obtain memory pre-allocated data; the trigger timing data and the memory pre-allocated data are processed by area division to obtain multi-area configuration data, and the multi-area configuration data is processed by parallel acquisition to obtain original multi-area image data; the original multi-area image data is processed by parallel enhancement to obtain enhanced image data, and the enhanced image data is processed by spectral fusion to obtain multi-spectral composite image data; the multi-spectral composite image data is processed by feature extraction to obtain multi-level feature data, and the multi-level feature data is processed by layered compression to obtain compressed feature data and feature index data.

[0006] In a second aspect, the present application provides an image acquisition device based on intelligent high-speed machine vision, and the image acquisition device based on intelligent high-speed machine vision comprises:

[0007] An acquisition module is used to collect and process the light intensity of multiple points in the detection area to obtain discrete light intensity data, and to perform light mapping processing on the discrete light intensity data to obtain continuous light distribution data and initial exposure parameter data;

[0008] a processing module, configured to perform multispectral analysis processing on the continuous illumination distribution data and the initial exposure parameter data to obtain basic illumination configuration data, and perform timing control processing on the basic illumination configuration data to obtain multispectral illumination sequence data;

[0009] A trigger module, used for performing trigger timing processing on the multi-spectral illumination sequence data and the initial exposure parameter data to obtain trigger timing data, and performing cache allocation processing on the trigger timing data to obtain memory pre-allocated data;

[0010] A division module, used for performing region division processing on the trigger timing data and the memory pre-allocated data to obtain multi-region configuration data, and performing parallel acquisition processing on the multi-region configuration data to obtain original multi-region image data;

[0011] An enhancement module, used for performing parallel enhancement processing on the original multi-region image data to obtain enhanced image data, and performing spectral fusion processing on the enhanced image data to obtain multi-spectral composite image data;

[0012] The extraction module is used to perform feature extraction processing on the multi-spectral composite image data to obtain multi-level feature data, and perform layered compression processing on the multi-level feature data to obtain compressed feature data and feature index data.

[0013] A third aspect of the present application provides a computer-readable storage medium, in which instructions are stored. When the computer-readable storage medium is run on a computer, the computer executes the above-mentioned image acquisition method based on intelligent high-speed machine vision.

[0014] In the technical solution provided by the present application, the technical solution of multi-spectral analysis and timing control processing is adopted to realize the coordinated work of multiple band light sources, which not only improves the integrity of urban monitoring image information, but also avoids interference between different bands, and effectively solves the image acquisition problem in complex scenes such as day and night alternation and weather changes. Through the cooperation of trigger timing processing and cache allocation processing, an accurate data acquisition and storage mechanism is established to ensure the data integrity in the high-speed acquisition process and meet the requirements of all-weather uninterrupted acquisition in urban video surveillance. The combination of regional division processing and parallel acquisition processing significantly improves the acquisition efficiency of the system, making large-scale, high-resolution urban road monitoring possible, and can cover multiple intersections, streets and other key areas at the same time. The application of parallel enhancement processing and spectral fusion processing not only improves the quality of monitoring images, but also realizes the effective integration of multiple spectral band information, enhances the information richness of images, especially the target recognition ability at night and in bad weather conditions. The innovative application of feature extraction processing and layered compression processing not only ensures the accurate preservation of important target feature information such as vehicles and pedestrians, but also realizes the optimization of storage space for massive monitoring data, achieving a good balance between data processing efficiency and storage efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying any creative work.

[0016] Figure 1 A schematic diagram of an embodiment of an image acquisition method based on intelligent high-speed machine vision in an embodiment of the present application;

[0017] Figure 2This is a schematic diagram of an embodiment of an image acquisition device based on intelligent high-speed machine vision in an embodiment of the present application. DETAILED DESCRIPTION

[0018] The present application embodiment provides an image acquisition method, device and storage medium based on intelligent high-speed machine vision. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described here can be implemented in an order other than the content illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0019] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiment of the present application, an embodiment of the image acquisition method based on intelligent high-speed machine vision includes:

[0020] Step S101, collecting and processing the multi-point light intensity of the detection area to obtain discrete light intensity data, and performing light mapping processing on the discrete light intensity data to obtain continuous light distribution data and initial exposure parameter data;

[0021] Step S102: performing multispectral analysis processing on the continuous illumination distribution data and the initial exposure parameter data to obtain basic illumination configuration data, and performing timing control processing on the basic illumination configuration data to obtain multispectral illumination sequence data;

[0022] Step S103, performing trigger timing processing on the multi-spectral illumination sequence data and the initial exposure parameter data to obtain trigger timing data, and performing cache allocation processing on the trigger timing data to obtain memory pre-allocated data;

[0023] Step S104, performing region division processing on the trigger timing data and the memory pre-allocated data to obtain multi-region configuration data, and performing parallel acquisition processing on the multi-region configuration data to obtain original multi-region image data;

[0024] Step S105, performing parallel enhancement processing on the original multi-region image data to obtain enhanced image data, and performing spectral fusion processing on the enhanced image data to obtain multi-spectral composite image data;

[0025] Step S106: performing feature extraction processing on the multi-spectral composite image data to obtain multi-level feature data, and performing layered compression processing on the multi-level feature data to obtain compressed feature data and feature index data.

[0026] It is understandable that the execution subject of the present application may be an image acquisition device based on intelligent high-speed machine vision, or a terminal or a server, which is not specifically limited here. The present application embodiment is described by taking a server as the execution subject as an example.

[0027] Specifically, the illumination information of the detection area is obtained by multi-point illumination intensity acquisition processing. Specifically, a plurality of photoelectric sensors are arranged in the detection area, and these sensors can convert the illumination intensity into electrical signal data. Discrete illumination intensity data is formed by processing these scattered light intensity data points. Subsequently, the illumination mapping processing technology is used to convert these discrete data points into a continuous illumination distribution surface. This process uses a bilinear interpolation algorithm to generate complete illumination distribution data by calculating the light intensity values ​​between adjacent sampling points. For example, when the light intensity values ​​of two adjacent sampling points are 80lux and 100lux respectively, the light intensity value at the middle position is approximately 90lux through interpolation calculation. At the same time, based on the illumination distribution data, the surface fitting is performed in combination with the least squares method to calculate the initial exposure parameter data, including key parameters such as sensitivity and shutter speed. After obtaining the continuous illumination distribution data and the initial exposure parameter data, multi-spectral analysis processing is performed. The process first divides the spectrum into multiple bands according to the wavelength range, such as visible light (400-700nm), near infrared (700-1000nm), etc., and then allocates energy to each band. Through the spectral energy allocation algorithm, the energy weight of each band is calculated to generate basic lighting configuration data. Next, the timing control processing is performed to switch the light sources of different bands according to the predetermined timing to realize the generation of multi-spectral lighting sequence data. Specifically, microsecond-level timing control is adopted, such as the visible light exposure time is 100 microseconds and the near infrared exposure time is 150 microseconds. Precise time control is used to ensure that there is no interference between light sources of different bands.

