Full-color low-light night vision security camera

Abstract

The invention discloses a full-color low-light night vision camera, which comprises a shell, a lens, a high-definition image sensor, an image transmission module, an intelligent monitoring module and a full-color low-light night vision imaging system, and is characterized in that the full-color low-light night vision imaging system is electrically connected with the high-definition image sensor and the image transmission module respectively; the intelligent monitoring module is electrically connected with the high-definition image sensor and the far-end intelligent terminal. The full-color low-light night vision imaging system executes the following imaging steps that background stray light is separated based on a preset algorithm; performing enhancement processing on the separated target image; stray light interference is reduced based on a background suppression technology; high-contrast full-color image reconstruction under a complex background is realized; and executing a target identification optimization step on the reconstructed image. According to the invention, the problem of difficult target identification caused by background stray light interference under a complex background condition can be solved.

Description

Technical Field

[0001] The invention relates to the technical field of night vision imaging, in particular to a full-color low-light-level night vision security camera. Background Art

[0002] The full-color low-light level night vision imaging method is a technology that uses low-intensity light, including faint starlight, moonlight and other weak light sources in the night environment, to achieve clear color imaging without using an active infrared light source. This technology enhances these existing light sources and combines advanced signal processing and image processing algorithms to generate bright color visual effects. It can provide images that are closer to natural colors at night or in dim environments. However, this method has obvious shortcomings when facing complex background conditions, especially when the background contains a large amount of stray light, such as non-target light emitted by street lamps, car taillights or reflective objects, which will seriously interfere with the night vision device's ability to capture and resolve targets, making the target information in the image blurred and difficult to identify, thereby causing challenges in target recognition. Summary of the invention

[0003] The purpose of the present invention is to provide a full-color low-light-level night-vision security camera to overcome the above-mentioned deficiencies in the prior art.

[0004] To achieve the above technical objectives, the technical solution of the present invention is implemented as follows:

[0005] A full-color low-light-level night vision imaging method, the method comprising the following steps:

[0006] Step 1: Separate background stray light based on a preset algorithm;

[0007] Step 2: Enhance the separated target image;

[0008] Step 3: Reduce stray light interference based on background suppression technology;

[0009] Step 4: Reconstruct high-contrast full-color images under complex backgrounds;

[0010] Step 5: Perform target recognition optimization steps on the reconstructed image.

[0011] Further, step one includes the following steps:

[0012] A convolutional neural network is used to analyze the intensity of stray light in the image;

[0013] Apply Gaussian mixture model to separate background stray light components from input images;

[0014] Adjust the accuracy of stray light segmentation based on adaptive threshold technology;

[0015] Identify the background stray light part based on the condition that the difference between the target area and the stray light intensity exceeds the threshold Ic, where I represents the stray light intensity, Ig represents the target image intensity, and Ic is a constant threshold. If I > Ig + Ic, it is determined that the background area contains significant stray light.

[0016] Further, the steps of applying the Gaussian mixture model to separate the background stray light component from the input image include:

[0017] Pre-set the working range of the image segmentation algorithm before separating the foreground target and the background stray light layer;

[0018] Traverse each pixel position in a specific order and update according to the likelihood score of the pixel belonging to the stray light layer at that position;

[0019] Based on the probability weight calculation model of each region in the iterative process, dynamically adjust the proportion weight value W of stray light in each sub-region until it falls within the set range;

[0020] Based on the condition that the overlapping ratio of the area of the target region and the background stray light region is lower than the threshold Smax, confirm the separation accuracy, where Starget represents the size of the effective region of the target image, Sscatter is the area of the identified stray light region. If Sscater / Starget < Smax holds, it is determined that the separation process is accurately completed.

[0021] Further, the steps of traversing each pixel position in a specific order and updating according to the likelihood score of the pixel belonging to the stray light layer at that position include:

[0022] Initialize the pixel matrix and assign a preliminary probability distribution;

[0023] Perform multi-dimensional fitting calculations on the positions of each undetermined element;

[0024] Based on the condition that the brightness change rate of the current element exceeds the estimated value, adjust the corresponding value in the probability weight matrix, that is, △B = Bcur - Bright_mean and |△B| > ε, where Bcur is the measured brightness value at the current position, Bright_mean is its average theoretical value, and ε represents the sensitivity coefficient, indicating that if the absolute brightness deviation is greater than a certain set level, the ownership of this point needs to be re-estimated;

[0025] Based on the condition that the probability that this point is correctly labeled as belonging to the target object rather than the stray light source is higher than the preset confidence level δ, it is determined that the pixel point has been correctly processed and further evaluation is stopped.

