Intelligent camera system based on image processing algorithm and optimization method

Abstract

The invention discloses an intelligent camera system and optimization method based on an image processing algorithm, and relates to the field of camera intelligent systems, and the system comprises an image collection module, a preprocessing module, a feature extraction module, a target recognition module, a decision module, a feedback control module, an illumination compensation module, an anomaly detection module and a data storage and analysis module. The method comprises the following steps: acquiring images in real time through a high-resolution sensor, denoising, enhancing and normalizing the images through preprocessing, generating feature vectors through LBP and HOG algorithms for feature extraction, identifying targets through a CNN model for target identification, generating instructions according to identification results and rules through decision making, adjusting camera parameters according to feedback control, and compensating to improve image quality through illumination compensation. The detection module detects an image abnormal area, and information collected by the data storage and analysis storage analyzes a mode rule. According to the method, clear images are collected under different illumination conditions, features are accurately extracted, targets are identified, intelligent decision is made, parameters are flexibly adjusted, abnormity can be detected, and analysis data can be stored to assist system optimization and upgrading.

Description

Technical Field

[0001] The present invention relates to the technical field of camera intelligent systems, and in particular to an intelligent camera system and an optimization method based on an image processing algorithm. Background Art

[0002] In today's digital and intelligent era, cameras have been widely used in many fields, such as security monitoring, intelligent transportation, industrial detection, smart home, etc. Traditional cameras mainly realize image acquisition and simple storage functions, which can no longer meet the growing demand for complex applications. The research and development of intelligent camera systems has become crucial.

[0003] In the field of security monitoring, with the continuous expansion of urban scale and the increasing demand for security, the monitoring scope is constantly expanding and the monitoring targets are becoming more diversified. Traditional cameras can only shoot at a fixed angle of view and cannot automatically adjust parameters according to changes in the monitoring scene. The detection and early warning capabilities of abnormal events are limited. When encountering complex lighting conditions, such as direct strong light, backlight, low light at night, etc., the captured images often have problems such as overexposure, underexposure, and blur, which greatly reduces the monitoring effect and makes it difficult to accurately identify the target object.

[0004] In intelligent transportation systems, cameras are used for vehicle detection, traffic flow statistics, and illegal behavior capture. However, existing cameras are prone to target loss and recognition errors when dealing with high-speed moving vehicles and complex traffic scenes. At the same time, changes in lighting during different time periods and weather conditions will also have a significant impact on image quality, affecting the accuracy and reliability of traffic data.

[0005] In terms of industrial inspection, the requirements for product quality inspection are getting higher and higher. Traditional cameras are difficult to accurately identify subtle defects and features of products and cannot meet the needs of high-precision inspection. In addition, in industrial production environments, there may be interference factors such as dust and vibration, which further reduce the working performance of the camera.

[0006] In the field of smart home, cameras are used for home security, elderly and child care, etc. However, traditional cameras lack intelligent interactive capabilities, cannot automatically adjust shooting modes and parameters according to user needs and scenarios, and cannot provide personalized services.

[0007] In terms of image processing algorithms, although there are many mature algorithms, a single algorithm can only solve specific problems and is difficult to cope with complex and changing actual scenarios. For example, a simple image denoising algorithm may blur the edge information of the image while removing noise; traditional target recognition algorithms have low recognition accuracy when facing complex backgrounds and similar targets.

[0008] In addition, the existing intelligent camera system lacks effective feedback mechanism and adaptive adjustment ability, and cannot adjust the camera parameters in real time according to image quality and actual needs. At the same time, there is a lack of effective storage and analysis methods for the large amount of image data collected, and it is impossible to fully explore the potential value in the data and provide support for system optimization and decision-making.

[0009] In summary, the development of an intelligent camera system and optimization method based on image processing algorithm has important practical significance, which can effectively solve the problems existing in the existing camera system and improve the intelligence level and application effect of the camera. Summary of the invention

[0010] The present invention proposes an intelligent camera system based on image processing algorithm to solve the problems mentioned in the above-mentioned prior art.