[0028] For the generated multi-spectral illumination sequence data and initial exposure parameter data, trigger timing processing is performed. By dividing the illumination sequence into time windows, the benchmark trigger timing is set, such as each acquisition cycle is 1000 microseconds, which contains multiple sub-cycles for image acquisition in different bands. A double-layer trigger structure is adopted, the outer trigger controls the entire acquisition cycle, and the inner trigger accurately controls the exposure timing of each band. Based on these timing data, cache allocation processing is performed to pre-divide the memory space to ensure the data storage requirements during high-speed acquisition. For example, for a 4K resolution image, each frame of raw data occupies about 32MB of space, and the system pre-allocates sufficient cache space accordingly. Based on the trigger timing data and memory pre-allocation data, regional division processing is performed. The entire image acquisition area is divided into multiple regions of interest, and the acquisition parameters can be set independently for each region. Through parallel acquisition processing technology, image acquisition is performed on multiple regions at the same time, which significantly improves the acquisition efficiency. In practical applications, for example, a 4K image is divided into four 1080p regions, each region is equipped with an independent data channel and cache space, to achieve true parallel acquisition.

[0029] The collected original multi-region image data is first enhanced in parallel. Through the multi-threaded parallel processing framework, the images of each region are enhanced by denoising, contrast adjustment and other operations. Then the spectral fusion processing is performed to fuse the image information collected in different bands. The wavelet transform algorithm is used for multi-scale decomposition, and then weighted fusion is performed according to the importance of features at different scale layers, and finally multi-spectral composite image data with richer information is generated. Finally, the multi-spectral composite image data is feature extracted and compressed and stored. First, the multi-scale feature extraction algorithm is used to extract the key feature information in the image, including edge, texture, color and other features, to form multi-level feature data. Then, according to the importance of the feature, a hierarchical compression strategy is adopted, with lossless compression for important features and lossy compression for minor features, and finally compressed feature data and feature index data are generated.

[0030] Taking urban road monitoring as an example, when monitoring an intersection area, the illumination distribution data of the monitoring area is first obtained. The light intensity in the center area of ​​the intersection is measured to be 1200 lux, and the average light intensity in the edge area is 800 lux. Through illumination mapping processing, a complete illumination distribution surface is generated, and the initial exposure parameters are calculated: sensitivity 400, shutter speed 1 / 1000 second. Then multi-spectral lighting control is performed, and visible light and near-infrared light are alternately irradiated with an interval of 200 microseconds. The system uses 5 parallel processing channels to simultaneously collect image data from different monitoring areas, and the image resolution of each area is 1920×1080 pixels. Through parallel enhancement processing and spectral fusion, a clear composite image is finally obtained, and key target feature information is extracted. For the extracted feature data, important target features such as vehicles and pedestrians are saved using lossless compression, and secondary background environment features are compressed using lossy compression. The compression ratio can reach 10:1, while maintaining the target detection accuracy above 98%.

[0031] In the embodiment of the present application, the technical scheme of multi-spectral analysis and timing control processing is adopted to realize the coordinated work of multiple band light sources, which not only improves the integrity of urban monitoring image information, but also avoids interference between different bands, and effectively solves the image acquisition problem in complex scenes such as day and night alternation and weather changes. Through the cooperation of trigger timing processing and cache allocation processing, an accurate data acquisition and storage mechanism is established to ensure the data integrity in the high-speed acquisition process and meet the demand for all-weather uninterrupted acquisition in urban video monitoring. The combination of regional division processing and parallel acquisition processing significantly improves the acquisition efficiency of the system, making large-scale, high-resolution urban road monitoring possible, and can cover multiple intersections, streets and other key areas at the same time. The application of parallel enhancement processing and spectral fusion processing not only improves the quality of monitoring images, but also realizes the effective integration of multiple spectral band information, enhances the information richness of images, especially the target recognition ability at night and in bad weather conditions. The innovative application of feature extraction processing and layered compression processing not only ensures the accurate preservation of important target feature information such as vehicles and pedestrians, but also realizes the optimization of storage space for massive monitoring data, achieving a good balance between data processing efficiency and storage efficiency.

[0032] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0033] (1) Grid partitioning is performed on the detection area to obtain multi-point sampling coordinate data, and photoelectric signal acquisition and processing is performed on each sampling point in the multi-point sampling coordinate data to obtain discrete light intensity data;

[0034] (2) performing signal denoising processing on the discrete light intensity data to obtain filtered light intensity data, and performing spatial mapping processing on the filtered light intensity data through a bilinear interpolation algorithm to obtain initial light distribution data;

[0035] (3) Performing surface fitting processing on the initial illumination distribution data by the least square method to obtain continuous illumination distribution surface data, and performing interval division processing on the continuous illumination distribution surface data to obtain illumination level data;

[0036] (4) performing parameter mapping processing on the light level data to obtain sensitivity parameter data and shutter speed parameter data, and combining the sensitivity parameter data and the shutter speed parameter data to obtain basic exposure parameter data;

[0037] (5) performing compensation processing on the basic exposure parameter data to obtain compensation coefficient data, and synthesizing the compensation coefficient data and the basic exposure parameter data to obtain target exposure parameter data;

[0038] (6) Parameter fusion is performed on the target exposure parameter data and the continuous illumination distribution surface data to obtain the continuous illumination distribution data and the initial exposure parameter data.

[0039] Specifically, grid partitioning refers to dividing the entire detection area into several rectangular grids at regular intervals, and the intersection of each grid is the sampling point. The specific operation is to evenly divide the detection area in the X-axis and Y-axis directions to form M×N grids. The size of each grid is determined according to the detection accuracy requirements, usually 10mm×10mm. The spatial coordinates (x, y) are recorded at each grid intersection to obtain multi-point sampling coordinate data. Photoelectric sensors are arranged at each sampling point. These sensors can convert the received light intensity into electrical signals, and obtain digital discrete light intensity data through analog-to-digital conversion.

[0040] After obtaining the discrete light intensity data, signal denoising is required to improve data reliability. The median filtering algorithm is used to process the original light intensity data. Specifically, a 5×5 window area is taken around each sampling point, and the 25 data points are sorted by size, and the middle value is selected as the filtered data of the point. The light intensity data obtained after filtering is smoother and can better reflect the actual light distribution. Next, the bilinear interpolation algorithm is used to spatially map the filtered discrete data points and calculate the light intensity values ​​between the grids. The specific process of bilinear interpolation is: for any point in the grid, the light intensity value of the point is obtained by using the light intensity values ​​of the four surrounding grid points and weighted by distance. For the initial light distribution data obtained, the least squares method is used for surface fitting to fit the discrete light intensity data points into a continuous three-dimensional surface. The core idea of ​​the least squares method is to minimize the sum of square errors from all data points to the fitting surface. Set the quadratic surface equation z=ax2 +by 2 +cxy+dx+ey+f, and the best fitting surface is obtained by solving the coefficients a, b, c, d, e, and f. Then the continuous light distribution surface is divided into intervals and divided into multiple levels according to the light intensity value, such as 0-1000 lux is divided into a weak light area, 1000-2000 lux is divided into a standard light area, and 2000 lux and above is divided into a strong light area, forming light level data.