[0026] Further, when performing multi-dimensional fitting calculations, the consistency between the current position and the pixels in its surrounding neighborhood needs to be considered simultaneously. The calculation method is as follows:

[0027] Collect information about each pixel to be evaluated and its immediate neighbors;

[0028] Map the collected data to four-dimensional space for similarity measurement and output the results to a new table;

[0029] The highest likelihood value is selected as the best fit value and the initial assessment grade or label is revised based on it;

[0030] Based on the local connectivity Connectness of the pixel point and the consistency coefficient ratio Ratio = Brightness_similarity_score / Noise_level_score, if the Connectivity is judged to be sufficient and the ratio is within the predetermined threshold λmin, it is identified as a good candidate and added to the further review list.

[0031] Furthermore, in order to avoid oversimplification or overcomplexity in the local connectivity analysis process, which may lead to recognition errors, the following design measures from the underlying logic are taken:

[0032] Perform a depth search on the selected area to confirm that no isolated pixels remain;

[0033] A binary logic gate circuit structure is used to check whether adjacent pixels in any specified area meet the established conditional connection state;

[0034] Only when the relationship between all units within the classification block to which the pixel belongs is maintained at the same level can it be judged to form a valid cluster and be regarded as a complete independent entity. Here, relationship means a kind of mutual connection or closeness index determined by comparing visual elements, including brightness and hue;

[0035] If the calculated result is higher than the preset threshold, the formula

[0036] Relativity_Ratio = Similarity(S_i, S_j) / Average_difference(D_i+D_k) is used to quantify and evaluate the degree of association between two continuous or indirectly connected parts, where S_i and S_j represent the general similarity scores between adjacent sample sets, and D_i and D_k represent the brightness difference components of the two relative to the global background.

[0037] A full-color low-light-level night vision camera comprises a housing, a lens, a high-definition image sensor, an image transmission module and an intelligent monitoring module, characterized in that it also comprises a full-color low-light-level night vision imaging system executed by the method as described above, the full-color low-light-level night vision imaging system is electrically connected to the high-definition image sensor and the image transmission module respectively, and the intelligent monitoring module is electrically connected to the high-definition image sensor and a remote intelligent terminal respectively.

[0038] Beneficial effects of the present invention: The present invention can solve the problem of difficulty in target recognition caused by background stray light interference under complex background conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flow chart of the full-color low-light-level night vision imaging method;

[0040] Figure 2 is a flow chart of steps for separating background stray light based on a preset algorithm;

[0041] Figure 3 is a flow chart of the steps of applying a Gaussian mixture model to separate background stray light components from an input image;

[0042] Figure 4 is a flowchart of the steps of traversing each pixel position in a specific order and updating the probability score of the pixel at that location belonging to the stray light layer;

[0043] Figure 5 It is a specific calculation flow chart when multi-dimensional fitting calculation needs to consider the consistency between the current position and its surrounding neighboring pixels at the same time;

[0044] Figure 6 It is a design flow chart from the underlying logic to ensure that the local connectivity analysis process avoids over-simplification or over-complexity that may lead to identification errors. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present invention are described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0046] like Figure 1 As shown, the full-color low-light-level night vision imaging method of the present invention involves obtaining high-quality color night vision images in low-light and stray light environments, ensuring that the target object and the environmental background can be accurately distinguished even under complex background conditions. This method mainly includes the following steps: separating the background stray light based on a preset algorithm; enhancing the separated target image; reducing the influence of stray light by means of background suppression technology; on this basis, improving the image contrast under complex backgrounds, thereby generating a clear and recognizable full-color image; and finally further analyzing and processing the obtained results to better identify the target.