[0011] In order to achieve the above object, the present invention adopts the following technical solution: an intelligent camera system based on image processing algorithm, comprising: Image acquisition module: uses image sensors to collect image data in real time and collect images under different lighting conditions; Image preprocessing module: denoise, enhance and normalize the collected images. The denoising method uses the bilateral filtering algorithm, and the enhancement method uses the histogram equalization algorithm. The normalization formula is: , I is the original image pixel value, I min and I max are the minimum and maximum values ​​of image pixels respectively; Feature extraction module: Use the local binary pattern (LBP) and histogram of oriented gradients (HOG) algorithms to extract the texture and shape features of the image and generate the feature vector F; Target recognition module: classifies and recognizes the feature vector F to determine the target object in the image and its category C; the model includes convolutional layer, pooling layer and fully connected layer. The convolutional layer extracts the deep features of the image through the convolution kernel, the pooling layer performs dimensionality reduction processing on the feature map, and the fully connected layer maps the extracted features to the target category; Decision module: Generates decision instruction D based on target recognition results combined with preset rules and strategies. The preset rules and strategies take into account the category, quantity, location, and motion state of the target objects. When the number of target objects exceeds the threshold, adjusts the camera's shooting range. Feedback control module: adjust the camera parameters according to the decision instruction D, and the adjustment formula is P new =P old +ΔP,P old is the original parameter value, and ΔP is the parameter adjustment amount calculated according to the decision instruction.

[0012] Furthermore, it also includes: Light compensation module: Real-time detection of ambient light intensity L. The light compensation module uses a light sensor to measure the ambient light intensity. When the light intensity does not meet the preset range [L min ,L max ], through the formula Perform illumination compensation on the image, I is the original image pixel value, L ref is the reference light intensity.

[0013] Furthermore, it also includes: Anomaly detection module: detects abnormal areas in the image by calculating the entropy value H of the image. The anomaly detection module divides the image into blocks and calculates the entropy value of each image block. The entropy value calculation formula is: , p(i) is the probability of the pixel with gray value i appearing in the image; when the entropy value H exceeds the preset threshold H threshold When the image is abnormal, the system marks the abnormal area, zooms in for observation, and triggers an alarm mechanism.

[0014] Furthermore, the convolutional neural network (CNN) in the target recognition module uses a pre-trained model and is fine-tuned through transfer learning; the pre-trained model is trained on an image dataset to learn image features and patterns; during the transfer learning process, the parameters of the pre-trained model are used as initial values ​​and fine-tuned on the target dataset; the cross entropy loss function is used during training. , is the true label, is the predicted label.

[0015] Furthermore, the preset rules and strategies in the decision module adopt fuzzy logic algorithm, which fuzzifies the category C, size S and position P factors of the target object and maps them to different fuzzy sets; through the preset fuzzy inference rules, the corresponding decision instructions D are generated according to the input fuzzy variables.

[0016] Furthermore, the feedback control module has adaptive adjustment capability, and dynamically adjusts the parameter adjustment amount ΔP according to the image quality evaluation index by monitoring the adjusted image effect; and adopts the image clarity evaluation function, specifically the Laplace variance method and the contrast evaluation function, to evaluate the image quality in real time, and automatically adjust the camera parameters when the image quality does not meet the requirements.

[0017] Furthermore, it also includes: Data storage and analysis module: stores collected image data, feature vectors, target recognition results and decision instruction information. The data storage and analysis module uses hard disk or cloud storage devices to discover potential patterns and rules in the data through data analysis algorithms, specifically cluster analysis and association rule mining.

[0018] Furthermore, it also includes: Image acquisition steps: Use image sensors to collect image data in real time. The image sensors have a resolution of 4K and a frame rate of 60fps. Image preprocessing steps: denoising, enhancing and normalizing the collected images. The normalization formula is: , denoising uses bilateral filtering algorithm, and enhancement uses histogram equalization algorithm; Feature extraction step: Use the local binary pattern (LBP) and histogram of oriented gradients (HOG) algorithms to extract the texture and shape features of the image and generate the feature vector F; Target recognition step: Based on a deep learning model, specifically a convolutional neural network (CNN), the feature vector F is classified and recognized to determine the target object and its category C in the image. The CNN model is trained using pre-training and transfer learning methods; Decision-making step: Generate decision instruction D based on target recognition results and combined with preset rules and strategies. The preset rules and strategies use fuzzy logic algorithm; Feedback control step: According to the decision instruction D, the parameters of the camera are adjusted, specifically the focal length, aperture, and shooting angle. The adjustment formula is P new =P old +ΔP.

[0019] Furthermore, before the image preprocessing step, it also includes real-time detection of the ambient light intensity L; using a light sensor for detection, when the light intensity does not meet the preset range [L min ,L max ], through the formula Perform lighting compensation on the image.