[0041] Based on the light level data, a parameter mapping relationship is established to calculate the corresponding sensitivity parameters and shutter speed parameters. In low-light areas, the sensitivity is increased (such as set to ISO 800) and the shutter speed is appropriately extended (such as set to 1 / 100 second); in strong light areas, the sensitivity is reduced (such as set to ISO 200) and the shutter speed is shortened (such as set to 1 / 1000 second). These parameters are combined to form basic exposure parameter data. Compensation processing is then performed to calculate the compensation coefficient based on the actual acquisition effect. If the image in a certain area is overexposed, the compensation coefficient is less than 1 (such as 0.8); if it is underexposed, the compensation coefficient is greater than 1 (such as 1.2). Multiply the compensation coefficient by the basic exposure parameter to obtain the target exposure parameter data.

[0042] Finally, the target exposure parameter data is parameter-fused with the continuous illumination distribution surface data to generate the final continuous illumination distribution data and initial exposure parameter data as the basis for subsequent image acquisition. Taking the surface inspection of industrial parts as a specific application scenario, the inspection area size is 100mm×100mm, which is divided into 10×10 grids with a grid spacing of 10mm, and 100 sampling points are obtained. In the measurement at a certain moment, the light intensity value of the central area is 2500lux, and the average light intensity of the edge area is 1200lux. Outliers are removed by median filtering, such as correcting the 3000lux outlier measured at a certain sampling point to the median of 2400lux of the surrounding points. The light intensity value of the points inside the grid is calculated by bilinear interpolation. For example, for a point in the grid with the light intensity values ​​of the surrounding four sampling points being 2200lux, 2400lux, 2300lux, and 2500lux respectively, the light intensity value of the point is calculated to be 2350lux. The illumination distribution surface obtained by least squares fitting is in the shape of a bowl with high center and low sides. The surface equation parameters are a=-0.02, b=-0.02, c=0, d=0, e=0, f=2500. According to the illumination level, ISO 200 and shutter speed 1 / 1000 second are used in the center strong light area, and ISO 400 and shutter speed 1 / 500 second are used in the edge standard light area. After compensation processing, the exposure parameters are multiplied by the compensation coefficient of 0.85 in the overexposed area, and finally the accurate exposure parameter configuration is obtained.

[0043] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0044] (1) performing spectrum analysis on the continuous illumination distribution data to obtain spectral distribution data, and performing band division processing on the spectral distribution data to obtain multi-band illumination data;

[0045] (2) combining the multi-band illumination data and the initial exposure parameter data to obtain band weight data, and performing energy calculation processing on the band weight data using a spectral energy allocation algorithm to obtain light source energy allocation data;

[0046] (3) performing pulse width modulation processing on the light source energy allocation data to obtain light source modulation signal data, and performing period segmentation processing on the light source modulation signal data to obtain basic lighting configuration data;

[0047] (4) Performing microsecond-level timing tag processing on the basic lighting configuration data to obtain time stamp sequence data, and performing staggered arrangement processing on the time stamp sequence data to obtain light source switching timing data;

[0048] (5) performing synchronization signal generation processing on the light source switching timing data to obtain trigger synchronization data, and combining the trigger synchronization data and the light source switching timing data to obtain timing control data;

[0049] (6) Performing time slice allocation processing on the timing control data to obtain spectrum acquisition sequence data, and performing verification processing on the spectrum acquisition sequence data to obtain multi-spectrum lighting sequence data.

[0050] Specifically, the continuous illumination distribution data is subjected to spectrum analysis. Spectral analysis refers to converting the illumination distribution data in the spatial domain to the frequency domain through Fourier transform to obtain the spectral intensity distribution in different wavelength ranges. The specific operation is to perform Fourier transform on the illumination data in the visible spectrum range of 380nm-780nm to obtain the spectral distribution data. Subsequently, the band division process is performed to divide the spectrum into multiple bands according to functional characteristics, such as the blue light band (380-495nm), the green light band (495-570nm), the red light band (620-780nm) and the near infrared band (780-1100nm). Each band records its spectral energy distribution characteristics to form multi-band illumination data. After obtaining the multi-band illumination data, it is combined with the initial exposure parameter data. In the combination process, the corresponding relationship between the spectral characteristics of each band and the initial exposure parameters is considered, and the importance weight of each band is calculated. For example, for urban night monitoring scenes, the near-infrared band has a better effect on target recognition in low-light environments, so it is given a higher weight (such as 0.4); the visible light band is more sensitive to capturing target detail features, and the weights of each visible light sub-band are: red light 0.2, green light 0.2, blue light 0.2. The band weight data obtained in this way is further processed by the spectral energy allocation algorithm, which calculates the actual energy value required for each band based on the weight value of each band and the maximum output power of the light source. If the maximum power of a light source is 100W, the near-infrared band is allocated 40W, and the other visible light bands are allocated 20W each, thereby obtaining the light source energy allocation data.

[0051] The light source energy allocation data is pulse-width modulated to convert the continuous energy allocation into a discrete time control signal. Pulse-width modulation (PWM) is a technology that controls the energy output of a light source by adjusting the duty cycle. The larger the duty cycle, the greater the energy output of the light source. According to the calculated energy allocation ratio, the PWM duty cycle of each band light source is set, such as a 40% duty cycle for the near-infrared light source and a 20% duty cycle for the visible light source. Then, these PWM signals are segmented according to a fixed cycle (such as 1000 microseconds) to generate basic lighting configuration data. The basic lighting configuration data is processed by microsecond-level timing mark to add an accurate timestamp to the opening and closing moments of each light source. The accuracy of the timestamp reaches the microsecond level (1us), recording the precise action time point of each band light source within a cycle. For example, the red light is turned on at 0us and turned off at 200us; the green light is turned on at 200us and turned off at 400us; and so on. These sequences with timestamps are interleaved to ensure that light sources of different bands work alternately to avoid spectral crosstalk and form light source switching timing data.

[0052] The light source switching timing data is processed by synchronization signal generation to generate trigger signals for synchronizing each light source and image acquisition device. These synchronization signals include the master clock signal (frequency is 1MHz) and the enable signals of each band. These trigger synchronization data are combined with the light source switching timing data to generate complete timing control data, which contains the switching timing and synchronization trigger signals of all bands of light sources. Finally, the timing control data is allocated to time slices, and a complete acquisition cycle (such as 1000us) is divided into multiple time segments, each of which is allocated to a light source of a specific band. A cyclic redundancy check (CRC) is performed on the data of each time segment to ensure the correctness and integrity of the data, and finally the multi-spectral lighting sequence data is obtained.