[0047] The background stray light separation step refers to identifying and stripping off the extra light components from the surrounding uneven light sources from the input image through a pre-established mathematical model, ensuring that the remaining data is closer to the original light signal emitted or reflected by the real object. For example, an advanced hybrid photon-electronic algorithm can be used, which uses spectral differences and statistical characteristics of brightness distribution to calculate and locate the contribution of stray particles, thereby excluding them. At this stage, the software will use advanced light field analysis to identify background spots and remove them.

[0048] For example, it is particularly important to apply this method in a scene where pedestrians or animals are passing through a starlit forest area. Although starlight provides a weak light source, it also brings a certain proportion of noise; coupled with the swaying leaves and the changing tree outlines, the light and shadow changes pose a great challenge to the detection system. However, relying on the processing results of this stage, the noise level introduced by these elements can be greatly reduced.

[0049] Next, the processed images are upgraded and optimized to highlight the main details while making the images more saturated and with sharp outlines. Here, you can choose a multi-layer depth enhancement solution to adapt to different light sources. On the one hand, the color of the target area is enhanced to make it more obvious and easier to observe; at the same time, the spatial layering is increased by refining the texture and contour lines; you can even try to combine AI enhancement algorithms to improve resolution.

[0050] For example, in the same night forest shot, by adjusting to the complex structure of the natural world around the target (assuming it is a wild animal), we can clearly see that the image becomes brighter and the colors become more vivid. The details of the animal's fur are fully revealed, and every movement can be captured even when it runs quickly.

[0051] Background suppression is an important step, which aims to eliminate weak reflections and inappropriate light sources remaining in the scene. It can formulate strategies to minimize the impact range based on the previously obtained information. By intelligently distinguishing effective lighting paths to protect the integrity of useful signals while eliminating irrelevant parts, this function enables the system to cope with extreme conditions such as fog, rain and snow.

[0052] In one embodiment, if the above scenario is transferred to a patrol mission on a city street where it often rains and contains water droplets, this skill will be extremely useful and necessary; because rain will scatter light spots under the illumination of car lights; street lights will reflect on wet ground to cause spot effects; all these problems will reduce the visual effect and make pedestrians or other obstacles difficult to distinguish. However, this chaotic situation can be greatly alleviated with the help of background suppression.

[0053] This is followed by contrast enhancement against complex backgrounds to ensure that the boundaries between the various components of the generated image are significant and clearly visible, helping to distinguish different features and potential threats.

[0054] Specifically, if the test is conducted in a multi-purpose park night security monitoring project, there will be many dynamic changes: such as neon billboards in the distance, decorative lamps on the balconies of surrounding residents, and occasional meteors. These will interfere with the operation of the system. The purpose of target separation can be achieved by implementing specific filters or local brightness equalization mechanisms; each object can be clearly outlined no matter how small or blurry it is.

[0055] Finally, a further recognition and correction process is implemented for the improved full-color gamut night vision material - mainly relying on the built-in neural network recognition model to learn from a large number of past successful case experience data as a reference, and make the most accurate matching judgment for each newly acquired picture to ensure that any suspicious activities can be reported and confirmed in a timely manner, whether it is an invasion of personal privacy or just harmless stray animal activities.

[0056] In summary, the above process allows us to continuously monitor, understand and analyze important objects or phenomena even in unfavorable lighting environments - this is the scientific explanation of how to overcome numerous difficulties through layer-by-layer iteration in the process of full-color low-light night vision.

[0057] like Figure 2 As shown, the separation of background stray light based on a preset algorithm of the present invention includes several key links. The first step is to use a convolutional neural network (CNN) to analyze the intensity of stray light in the image. This stage mainly processes the pixels and their neighborhoods in the image to capture the spatial distribution characteristics of stray light and use it as the basis for preliminary judgment.

[0058] The next step is to apply a Gaussian mixture model (GMM) to the input image to further separate the background stray light components from the image. The Gaussian mixture model estimates the probability distribution of the background stray light component in a statistical sense to achieve the effect of accurately extracting this part of the content.

[0059] The next step is to use adaptive thresholding technology to adjust the accuracy of the above segmentation process. This means that the standard line used to separate background stray light and valid image information will be adjusted in real time according to different characteristics of the current image, such as brightness or contrast changes, to ensure that the separated stray light fragments are as close to the real situation as possible.