[0020] Furthermore, after the target recognition step, the entropy value H of the image is calculated to detect abnormal areas in the image; the image is divided into blocks and the entropy value of each image block is calculated respectively. The entropy value calculation formula is: , when the entropy value H exceeds the preset threshold H threshold When an abnormality is detected in the image, the corresponding processing mechanism is triggered, including marking, magnification observation and alarm.

[0021] Compared with the prior art, the present invention has the following beneficial effects: In terms of image acquisition and preprocessing, the use of high-resolution image sensors can stably acquire clear, smear-free images under different lighting conditions. The image preprocessing module uses bilateral filtering and histogram equalization algorithms, combined with normalization processing, to effectively remove noise, enhance contrast, and improve image quality, providing a good data foundation for subsequent feature extraction and target recognition.

[0022] The illumination compensation module detects the ambient light intensity in real time and processes the image through the illumination compensation formula. It can improve the image brightness and contrast under different lighting conditions, avoid problems such as overexposure and underexposure, and ensure that clear and accurate images can be captured in complex environments such as strong light, weak light, and backlight.

[0023] The feature extraction module integrates the local binary pattern (LBP) and directional gradient histogram (HOG) algorithms to extract the texture and shape features of the image, generate more representative feature vectors, and improve the accuracy of target recognition. The target recognition module is based on the deep learning convolutional neural network (CNN), and uses pre-training and transfer learning methods to quickly and accurately identify the target objects and their categories in the image.

[0024] The decision module uses fuzzy logic algorithms, combined with preset rules and strategies, to generate flexible and intelligent decision instructions based on the target recognition results. The feedback control module accurately adjusts the camera parameters according to the decision instructions, and has adaptive adjustment capabilities. It can dynamically optimize parameters according to image quality evaluation indicators to ensure that the shooting effect always meets the requirements.

[0025] The anomaly detection module detects abnormal areas by calculating the entropy value of the image. When the entropy value exceeds the preset threshold, the anomaly is determined in time and the corresponding mechanism is triggered, which improves the security and reliability of the system.

[0026] The data storage and analysis module stores and analyzes collected image data, feature vectors, target recognition results, decision instructions and other information, and uses algorithms such as cluster analysis and association rule mining to discover potential patterns and rules in the data, providing a strong basis for system optimization and decision-making, and realizing continuous improvement and upgrading of the intelligent camera system. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A schematic block diagram of an intelligent camera system based on an image processing algorithm proposed by the present invention; Figure 2 A schematic block diagram of an intelligent camera method based on an image processing algorithm proposed by the present invention; Figure 3 Schematic diagram of image quality comparison under different light intensities; Figure 4 This is a schematic diagram showing how the target recognition accuracy changes with the number of training rounds; Figure 5 It is a schematic diagram of the relationship between the anomaly detection entropy value and the anomaly determination result; Figure 6 Schematic diagram for comparing the effects of feedback control parameter adjustment. DETAILED DESCRIPTION

[0028] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0029] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0030] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In addition, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, and it can be the internal connection of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0031] Reference Figures 1 to 6 :An intelligent camera system based on image processing algorithm, comprising: Image acquisition module: Equipped with a high-resolution image sensor, it can collect image data in real time with excellent performance. Specifically, this image sensor has an ultra-high-definition resolution of no less than 4K, which can accurately capture every detail in the picture, and even subtle textures and features can be clearly presented. In terms of frame rate, the sensor can reach 60fps (60 frames per second), which means that in dynamic scenes, it can record the picture smoothly and without stuttering, avoiding the picture lag and smear caused by insufficient frame rate, and is suitable for capturing fast-moving objects. The performance under different lighting conditions is a highlight of this image acquisition module. Whether it is an outdoor environment with direct strong light, a dark indoor scene with weak light, or a complex backlight condition, it can stably capture clear and smear-free images. This is due to the advanced photosensitive element used inside it - the back-illuminated CMOS sensor. The photosensitive element increases the photosensitive area by optimizing the pixel structure, so that more photons can be captured under low light conditions, thereby improving the brightness and clarity of the image. In addition, low-noise amplifiers and high-precision analog-to-digital converters (ADCs) are used in circuit design. The low-noise amplifier can effectively suppress noise interference in the circuit and ensure the purity of the signal; the high-precision analog-to-digital converter can accurately convert the analog signal into a digital signal, further improving the signal-to-noise ratio of the image and making the captured image more realistic and delicate.