[0053] Taking the monitoring of urban traffic intersections as an example, after spectrum analysis of the continuous illumination distribution data in the monitoring area, the energy distribution of four main bands is identified: the near infrared band (800-1000nm) accounts for 35% of the total energy, the red light band (620-780nm) accounts for 25%, the green light band (495-570nm) accounts for 20%, and the blue light band (380-495nm) accounts for 20%. Based on this distribution feature and the initial exposure parameters (ISO 400, shutter speed 1 / 500 seconds), the weights of each band are calculated: near infrared 0.35, red light 0.25, green light 0.2, blue light 0.2. The total power of the light source is 200W, and the energy of each band is distributed accordingly: near infrared 70W, red light 50W, green light 40W, blue light 40W. In the 1000us acquisition cycle, it is divided into 4 time slices through PWM control: 0-350us is allocated to the near-infrared light source (duty cycle 35%, used for target detection at night and in low-light environments), 350-600us is allocated to the red light source (duty cycle 25%, enhancing target contour information), 600-800us is allocated to the green light source (duty cycle 20%, enhancing environmental details), and 800-1000us is allocated to the blue light source (duty cycle 20%, supplementing scene hierarchy). 10us is reserved before the start of each time slice for light source stabilization, and 10us is reserved before the end for light source shutdown, thereby ensuring the quality of the acquired image. The multispectral illumination sequence data finally generated contains a complete four-band illumination timing, and each band has an accurate timestamp and synchronization trigger signal.

[0054] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0055] (1) performing time window segmentation processing on the multi-spectral illumination sequence data to obtain reference time window data, and performing trigger interval calculation processing on the reference time window data to obtain trigger interval sequence data;

[0056] (2) performing synchronous calibration processing on the trigger interval sequence data and the initial exposure parameter data to obtain external trigger reference data, and performing pulse edge detection processing on the external trigger reference data to obtain coarse positioning trigger data;

[0057] (3) performing phase compensation processing on the coarse positioning trigger data to obtain trigger compensation data, and performing fine-tuning processing on the trigger compensation data through a timing compensation algorithm to obtain precise trigger data;

[0058] (4) performing trigger priority allocation processing on the precise trigger data to obtain trigger priority data, and performing trigger timing combination processing on the trigger priority data to obtain trigger timing data;

[0059] (5) Performing storage space calculation processing on the trigger timing data to obtain cache capacity data, and performing storage unit division processing on the cache capacity data to obtain storage block data;

[0060] (6) Performing data stream channel allocation processing on the storage block data to obtain cache channel data, and performing storage address mapping processing on the cache channel data to obtain memory pre-allocated data.

[0061] Specifically, the multi-spectral illumination sequence data is processed by time window segmentation. Time window segmentation refers to dividing the entire acquisition cycle into multiple time periods, each of which corresponds to a specific spectral band acquisition process. Specifically, the standard acquisition cycle of 1000 microseconds is divided into 4 reference time windows, each of which corresponds to the acquisition period of near infrared, red light, green light and blue light. Each time window is further subdivided into preparation time (10 microseconds), exposure time and buffer time (10 microseconds), so as to obtain the reference time window data. Then the trigger interval between adjacent time windows is calculated to ensure the precise synchronization of light source switching and image acquisition, and generate trigger interval sequence data. The trigger interval sequence data is synchronously calibrated with the initial exposure parameter data, and the calibration process takes into account the exposure time requirements of different bands. For example, the near infrared band requires an exposure time of 350 microseconds, and the red light band requires an exposure time of 250 microseconds. By adjusting the trigger timing, it is ensured that each band obtains sufficient exposure time. The calibrated data forms the external trigger reference data, and then the pulse edge detection is performed to identify the rising and falling edge moments of each trigger signal, and record them as coarse positioning trigger data.

[0062] Phase compensation is performed on the coarse positioning trigger data to compensate for the timing deviation caused by signal transmission delay. In practical applications, the response time of light sources in different bands is different. For example, the startup delay of the near-infrared light source is 2 microseconds, and the startup delay of the red light source is 1.5 microseconds. These differences need to be compensated to obtain trigger compensation data. The compensation data is fine-tuned through the timing compensation algorithm, and the influence of factors such as temperature drift and device aging are considered to finally generate accurate trigger data. Priority is assigned to the accurate trigger data, and the trigger priority is set according to the importance of different bands. In metal surface defect detection, the near-infrared band is more important for the detection of tiny cracks, so the highest priority 1 is assigned, followed by the red light band priority 2, and the green and blue light bands are priorities 3 and 4 respectively. A complete trigger timing is generated according to the priority data, including the generation time, duration and interval time of the trigger signal, to form trigger timing data.

[0063] For the trigger timing data, calculate the required storage space size. Taking the image with 4K resolution (4096×3072 pixels) and 12-bit color depth as an example, the data volume of a single frame image is 18MB. Considering the acquisition requirements of 4 bands, and reserving 20% ​​of the buffer space, the required cache capacity is calculated to be 86.4MB. Divide these storage spaces into multiple storage units, each unit corresponds to the image data of a band, and obtain the storage block data. Finally, allocate data stream channels for the storage block data. Allocate a high-speed data channel for each band separately, and the channel bandwidth is determined according to the data volume and transmission time requirements. For example, if the data volume of the near-infrared band image is 18MB and the transmission time requirement is 200 microseconds, the required channel bandwidth is 90MB / s. After completing the channel allocation, establish a mapping relationship from virtual address to physical address to ensure that the data can be accessed quickly and accurately, and finally generate memory pre-allocated data.

[0064] Taking urban night traffic monitoring as a specific application scenario, within the 1000 microsecond acquisition cycle, the near-infrared band (300-320 microseconds) is allocated a 350 microsecond window, including 10 microseconds of preparation time, 330 microseconds of exposure time and 10 microseconds of buffer time. The trigger timing is determined by synchronous calibration: the near-infrared light source is triggered at 0 microseconds, considering a 2-microsecond startup delay, and starts exposure at 12 microseconds, which is used for target detection under low-light conditions; the red light band is triggered at 350 microseconds, considering a 1.5-microsecond delay, and starts exposure at 361.5 microseconds, which is used to enhance the edge features of targets such as vehicles. In terms of storage space allocation, each band is allocated 20MB of cache space, the data write address range of the near-infrared channel is 0-20MB, the red light channel is 20-40MB, the green light channel is 40-60MB, and the blue light channel is 60-80MB, and a 2MB safety interval is reserved between each channel. In actual tests, the complete four-band image acquisition process is carried out strictly according to the timing, and the entire process from triggering to data storage is controlled within 950 microseconds, meeting the requirements of real-time video acquisition in urban monitoring.

[0065] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0066] (1) Performing time domain segmentation processing on the trigger timing data to obtain acquisition timing segmentation data, and matching the acquisition timing segmentation data with the memory pre-allocated data to obtain time-space corresponding data;

[0067] (2) dividing the time-space corresponding data into regions of interest to obtain region boundary data, and dividing the region boundary data into blocks using a region segmentation algorithm to obtain region block data;

[0068] (3) performing acquisition resource configuration processing on the regional block data to obtain regional resource data, and performing channel allocation processing on the regional resource data to obtain parallel channel data;

[0069] (4) performing time slot allocation processing on the parallel channel data to obtain parallel acquisition timing data, and performing thread scheduling processing on the parallel acquisition timing data to obtain multi-region configuration data;

[0070] (5) collecting and processing each region in the multi-region configuration data through an independent data stream to obtain regional sub-image data, and performing cache writing processing on the regional sub-image data to obtain cache image data;

[0071] (6) Performing data stream merging processing on the cached image data to obtain merged image data, and performing image reorganization processing on the merged image data to obtain original multi-region image data.