[0060] Specifically, a judgment condition is also used in this process, that is, whether the difference between the target area and the stray light intensity exceeds the set threshold \(Ic\). Here \(Ic\) is a constant, \(I\) represents the actual observed stray light intensity level, and \(Ig\) refers to the normal light intensity that the target itself should have. If the difference \(I>Ig+Ic\), it indicates that there is significant stray light pollution in this area.

[0061] To illustrate this point, in one embodiment, assuming that the visibility of faint objects in a surveillance video is to be enhanced, a pre-trained convolutional neural network is first used to identify areas in each frame that potentially contain stray light, and the extent of this impact is preliminarily quantified; a more accurate background model is then separated through GMM; and the segmentation boundaries in the image are adaptively adjusted to improve detection accuracy; when the system calculates that the \(I\) value at a certain point position is greater than the corresponding \(Ig+Ic\) (for example, when \(Ic\) is set to an empirical value of 10 under a specific environment), it can be determined that this is additional radiation caused by the background light source and is removed, making the subject more clearly visible.

[0062] like Figure 3 As shown, the step of applying the Gaussian mixture model to separate the background stray light component from the input image of the present invention includes: before separating the foreground target and the background stray light layer, it is necessary to pre-set the working range of the image segmentation algorithm. This step is intended to define the area of ​​interest or the concentration of key pixels during image processing, ensuring that the segmentation operation focuses on the valid pixels in a specific area rather than covering the entire image.

[0063] Traverse each pixel position and update it in real time based on the probability score of the pixel belonging to the stray light component. During this period, for each pixel, a statistical probability of the pixel belonging to the background or stray light is calculated, and the knowledge of the pixel attributes is gradually updated based on this score. This step helps to locate stray light more accurately and increase the possibility of it being correctly identified. For example, when processing full-color low-light night vision images, if the pixels in a certain area show consistent but unexpected color intensity or distribution patterns, then these pixels will receive higher scores, indicating that they may constitute part of the stray light component.

[0064] Then, the weight value W of the proportion of background stray light in individual sub-regions is dynamically adjusted according to the probability information in the iterative process. The sub-regions here can be understood as smaller fragments that are segmented out. After each iteration, the weights are adaptively changed based on the pixel-level evaluation results obtained in the previous step. The W value is usually controlled in the interval [0,1] and approaches the final stable value as the iteration progresses. Specifically in the actual scene, this means that as the algorithm runs, the possibility of pixels of different brightness or hue being correctly classified as stray light will be more accurate.

[0065] Verify the accuracy when the separation is completed. When the proportion of the overlapping part between the target area and the detected background stray light in the entire foreground area is lower than the predetermined maximum ratio \(S_{\text{max}}\), it is determined that the operation of this time meets the standard accurately. Here, the parameter \(S_{\text{target}}\) refers to the actual size of the foreground object, that is, the area of the meaningful main part in the image, and \(S_{\text{scatter}}\) is the size of the part separated and marked as stray light. The formula \(S_{\text{scatter}} / S_{\text{target}} < S_{\text{max}}\) checks whether the part determined as the background affects the existence of the main observed object to a large extent. The value of \(S_{\text{max}}\) is set to prevent too much useful information from being removed as stray light, resulting in the loss of useful signals or incomplete display of the image. In one implementation, \(S_{\text{max}}\) is set to be equal to 0.1 or lower, meaning that only a very small amount of interfering light affecting the main visual focus is accepted.

[0066] Taking an embodiment as an example, in the shooting of a night outdoor scene, assume that the camera captures the picture of a car in the distance and its surrounding environment. Due to the night vision enhancement feature, the camera may receive various interfering lights in the environment, such as the weak omnidirectional light emitted by street lamps. Applying the above algorithm will help distinguish the body and its contour of the car as the real target object and separate the various omnidirectional light marks in the environment, ensuring that the final picture presented to the observer minimizes the irrelevant stray factors. Through the above process, the contrast of the main observed object is effectively enhanced, and the imaging quality under night vision conditions is improved.

[0067] As Figure 4 shown, the step of traversing each pixel position in a specific order and updating according to the possibility score of the pixel belonging to the stray light layer of the present invention includes the following parts: First, initialize the pixel matrix and assign a preliminary probability distribution. This step aims to construct the basic structure and initial possibility of the image and prepare for subsequent processing. At the beginning, each pixel is assigned a set of probabilities of belonging to or not belonging to stray light set based on prior knowledge. These probability distributions will be used as the starting point for the subsequent steps and will be updated and improved in continuous iteration.