[0032] Image preprocessing module: optimizes the collected original image to lay the foundation for subsequent in-depth analysis and processing. In terms of denoising, this module uses bilateral filtering algorithm. This algorithm takes two key factors into consideration when processing images. The first is spatial distance, that is, the physical position relationship between adjacent pixels. Pixels with closer distances have greater mutual influence during the filtering process; the second is pixel value difference, that is, the degree of grayscale or color difference between different pixels. When calculating the filtering weight, bilateral filtering will make adaptive adjustments based on these two factors. For areas in the image where the noise distribution is relatively random and the grayscale changes are small, the algorithm will enhance the smoothing effect to effectively remove the noise; while in areas where the pixel value changes drastically, such as the edge of the image, bilateral filtering will reduce the degree of smoothing, thereby retaining the edge information of the image to the greatest extent, avoiding edge blurring during the denoising process, and ensuring that the contour and details of the image can be clearly presented. The image enhancement link uses the histogram equalization algorithm. The core of this algorithm is to conduct in-depth analysis and adjustment of the grayscale histogram of the image. The grayscale histogram shows the distribution of the number of pixels corresponding to each grayscale level in the image. By redistributing the grayscale values ​​of the image, the pixel value range originally concentrated on a few grayscale levels is expanded to a wider range, making the grayscale distribution of the image more uniform. In this way, the originally darker or brighter areas can obtain richer grayscale levels, thereby significantly enhancing the contrast of the image, making the details in the image clearer and more discernible, and presenting better visual effects, whether it is the texture of the dark part or the highlight of the bright part. The normalization formula is , I is the original image pixel value, I min and I max are the minimum and maximum values ​​of the image pixels, respectively. By calculating this formula, the dynamic range of the pixel values ​​in the original image can be mapped to the standard interval of [0,1]. This normalization process is of great significance. On the one hand, it can eliminate the pixel value differences between different images caused by factors such as shooting equipment and lighting conditions, so that the image data has a unified scale; on the other hand, the normalized data format is more in line with the input requirements of subsequent deep learning models or other image processing algorithms, which helps to improve the accuracy and efficiency of processing.

[0033] Feature extraction module: Extract representative features from the original image to provide effective information for subsequent image classification, target recognition and other tasks. This module uses the local binary pattern (LBP) and directional gradient histogram (HOG) algorithms to extract the texture and shape features of the image respectively, and finally generate the feature vector F. The local binary pattern (LBP) algorithm is a classic texture feature extraction method. In its processing, a neighborhood of a specific size is delineated with each pixel in the image as the center. Then, the gray value of the central pixel is used as the threshold and compared with the gray values ​​of other pixels in the neighborhood. If the gray value of the neighborhood pixel is greater than or equal to the gray value of the central pixel, it is encoded as 1; otherwise, it is encoded as 0. Connecting these binary codes in a certain order (such as clockwise) forms a binary number. This binary number can reflect the local texture pattern around the pixel, and different binary codes correspond to different texture features. By performing such operations on all pixels in the image, an LBP feature image can be obtained, which contains rich texture information, thereby achieving an effective description of the local texture of the image. The Histogram of Directed Gradients (HOG) algorithm is mainly used to extract the shape and edge features of an image. First, the image is grayed to remove color information and reduce the amount of calculation. Next, the gradient magnitude and direction of each pixel in the image are calculated. The gradient magnitude reflects the degree of intensity change of the pixel at that position, while the gradient direction indicates the direction of the intensity change. Then, the image is divided into several local areas (called cells), and the histogram of the gradient direction is statistically calculated in each cell. The horizontal axis of the histogram represents the different intervals of the gradient direction, and the vertical axis represents the cumulative value of the gradient magnitude in each interval. These histograms reflect the distribution of edges and shapes in the region. Finally, several adjacent cells are combined into a block, and the histogram in each block is normalized to enhance the algorithm's robustness to changes in illumination and local contrast. The histograms of all blocks are connected in sequence to obtain the HOG feature description of the entire image. In order to obtain more comprehensive and representative image features, the features extracted by the LBP algorithm and the HOG algorithm are fused. Since LBP mainly describes the texture information of the image, while HOG focuses on extracting the shape and edge features of the image, the two complement each other. Through reasonable fusion strategies, such as simple concatenation or weighted fusion, the two features are combined to generate the final feature vector F. This feature vector combines the texture and shape information of the image, can better represent the essential characteristics of the original image, and provide more powerful support for subsequent analysis and processing.