[0072] Specifically, the trigger timing data is processed by time domain segmentation. Time domain segmentation refers to dividing the continuous trigger timing according to the time axis, and each time period corresponds to a complete image acquisition process. The specific operation is to divide the 1000 microsecond acquisition cycle into multiple sub-cycles, each sub-cycle contains the rising edge, stabilization time, acquisition time and falling edge of the trigger signal, forming acquisition timing segmented data. These segmented data are then matched with the pre-allocated data in the memory to establish a correspondence between the time point and the storage space. For example, the data collected in the first sub-cycle (0-250 microseconds) is allocated to a storage space of 0-20MB, and the data in the second sub-cycle (250-500 microseconds) is allocated to a storage space of 20-40MB, thereby obtaining time-space corresponding data. The time-space corresponding data is processed by region of interest division, and the entire field of view is divided into multiple key focus areas according to the detection requirements. For example, in a 4K resolution (4096×3072 pixels) field of view, the image is divided into 16 regions, each of which is 1024×768 pixels in size, to obtain region boundary data. These boundary data are processed by the region segmentation algorithm, and considering the overlap and boundary transition between regions, each region is additionally extended with an overlap band of 32 pixels to ensure the continuity of image information between regions, and finally the region block data is obtained.

[0073] The acquisition resource configuration is performed for the regional block data, and independent acquisition resources are allocated to each region, including the readout port of the image sensor, the data cache, the processing unit, etc. In the scenario of 16 regions, 4 readout ports are allocated to 4 groups of regions, each group contains 4 adjacent regions, and regional resource data is formed. Subsequently, these resources are allocated to channels, and 4 independent data transmission channels are established. Each channel is responsible for the data transmission of a group of regions, and the transmission bandwidth is 1GB / s, and parallel channel data is obtained. Time slot allocation is performed for the parallel channel data, and the time resources within the acquisition cycle are allocated to each channel. In the 1000 microsecond acquisition cycle, each channel is allocated an exclusive time slice of 250 microseconds for data transmission and processing. Through the thread scheduling mechanism, an independent processing thread is allocated to each channel to realize parallel data processing and generate multi-region configuration data. The specific scheduling strategy is to adopt a polling method, and each thread completes data acquisition, transmission and cache writing operations within its time slice.

[0074] Each region in the multi-region configuration data is collected through an independent data stream. The collection process of each region has independent control parameters, including exposure time, gain value, etc. For example, for areas with strong surface reflection, the exposure time is reduced to 100 microseconds and the gain is reduced to 1.5 times; for shadow areas, the exposure time is increased to 200 microseconds and the gain is increased to 2.5 times, thereby obtaining regional sub-image data. These data are processed by cache write to achieve temporary storage of data. The data of each region occupies about 3MB cache space, and all regions have a total of 48MB cache space to form cache image data. Finally, the cache image data is merged by data stream merging. The image data of 16 regions are reorganized according to the spatial position relationship, the redundant information of the overlapping area is eliminated, and the boundary smoothing is performed to ensure the continuity of the spliced ​​image, and the merged image data is obtained. The merged data is reorganized to restore the integrity and spatial relationship of the original image, and finally the original multi-region image data is obtained.

[0075] Taking the monitoring of large urban intersections as an example, the monitoring area covers a 200m×150m road intersection area. A 4K camera is used for acquisition, and the field of view is divided into 16 areas, each corresponding to a monitoring range of 12.5m×9.375m. In the actual acquisition process, firstly, according to the time sequence allocation, the first group of 4 areas (the central area of ​​the main intersection) are acquired in 0-250 microseconds, and the data is written into 0-12MB storage space; the second group of areas (pedestrian crossing areas) are acquired in 250-500 microseconds, and the data is written into 12-24MB space; and so on. Each area uses independent parameter configuration, such as 150 microsecond exposure time for dense traffic areas and 180 microsecond exposure time for open road areas. Data is transmitted simultaneously through 4 parallel channels, and the peak transmission rate of each channel reaches 800MB / s. Finally, the image data of the 16 areas are merged and reorganized to obtain a complete 4K resolution intersection monitoring image. The total acquisition process takes no more than 1000 microseconds.

[0076] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0077] (1) performing signal-noise separation processing on each region in the original multi-region image data to obtain regional noise feature data, and performing noise suppression processing on the regional noise feature data to obtain enhanced image data;

[0078] (2) performing grayscale histogram analysis on the enhanced image data to obtain grayscale distribution data, and performing contrast adjustment on the grayscale distribution data using a histogram equalization algorithm to obtain contrast enhancement data;

[0079] (3) performing spectral separation processing on the contrast enhancement data according to the wavelength range to obtain multi-band image data, and performing frequency domain transformation processing on the multi-band image data to obtain spectral component data;

[0080] (4) performing band weight allocation processing on the spectral component data to obtain band weight data, and performing coefficient calculation processing on the band weight data through a wavelet transform algorithm to obtain fusion coefficient data;

[0081] (5) performing energy superposition processing on the fusion coefficient data and the multi-band image data to obtain spectral superposition data, and performing energy balance processing on the spectral superposition data to obtain balanced spectrum data;

[0082] (6) Performing inverse transformation processing on the balanced spectrum data to obtain spatial domain image data, and performing color mapping processing on the spatial domain image data to obtain multi-spectral composite image data.

[0083] Specifically, the original multi-region image data is processed for signal-noise separation. This process uses advanced noise analysis technology to decompose the image signal into two parts: effective signal and noise. For each regional image, the local variance and signal-to-noise ratio are calculated to identify noise features. For example, in a regional image of 1024×768 pixels, the local variance is calculated through a sliding window of 8×8 pixels, and the area that is 3 times higher than the background variance is marked as a noise area to form regional noise feature data. Subsequently, these noise feature data are suppressed, and the adaptive median filtering method is used. The window size is dynamically adjusted from 3×3 to 7×7 to suppress noise of different intensities in a directional manner to obtain enhanced image data. The enhanced image data is subjected to grayscale histogram analysis and processing to count the pixel distribution of each grayscale level in the image. In a 12-bit image (0-4095 grayscale levels), the number of pixels at each grayscale level is calculated to generate histogram data. For example, in a metal surface image, the background area is concentrated in the grayscale of 2000-2500, and the defect area is distributed in the grayscale of 1000-1500. Grayscale distribution data is obtained based on this distribution feature. The grayscale distribution is remapped through the histogram equalization algorithm, and the concentrated grayscale values ​​are expanded to a wider range to enhance the image contrast. The specific operation is to calculate the cumulative distribution function, map the original grayscale value to the new grayscale range, and obtain contrast enhancement data.

[0084] The contrast enhancement data is spectrally separated to separate the image into multiple spectral components according to different wavelength ranges. In this scheme, the visible spectrum range of 380-780nm is divided into three bands: red light (620-780nm), green light (495-570nm), and blue light (380-495nm), plus the near-infrared band (780-1100nm), to obtain multi-band image data. The image data of each band is subjected to a two-dimensional Fourier transform to convert the spatial domain information into the frequency domain, and the energy distribution of each frequency component is analyzed to obtain the spectral component data. The spectral component data is assigned band weights, and the weight coefficients are determined according to the contribution of different bands to the target features. For example, in urban night traffic monitoring, the near-infrared band contributes more to the recognition of vehicles and pedestrians in dark light environments, and is assigned a weight of 0.4; the visible light band is more sensitive to the recognition of key information such as license plates and road markings, and the three visible light bands are each assigned a weight of 0.2 to obtain band weight data. The wavelet transform algorithm is used to perform multi-scale decomposition on the weight data, and the Daubechies wavelet basis is used for three-layer decomposition to calculate the low-frequency approximate coefficients and high-frequency detail coefficients to generate fusion coefficient data.