[0068] Next, a multidimensional fitting calculation is performed on the position of each element to be determined, with the goal of accurately measuring the relationship between the properties of the element at the current position and the surrounding environment. This step involves analyzing the state of each pixel relative to its surrounding elements in order to more accurately understand the probability that it belongs to background stray light. By examining the interaction between its neighboring pixels and the degree of consistency with the overall brightness, it is possible to better assess whether a certain area is image information of interest or insignificant stray light.

[0069] Subsequently, when dynamically adjusting the corresponding probability based on whether the brightness change rate at the current position exceeds a preset threshold, the formula \(\Delta B = B_{\text{cur}}-Bright_{\text{mean}}\), where \(\Delta B\) reflects the difference between the actual observed brightness value \(B_{\text{cur}}\) and the assumed uniform background value \(Bright_{\text{mean}}\). The coefficient \(\epsilon\) defines a sensitive limit: only when the brightness deviation exceeds the \(\epsilon\) threshold (such as within the range of 0.01 to 0.05), the system will recognize that this point may be interfered with or contain valid target information, and mark the higher probability as a useful signal in the probability matrix.

[0070] Finally, by ensuring that a higher certainty standard, namely the confidence level \(\delta\), is given, once the probability that a point belongs to valid information rather than regular background exceeds the threshold, the pixel is considered to have been fully evaluated and correctly classified, and additional testing can be stopped. This design ensures that only truly reliable detection results will be adopted into the final output image, reducing the chance of misclassification, thereby improving the quality of night vision images.

[0071] For example, in one embodiment, such an evaluation is performed on a video frame of a city scene captured under night conditions. The system initially makes a broad, hypothetical initial guess for all detected targets; as processing proceeds, the algorithm focuses on areas where the brightness is significantly different from the surrounding or expected model predictions. When a strong light spot caused by direct street lighting is identified, even if it may be mistaken for the object itself in the early stages, by adjusting the weights in subsequent steps, such interference items can be effectively distinguished and eliminated; on the contrary, for pedestrians and vehicles that do exist in low-light environments, their brightness deviation values ​​are not large and the consistency is high, so that a relatively clear and accurate picture effect can be obtained even in poor lighting conditions. Therefore, choosing an appropriate size such as \(\epsilon=0.03,\delta=0.95\) helps to balance the conflicting needs between sensitivity and reliability.

[0072] like Figure 5As shown, the multi-dimensional fitting calculation of the present invention needs to consider the consistency between the current position and its surrounding neighboring pixels at the same time, and the calculation method is as follows: collect information on each pixel to be evaluated and its direct neighbors; map the collected data to a four-dimensional space for similarity measurement and output the result to a new table; select the highest likelihood value as the best fitting value, and modify the initial evaluation level or label based on this; based on the local connectivity Connectness of the pixel point and the consistency coefficient ratio Ratio = Brightness_similarity_score / Noise_level_score, if the Connectness is judged to be sufficient and the ratio is within a predetermined threshold λmin, it is identified as a good candidate and added to the further review list.

[0073] The process of collecting information about each pixel to be evaluated and its immediate neighbors is primarily to ensure that the algorithm fully considers the current pixel being processed and its surroundings. By collecting the intensity of the pixel and other related information (such as color, etc.), the image data can be understood at a higher level. Specifically, in a low-light night vision environment, this means that in addition to the central pixel, the brightness and chromaticity information of its four adjacent pixels also needs to be read. In one embodiment, when capturing a color scene in low light, it is necessary to record the brightness level of each pixel and its four adjacent pixels, while considering the correlation metrics between them. Such information will be used to determine the common performance pattern of these pixels.

[0074] Subsequently, the above data is converted and embedded in a four-dimensional space for further processing. The purpose of this is to be able to use more dimensional expressions to enhance the connection between similar pixels, thereby improving noise detection and filtering capabilities. Specifically, pixels with small brightness differences but similar colors will be closer in this four-dimensional representation. When applied to night vision scenes in the above example, the four-dimensional measurement of brightness and noise can better capture the changing details of weak light sources.

[0075] After obtaining the quantification of similarity, the highest probability distribution is selected to obtain the closest match result, and the originally given location estimate or its attribute label is modified accordingly. In the context of the above case, it means selecting a representative result from all evaluated pixel combinations and possibly updating the original label status of this location; for example, a part originally considered to be a noise point may be changed to a part of the natural landscape based on the comprehensive evaluation of neighboring pixels.

[0076] The last step is to make an evaluation decision for each pixel at each position based on the local connectivity property Connectness and the comparison coefficient composed of pixel brightness similarity and noise level, namely Ratio. When the local connectivity is considered sufficient and the ratio Brightness_similarity_score / Noise_level_score is greater than a predetermined threshold λmin, it indicates that the potential value with higher quality is worth retaining for the next level of refinement analysis.

[0077] For example, if the central pixel of a certain image area in a low-light environment and the four directly adjacent points around it have similar low-frequency changes (well connected) and the degree of brightness consistency is greater than the noise fluctuation is higher than the minimum threshold value λmin specified by the system, it can be judged that the visual elements in this block are likely to represent valid scene structures rather than interfering noise sources, so they should be identified as high-quality data points for the ultimate presentation of clearer and more realistic nighttime color visual images. In addition, the optimal setting of λmin depends on the actual application requirements, but generally a higher setting can effectively reduce false alarms. For example, setting it near a certain empirical value of λmin = 2 or 3 may be a suitable choice. This ratio calculation method can help identify truly important image elements rather than being misled by random or recurring noise patterns. Through such a mechanism, the system can focus more on capturing actual signals rather than misinterpreting noise during processing.

[0078] Next, the present invention describes the following design measures from the underlying logic to ensure that the local connectivity analysis process avoids oversimplification or overcomplexity leading to recognition errors:

[0079] A deep search is performed on the selected area to confirm that no isolated pixels remain; this measure refers to a comprehensive and in-depth scan within a specific area of ​​interest to ensure that no pixel exists independently of other pixels. The deep search verifies connectivity by traversing each pixel to avoid isolated points or noise interfering with the accuracy of the final image recognition. Specifically, in a night vision camera application scenario, this step helps to filter out abnormal independent noise points or non-target related single bright spots that may be introduced into the night view due to sensor false alarms or other external disturbances, making the remaining parts clearer and more reliable.

[0080] like Figure 6As shown, a binary logic gate circuit structure is used to check whether the adjacent pixels in any specified area meet the established connection status; this measure relies on simple and effective logic gates, especially AND gate operations to achieve fast and accurate judgment. If parameters such as the contrast and brightness distribution between multiple consecutive pixels meet certain preset standards, it is considered that these pixels belong to the same object or are adjacent to each other. In one embodiment, this process is used to acquire targets under weak light, such as when animals move at night, and small changes in the dark background can be captured more accurately and the coherence and boundary consistency of these objects can be maintained, thereby improving the overall imaging system resolution and making it more useful and efficient for motion detection or other subsequent processing steps.

[0081] Then, only when the relationship between all units within the classification block to which the pixel belongs is maintained at the same level (Relativity_Level>=Threshold) can it be determined to form a valid cluster and be regarded as a complete independent entity. This operation means that the current cluster can only be considered to be a combination of pixels with high correlation and consistency when the evaluation results of brightness, hue and other measurable attributes reach a pre-specified baseline. Based on the above example, if the characteristic values ​​between the various parts of a suspected wild animal individual image reach or exceed the pre-defined trust threshold (such as small brightness difference, complete and coherent contours, etc., which are significantly better than the noise or non-target area of ​​the surrounding environment), it will be identified as a single entity instead of being split into fragmented and irregular small fragments.

[0082] Then by using the formula

[0083] Relativity_Ratio = Similarity(S_i, S_j) / Average_difference(D_i+D_k) is used to quantify and evaluate the closeness of the association between two continuous or indirectly connected parts. In this formula, the Similarity function estimates the level of dissimilarity or consistency between two given point sets (for example, continuous segment A and adjacent segment B), and D_i and D_k represent the deviation of their respective relative background average luminance levels. The Relativity_Ratio obtained by this ratio represents the degree to which the two are visually connected. If the ratio is high, it indicates that there is stronger evidence that this is a consistent part rather than a segmented area or an edge seam between different entities. The establishment of this formula is mainly based on empirical rules to measure the contrast intensity while taking into account the intrinsic pattern similarity, thereby improving the overall processing capabilities for night image segmentation tasks, especially those dealing with target recognition challenges in low-light scenes.