[0085] The fusion coefficient data and multi-band image data are subjected to energy superposition processing, and the image information of different bands is linearly combined according to the weight coefficient. The value of each pixel is determined by the weighted sum of the four bands. The energy normalization of each band needs to be considered to avoid saturation of the superimposed image. During the superposition process, the image of each band is first normalized and then superimposed according to the weight to obtain the spectral superposition data. Subsequently, energy balance processing is performed to ensure that the overall brightness and contrast of the image are moderate through histogram stretching and brightness adjustment, and balanced spectrum data is generated. Finally, the balanced spectrum data is inversely transformed to convert the information in the frequency domain back to the spatial domain. The inverse Fourier transform is used to restore the spectrum information to the spatial image while retaining the multispectral features after fusion. The spatial domain image is color mapped to map the multispectral information to the RGB color space to generate visually clear multispectral composite image data.

[0086] Taking urban night road monitoring as an example, the monitoring area covers a 10m×10m road section, and a 4K camera is used for image acquisition. There is Gaussian noise in the original image, with a standard deviation of about 50 gray levels. Through noise separation and suppression processing, the noise level is reduced to less than 10 gray levels. The original grayscale distribution of the image is concentrated in the range of 1800-2200, and after histogram equalization, it is expanded to the range of 1000-3000, which significantly improves the image contrast. In spectral separation, the near-infrared band (800-1000nm) reflects the target contour information in low-light environments, the red light band (620-780nm) highlights the characteristics of vehicles and pedestrians, and the green and blue light bands provide road markings and environmental details. Through wavelet transform and energy superposition, the characteristic information of each band is effectively integrated, and the multi-spectral composite image finally generated clearly shows the multi-dimensional characteristics of the night scene, which can realize the accurate identification of vehicles and pedestrians, and the target positioning accuracy reaches the centimeter level.

[0087] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0088] (1) performing scale decomposition processing on the multispectral composite image data to obtain multiscale image data, and performing feature point detection processing on the multiscale image data to obtain feature point set data;

[0089] (2) performing descriptor calculation processing on the feature point set data to obtain feature description data, and performing hierarchical clustering processing on the feature description data to obtain multi-level feature data;

[0090] (3) The multi-level feature data are graded according to their importance to obtain feature level data, and the feature level data are entropy calculated to obtain data redundancy data;

[0091] (4) performing threshold segmentation processing on the data redundancy data to obtain compression level data, and performing hierarchical coding processing on the compression level data to obtain hierarchical coding data;

[0092] (5) encoding the hierarchical coded data of the high-importance level using a lossless compression algorithm to obtain lossless compressed data, and compressing the hierarchical coded data of the low-importance level using a lossy compression algorithm to obtain lossy compressed data;

[0093] (6) Performing data packet organization processing on the lossless compressed data and the lossy compressed data to obtain compressed feature data, and performing index structure construction processing on the compressed feature data to obtain feature index data.

[0094] Specifically, the multispectral composite image data is scaled and decomposed. The Gaussian pyramid decomposition method is used to downsample the original image layer by layer to generate images at multiple scale levels. The mathematical expression of scale decomposition is:

[0095] L(x,y,σ)=G(x,y,σ)*I(x,y)

[0096] Where L(x,y,σ) represents the Gaussian blurred image of the σth layer, G(x,y,σ) is the Gaussian kernel function, I(x,y) is the input image, and * represents the convolution operation. Through 4-layer decomposition, multi-scale image data with original resolution (4096×3072), 1 / 2 resolution, 1 / 4 resolution, and 1 / 8 resolution are obtained. FAST corner point detection is performed on the image of each scale layer to extract feature points. The FAST algorithm identifies corner point features by calculating the grayscale difference between 16 pixels on the circumference and the center point, and obtains feature point set data.

[0097] The descriptor is calculated for the feature point set data using the BRIEF descriptor algorithm. The calculation formula of the BRIEF descriptor is:

[0098] τ(p;x,y)=1,if I(x) <I(y);

[0099] τ(p;x,y)=0,otherwise;

[0100] Where p represents a feature point, and x and y are pairs of sampling points around the feature point. For each feature point, the comparison results of 256 pairs of sampling points are calculated to generate a 256-bit binary description vector to form feature description data. These feature description data are grouped using a hierarchical clustering algorithm, and the cluster distance calculation uses the Hamming distance:

[0101] H(di,dj)=∑|bi,k-bj,k|

[0102] Among them, di and dj are two descriptors, bi,k and bj,k are the corresponding binary bits, and multi-level feature data is obtained.

[0103] The characteristic points are graded according to their response strength R. The response strength calculation formula is:

[0104] R=det(M)-e·trace(M)2

[0105] Where M is the Harris matrix around the feature point, e is the empirical coefficient (value is 0.04), det(M) represents the determinant of the matrix M, trace(M) represents the trace of the matrix M, and the features are sorted from large to small according to the R value, and divided into three levels: high, medium, and low to obtain the feature level data.

[0106] Calculate the information entropy for each level:

[0107]

[0108] where p i is the probability of the i-th bit appearing in the feature descriptor, and the data redundancy data is obtained.

[0109] The compression threshold is set based on the entropy value. Features with an entropy value higher than 0.8 are classified as important features and require lossless compression; features with an entropy value between 0.5 and 0.8 allow lossy compression; features with an entropy value lower than 0.5 can use lossy compression with a high compression ratio to obtain compression level data. A hierarchical coding strategy is used for different compression levels:

[0110]

[0111] Among them, C(f) represents the final encoding result of feature f, B(f) represents the binary encoding of feature f, E(f) represents the entropy value of feature f, Q represents the quantization function, q1 represents the quantization parameter of medium-importance features (the quantization step is 1 / 8 of the original data), and q2 represents the quantization parameter of low-importance features (the quantization step is 1 / 16 of the original data). Huffman coding is used for lossless compression of high-importance levels. The average code length of Huffman coding is calculated as:

[0112] L=∑(Li·Hi)

[0113] Where Li is the probability of symbol occurrence, Hi is the encoding length, and lossless compressed data is obtained. The JPEG2000 algorithm is used for lossy compression of low importance levels, and the compression ratio can reach 20:1, obtaining lossy compressed data. Finally, the compressed data is organized into a data packet format: the packet header contains metadata such as feature level and compression method, and the data body contains compressed feature data to obtain compressed feature data. An index structure based on B+ tree is established, with the index key value being the spatial position and importance level of the feature, and the leaf node storing the physical address of the compressed data to obtain feature index data.