[0084] Assume that the values ​​of S_i and S_j are between 0 and 1, indicating that they are completely different when close to zero, and extremely consistent when close to one; while the values ​​of D_i and D_k of global brightness are negative numbers, indicating that the corresponding parts are relatively bright, and positive values ​​are the opposite. Ideally, the threshold is selected so that the formula calculation result is significantly higher than the average ratio of the background area (assuming it is greater than a certain threshold such as 0.6), so that the system can stably identify the target even in very low light conditions, reduce the probability of misclassification caused by over-abstraction or over-refinement, and optimize the accuracy of image analysis.

[0085] The full-color low-light-level night vision imaging method of the present invention includes the following key technical steps:

[0086] The first step is to efficiently separate the background stray light by using a specific preset algorithm during the image acquisition process. This separation step is one of the keys to the entire technical process. Through it, we can ensure that the stray light in the background is initially eliminated. This step can not only effectively reduce the amount and complexity of subsequent processing data, but also provide a pure foundation for further target recognition work.

[0087] In the second step, a series of advanced enhancement processing methods are applied to the target image area after it is extracted from the background stray light or separated from the stray light, aiming to improve the image quality, especially the display effect of object features in a weak light source environment.

[0088] The third step is to further strengthen the suppression of stray light interference with the help of innovative background suppression technology. The core of background suppression is to weaken or even remove the adverse factors caused by stray light without damaging or minimizing the damage to the main information of the target object, thereby effectively improving the visual readability and analytical usability of the image.

[0089] Through the first three measures, this method successfully overcomes the problems of image quality degradation and unstable target detection accuracy caused by stray light interference in the face of complex and changing backgrounds in existing technical solutions. The fourth step, based on the joint cooperation of the first three, realizes the high-quality, contrast-rich full-color image reconstruction process even in extremely harsh and complex working scenes, further improving the overall quality and readability of the image;

[0090] Finally, an important additional link is added to the entire imaging process - for the reconstructed picture formed by the aforementioned steps, the present invention introduces a set of optimization mechanisms, focusing on how to better implement target recognition for intensive training. This operation is intended to enable the final imaging system to not only accurately distinguish the boundaries between different individuals, objects or phenomena, but also to pay more attention to whether the algorithm model can be dynamically adjusted according to the specific situation, so that it can continuously adapt to the ever-changing application requirements and environmental challenges, so that the entire set of full-color low-light night vision imaging solutions can truly achieve a level of stability, reliability and intelligent characteristics. In this way, the interference problem caused by background stray light can be effectively improved in complex backgrounds. In summary, this series of technical design ideas and operating methods effectively respond to and overcome many difficulties and challenges often encountered in actual environments.

[0091] The present invention also discloses a full-color low-light-level night vision camera, comprising a housing, a lens, a high-definition image sensor, an image transmission module and an intelligent monitoring module, characterized in that it also comprises a full-color low-light-level night vision imaging system executed by the method as described above, the full-color low-light-level night vision imaging system being electrically connected to the high-definition image sensor and the image transmission module respectively, and the intelligent monitoring module being electrically connected to the high-definition image sensor and the remote intelligent terminal respectively.

[0092] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute the instructions), or in combination with these instruction execution systems, devices or apparatuses. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in combination with these instruction execution systems, devices or apparatuses. More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk box (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium and then editing, interpreting or processing in other suitable ways if necessary, and then stored in a computer memory.

[0093] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A full-color low-light-level night vision imaging method, characterized in that: The method includes the following steps: Step 1: Separate the background stray light based on a preset algorithm; Step 2: Enhance the separated target image; Step 3: Reduce the stray light interference based on background suppression technology; Step 4: Achieve the reconstruction of a high-contrast full-color image under a complex background; Step 5: Execute the target recognition optimization step on the reconstructed image.

2. The method according to claim 1, characterized in that Step 1 includes the following steps: Use a convolutional neural network to analyze the stray light intensity in the image; Apply a Gaussian mixture model to separate the background stray light component from the input image; Adjust the accuracy of stray light segmentation based on adaptive threshold technology; Identify the background stray light part based on the condition that the difference between the target area and the stray light intensity exceeds the threshold Ic, where I represents the stray light intensity, Ig represents the target image intensity, and Ic is a constant threshold. If I > Ig + Ic, it is determined that the background area contains significant stray light.

3. The method according to claim 2, characterized in that The step of applying a Gaussian mixture model to separate the background stray light component from the input image includes: Pre-set the working range of the image segmentation algorithm before separating the foreground target and the background stray light layer; Traverse each pixel position in a specific order and update according to the likelihood score of the pixel belonging to the stray light layer at that position; Based on the probability weight calculation model of each region in the iterative process, dynamically adjust the proportion weight value W of stray light in each sub-region until it falls within the set range; Confirm the separation accuracy based on the condition that the overlapping ratio of the area of the target region and the background stray light region is lower than the threshold Smax, where Starget represents the size of the effective region of the target image, Sscatter is the area of the identified stray light region. If Sscater / Starget < Smax holds, it is determined that the separation process is accurately completed.

4. The method according to claim 3, characterized in that The step of traversing each pixel position in a specific order and updating according to the likelihood score of the pixel belonging to the stray light layer at that position includes: Initialize the pixel matrix and assign a preliminary probability distribution; Perform multi-dimensional fitting calculations on the positions of each undetermined element; Adjust the corresponding value in the probability weight matrix based on the condition that the brightness change rate of the current element exceeds the estimated value, that is, △B = Bcur - Bright_mean and |△B| > ε, where Bcur is the measured value of the brightness at the current position, Bright_mean is its average theoretical value, and ε represents the sensitivity coefficient, indicating that if the absolute brightness deviation is greater than a certain set level, the ownership of this point needs to be re-estimated; It is determined that the pixel point has been correctly processed and further evaluation is stopped only when the probability that this point is correctly labeled as belonging to the target object rather than the stray light source is higher than the preset confidence level δ.

5. The method according to claim 4, characterized in that When performing multi-dimensional fitting calculations, the consistency between the current position and the pixels in its surrounding neighborhood needs to be considered simultaneously, and its calculation method is as follows: Collect the information of each pixel to be evaluated and its direct neighbors; Map the collected data to a four-dimensional space for similarity measurement and output the result to a new table; Select the highest likelihood value as the best fitting value and correct the initial evaluation level or label based on this. Based on the local connectivity Connectness of the pixel point and the consistency coefficient ratio Ratio = Brightness_similarity_score / Noise_level_score, if the Connectivity is judged to be sufficient and the ratio is within the predetermined threshold λmin, it is identified as a good candidate and added to the further review list.

6. The method according to claim 5, characterized in that In order to avoid oversimplification or overcomplexity in the local connectivity analysis process, which may lead to recognition errors, the following design measures from the underlying logic are taken: Perform a depth search on the selected area to confirm that no isolated pixels remain; A binary logic gate circuit structure is used to check whether adjacent pixels in any specified area meet the established conditional connection state; Only when the relationship between all units within the classification block to which the pixel belongs is maintained at the same level can it be judged to form a valid cluster and be regarded as a complete independent entity. Here, relationship means a kind of mutual connection or closeness index determined by comparing visual elements, including brightness and hue; If the calculated result is higher than the preset threshold, the formula Relativity_Ratio = Similarity(S_i, S_j) / Average_difference(D_i+D_k) is used to quantify and evaluate the degree of association between two continuous or indirectly connected parts, where S_i and S_j represent the general similarity scores between adjacent sample sets, and D_i and D_k represent the brightness difference components of the two relative to the global background.

7. A full-color low-light-level night vision camera, comprising a housing, a lens, a high-definition image sensor, an image transmission module and an intelligent monitoring module, characterized in that: It also includes a full-color low-light-level night vision imaging system executed by the method described in any one of claims 1-6, the full-color low-light-level night vision imaging system is electrically connected to the high-definition image sensor and the image transmission module, and the intelligent monitoring module is electrically connected to the high-definition image sensor and the remote intelligent terminal, respectively.

Images