[0114] Taking the monitoring of urban main roads as an example, the original 4K image size is 48MB, and 15MB of multi-scale data is obtained through 4-layer pyramid decomposition. The FAST algorithm detects 5,000 feature points, and the BRIEF descriptor generates 1.28MB of feature data. After hierarchical clustering, three levels of features are obtained: 1,000 key features (entropy value>0.8, mainly including important targets such as vehicles and pedestrians), 2,000 secondary features (0.5≤entropy value≤0.8, including scene elements such as road markings and traffic signs), and 2,000 common features (entropy value<0.5, mainly background environment information). Key features are compressed to 250KB using Huffman coding, secondary features are compressed to 200KB using JPEG2000, and common features are compressed to 100KB. Finally, compressed feature data of 600KB is generated, realizing efficient storage and fast retrieval of surveillance video feature data.

[0115] The above describes the image acquisition method based on intelligent high-speed machine vision in the embodiment of the present application. The following describes the image acquisition device based on intelligent high-speed machine vision in the embodiment of the present application. Figure 2 In the embodiment of the present application, an image acquisition device based on intelligent high-speed machine vision comprises:

[0116] The acquisition module 201 is used to collect and process the multi-point light intensity of the detection area to obtain discrete light intensity data, and perform light mapping processing on the discrete light intensity data to obtain continuous light distribution data and initial exposure parameter data;

[0117] The processing module 202 is used to perform multi-spectral analysis processing on the continuous illumination distribution data and the initial exposure parameter data to obtain basic lighting configuration data, and perform timing control processing on the basic lighting configuration data to obtain multi-spectral lighting sequence data;

[0118] The trigger module 203 is used to perform trigger timing processing on the multi-spectral illumination sequence data and the initial exposure parameter data to obtain trigger timing data, and perform cache allocation processing on the trigger timing data to obtain memory pre-allocated data;

[0119] A division module 204 is used to perform region division processing on the trigger timing data and the memory pre-allocated data to obtain multi-region configuration data, and perform parallel acquisition processing on the multi-region configuration data to obtain original multi-region image data;

[0120] The enhancement module 205 is used to perform parallel enhancement processing on the original multi-region image data to obtain enhanced image data, and perform spectral fusion processing on the enhanced image data to obtain multi-spectral composite image data;

[0121] The extraction module 206 is used to perform feature extraction processing on the multi-spectral composite image data to obtain multi-level feature data, and perform hierarchical compression processing on the multi-level feature data to obtain compressed feature data and feature index data.

[0122] Through the cooperation of the above components, the technical solutions of multi-spectral analysis and timing control processing are adopted to realize the coordinated work of multiple bands of light sources, which not only improves the integrity of urban monitoring image information, but also avoids interference between different bands, and effectively solves the image acquisition problems in complex scenes such as day and night alternation and weather changes. Through the cooperation of trigger timing processing and cache allocation processing, an accurate data acquisition and storage mechanism is established to ensure the integrity of data in the high-speed acquisition process, and meet the needs of all-weather uninterrupted acquisition in urban video surveillance. The combination of regional division processing and parallel acquisition processing significantly improves the acquisition efficiency of the system, making large-scale and high-resolution urban road monitoring possible, and can cover multiple intersections, streets and other key areas at the same time. The application of parallel enhancement processing and spectral fusion processing not only improves the quality of monitoring images, but also realizes the effective integration of multiple spectral band information, enhances the information richness of images, especially the target recognition ability at night and in bad weather conditions. The innovative application of feature extraction processing and layered compression processing not only ensures the accurate preservation of important target feature information such as vehicles and pedestrians, but also optimizes the storage space of massive monitoring data, achieving a good balance between data processing efficiency and storage efficiency.

[0123] The present application also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the image acquisition method based on intelligent high-speed machine vision.

[0124] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0125] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk and other media that can store program codes.

[0126] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An image acquisition method based on intelligent high-speed machine vision, characterized in that: The image acquisition method based on intelligent high-speed machine vision includes: Collecting and processing the multi-point light intensity of the detection area to obtain discrete light intensity data, and performing light mapping processing on the discrete light intensity data to obtain continuous light distribution data and initial exposure parameter data; Performing multispectral analysis processing on the continuous illumination distribution data and the initial exposure parameter data to obtain basic lighting configuration data, and performing timing control processing on the basic lighting configuration data to obtain multispectral lighting sequence data; Performing trigger timing processing on the multi-spectral illumination sequence data and the initial exposure parameter data to obtain trigger timing data, and performing cache allocation processing on the trigger timing data to obtain memory pre-allocated data; Performing region division processing on the trigger timing data and the memory pre-allocated data to obtain multi-region configuration data, and performing parallel acquisition processing on the multi-region configuration data to obtain original multi-region image data; Performing parallel enhancement processing on the original multi-region image data to obtain enhanced image data, and performing spectral fusion processing on the enhanced image data to obtain multi-spectral composite image data; The multi-spectral composite image data is subjected to feature extraction processing to obtain multi-level feature data, and the multi-level feature data is subjected to layered compression processing to obtain compressed feature data and feature index data.

2. The image acquisition method based on intelligent high-speed machine vision according to claim 1 is characterized in that: The multi-point light intensity of the detection area is collected and processed to obtain discrete light intensity data, and the discrete light intensity data is subjected to light mapping processing to obtain continuous light distribution data and initial exposure parameter data, including: Performing grid partitioning processing on the detection area to obtain multi-point sampling coordinate data, and performing photoelectric signal acquisition processing on each sampling point in the multi-point sampling coordinate data to obtain discrete light intensity data; Performing signal denoising processing on the discrete light intensity data to obtain filtered light intensity data, and performing spatial mapping processing on the filtered light intensity data through a bilinear interpolation algorithm to obtain initial light distribution data; Performing surface fitting processing on the initial illumination distribution data by least square method to obtain continuous illumination distribution surface data, and performing interval division processing on the continuous illumination distribution surface data to obtain illumination level data; Performing parameter mapping processing on the light level data to obtain sensitivity parameter data and shutter speed parameter data, and combining the sensitivity parameter data and shutter speed parameter data to obtain basic exposure parameter data; Performing compensation processing on the basic exposure parameter data to obtain compensation coefficient data, and synthesizing the compensation coefficient data and the basic exposure parameter data to obtain target exposure parameter data; Parameter fusion is performed on the target exposure parameter data and the continuous illumination distribution surface data to obtain continuous illumination distribution data and initial exposure parameter data.

3. The image acquisition method based on intelligent high-speed machine vision according to claim 1 is characterized in that: The performing multi-spectral analysis processing on the continuous illumination distribution data and the initial exposure parameter data to obtain basic illumination configuration data, and performing timing control processing on the basic illumination configuration data to obtain multi-spectral illumination sequence data, includes: Performing spectrum analysis processing on the continuous illumination distribution data to obtain spectrum distribution data, and performing band division processing on the spectrum distribution data to obtain multi-band illumination data; Combining the multi-band illumination data and the initial exposure parameter data to obtain band weight data, and performing energy calculation processing on the band weight data using a spectral energy allocation algorithm to obtain light source energy allocation data; Performing pulse width modulation processing on the light source energy allocation data to obtain light source modulation signal data, and performing period segmentation processing on the light source modulation signal data to obtain basic lighting configuration data; Performing microsecond-level timing tag processing on the basic lighting configuration data to obtain time stamp sequence data, and performing staggered arrangement processing on the time stamp sequence data to obtain light source switching timing data; Performing synchronization signal generation processing on the light source switching timing data to obtain trigger synchronization data, and combining the trigger synchronization data and the light source switching timing data to obtain timing control data; The timing control data is subjected to time slice allocation processing to obtain spectrum acquisition sequence data, and the spectrum acquisition sequence data is subjected to verification processing to obtain multi-spectrum lighting sequence data.

4. The image acquisition method based on intelligent high-speed machine vision according to claim 1 is characterized in that: The trigger timing processing is performed on the multi-spectral illumination sequence data and the initial exposure parameter data to obtain the trigger timing data, and the cache allocation processing is performed on the trigger timing data to obtain the memory pre-allocated data, including: Performing time window segmentation processing on the multi-spectral illumination sequence data to obtain reference time window data, and performing trigger interval calculation processing on the reference time window data to obtain trigger interval sequence data; Performing synchronous calibration processing on the trigger interval sequence data and the initial exposure parameter data to obtain external trigger reference data, and performing pulse edge detection processing on the external trigger reference data to obtain coarse positioning trigger data; Performing phase compensation processing on the coarse positioning trigger data to obtain trigger compensation data, and performing fine-tuning processing on the trigger compensation data through a timing compensation algorithm to obtain precise trigger data; Performing trigger priority allocation processing on the precise trigger data to obtain trigger priority data, and performing trigger timing combination processing on the trigger priority data to obtain trigger timing data; Performing storage space calculation processing on the trigger timing data to obtain cache capacity data, and performing storage unit division processing on the cache capacity data to obtain storage block data; The storage block data is processed by data stream channel allocation to obtain cache channel data, and the cache channel data is processed by storage address mapping to obtain memory pre-allocated data.

5. The image acquisition method based on intelligent high-speed machine vision according to claim 1, characterized in that: The trigger timing data and the memory pre-allocated data are processed by area division to obtain multi-area configuration data, and the multi-area configuration data are processed by parallel acquisition to obtain original multi-area image data, including: Performing time domain segmentation processing on the trigger timing data to obtain acquisition timing segmentation data, and performing matching processing on the acquisition timing segmentation data and memory pre-allocated data to obtain time-space corresponding data; Performing a region-of-interest division process on the time-space corresponding data to obtain region boundary data, and performing a block process on the region boundary data using a region segmentation algorithm to obtain region block data; Performing acquisition resource configuration processing on the regional block data to obtain regional resource data, and performing channel allocation processing on the regional resource data to obtain parallel channel data; Performing time slot allocation processing on the parallel channel data to obtain parallel acquisition time series data, and performing thread scheduling processing on the parallel acquisition time series data to obtain multi-region configuration data; Performing acquisition processing on each region in the multi-region configuration data through an independent data stream to obtain regional sub-image data, and performing cache writing processing on the regional sub-image data to obtain cache image data; The cached image data is subjected to data stream merging processing to obtain merged image data, and the merged image data is subjected to image reorganization processing to obtain original multi-region image data.

6. The image acquisition method based on intelligent high-speed machine vision according to claim 1, characterized in that: The performing parallel enhancement processing on the original multi-region image data to obtain enhanced image data, and performing spectral fusion processing on the enhanced image data to obtain multi-spectral composite image data, comprises: Performing signal-noise separation processing on each region in the original multi-region image data to obtain regional noise feature data, and performing noise suppression processing on the regional noise feature data to obtain enhanced image data; Performing grayscale histogram analysis processing on the enhanced image data to obtain grayscale distribution data, and performing contrast adjustment processing on the grayscale distribution data through a histogram equalization algorithm to obtain contrast enhancement data; Performing spectral separation processing on the contrast enhancement data according to the wavelength range to obtain multi-band image data, and performing frequency domain transformation processing on the multi-band image data to obtain spectral component data; Performing band weight allocation processing on the spectral component data to obtain band weight data, and performing coefficient calculation processing on the band weight data through a wavelet transform algorithm to obtain fusion coefficient data; Performing energy superposition processing on the fusion coefficient data and the multi-band image data to obtain spectral superposition data, and performing energy balance processing on the spectral superposition data to obtain balanced spectrum data; The balanced spectrum data is subjected to inverse transformation processing to obtain spatial domain image data, and the spatial domain image data is subjected to color mapping processing to obtain multi-spectral composite image data.

7. The image acquisition method based on intelligent high-speed machine vision according to claim 1, characterized in that: The step of performing feature extraction processing on the multi-spectral composite image data to obtain multi-level feature data, and performing layered compression processing on the multi-level feature data to obtain compressed feature data and feature index data includes: Performing scale decomposition processing on the multi-spectral composite image data to obtain multi-scale image data, and performing feature point detection processing on the multi-scale image data to obtain feature point set data; Performing descriptor calculation processing on the feature point set data to obtain feature description data, and performing hierarchical clustering processing on the feature description data to obtain multi-level feature data; The multi-level feature data are graded according to their importance to obtain feature level data, and the feature level data are entropy calculated to obtain data redundancy data; Performing threshold segmentation processing on the data redundancy data to obtain compression level data, and performing hierarchical coding processing on the compression level data to obtain hierarchical coding data; The hierarchical coded data of the high-importance level is coded by a lossless compression algorithm to obtain lossless compressed data, and the hierarchical coded data of the low-importance level is compressed by a lossy compression algorithm to obtain lossy compressed data; The lossless compressed data and the lossy compressed data are processed by data packet organization to obtain compressed feature data, and the compressed feature data are processed by index structure construction to obtain feature index data.

8. An image acquisition device based on intelligent high-speed machine vision, used to implement the image acquisition method based on intelligent high-speed machine vision as described in any one of claims 1 to 7, characterized in that: The image acquisition device based on intelligent high-speed machine vision comprises: An acquisition module is used to collect and process the light intensity of multiple points in the detection area to obtain discrete light intensity data, and to perform light mapping processing on the discrete light intensity data to obtain continuous light distribution data and initial exposure parameter data; a processing module, configured to perform multispectral analysis processing on the continuous illumination distribution data and the initial exposure parameter data to obtain basic illumination configuration data, and perform timing control processing on the basic illumination configuration data to obtain multispectral illumination sequence data; A trigger module, used for performing trigger timing processing on the multi-spectral illumination sequence data and the initial exposure parameter data to obtain trigger timing data, and performing cache allocation processing on the trigger timing data to obtain memory pre-allocated data; A division module, used for performing region division processing on the trigger timing data and the memory pre-allocated data to obtain multi-region configuration data, and performing parallel acquisition processing on the multi-region configuration data to obtain original multi-region image data; An enhancement module, used for performing parallel enhancement processing on the original multi-region image data to obtain enhanced image data, and performing spectral fusion processing on the enhanced image data to obtain multi-spectral composite image data; The extraction module is used to perform feature extraction processing on the multi-spectral composite image data to obtain multi-level feature data, and perform layered compression processing on the multi-level feature data to obtain compressed feature data and feature index data.

9. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the image acquisition method based on intelligent high-speed machine vision as described in any one of claims 1 to 7 is implemented.

Images