[0034] Target recognition module: With the help of deep learning models, especially convolutional neural networks (CNNs), the feature vector F extracted from the image is deeply analyzed to determine the target object in the image and its category C. As a powerful deep learning architecture, the unique hierarchical design of convolutional neural networks (CNNs) enables them to automatically learn rich feature representations from image data. The model contains multiple hierarchical structures with different functions. The convolution layer is one of the core components of CNN. In the convolution layer, the convolution kernel (also called filter) slides on the image and extracts the deep features of the image by convolution operations with the local area of ​​the image. Each convolution kernel has specific weight parameters, which are continuously adjusted and optimized during the training process to capture different types of features in the image, such as edges, corners, textures, etc. Different convolution kernels can learn different feature patterns. Multiple convolution kernels work in parallel to extract diverse features, thereby constructing a feature map of the image. As the number of network layers increases, the features extracted by the convolution layer gradually transition from simple low-level features to complex high-level semantic features. The pooling layer usually follows the convolution layer, and its main function is to reduce the dimensionality of the feature map. Common pooling operations include maximum pooling and average pooling. Taking maximum pooling as an example, it divides the feature map into several non-overlapping sub-regions, and selects the maximum value in each sub-region as the output of the region, thereby reducing the size of the feature map and the number of parameters. This dimensionality reduction operation can not only reduce the amount of calculation, improve the training and reasoning speed of the model, but also enhance the robustness of the model to image translation, scaling and rotation to a certain extent. The fully connected layer is responsible for integrating the features extracted by the previous convolutional layer and pooling layer and mapping them to different target categories. In the fully connected layer, each neuron is connected to all neurons in the previous layer, and the feature vector is converted into a category probability distribution through weighted summation and activation function operations. For example, in a multi-classification problem, the number of output nodes of the fully connected layer is equal to the number of target categories, and the value of each node represents the probability that the input image belongs to that category. In order to improve the accuracy of target recognition, the CNN model needs to be trained on a large number of labeled image datasets. During the training process, the error between the predicted result and the true label is calculated through the back propagation algorithm, and the error is back propagated to each layer of the network to adjust and optimize the parameters of the model (such as the weight of the convolution kernel, the weight and bias of the fully connected layer, etc.), so that the predicted result of the model is closer and closer to the true value. After multiple rounds of training, the CNN model can learn the complex mapping relationship between image features and target categories, thereby realizing accurate recognition and classification of target objects in the input image.

[0035] Decision-making module: Based on the output of the target recognition module, the pre-set rules and strategies are integrated to generate accurate and effective decision instructions D to achieve intelligent response and regulation of specific scenarios. When constructing preset rules and strategies, it is necessary to comprehensively consider various key factors of the target object. The category of the target object is one of the important bases, and different categories of targets often correspond to different processing methods. For example, in a security monitoring scenario, if the target object is identified as a human, the system may further analyze whether its behavior is abnormal; if it is identified as a vehicle, it will focus on its driving trajectory and speed. The number of target objects should not be ignored. By counting the number of target objects, the system can determine the status of the current scene. For example, in the monitoring of crowded places, when the number of people detected exceeds the pre-set safety threshold, the decision-making module can trigger the crowd diversion plan or send early warning information to the management personnel so that timely measures can be taken to avoid safety accidents such as crowding and trampling. The location information of the target object is also critical. The system can accurately obtain the coordinate position of the target object in the image or actual scene, and then determine whether it is in a specific area. Taking warehouse monitoring as an example, if the goods are moved to a non-designated storage area, the decision module can generate an alarm instruction based on this position change to remind the staff that the goods may be misplaced or stolen. In addition, the movement state of the target object, including parameters such as speed and direction, is also an important reference for decision making. In the traffic monitoring system, when the vehicle speed exceeds the speed limit or the driving direction is abnormal (such as driving in the opposite direction), the decision module will quickly generate corresponding penalties or warning instructions to ensure traffic order and safety. Taking the actual application scenario as an example, when the target recognition module detects a specific intrusion target object entering the monitoring area, the decision module will first confirm the category and location information of the target. If it is judged as an illegal intrusion behavior, an alarm instruction will be immediately generated according to the preset rules. The instruction will not only trigger the local sound and light alarm device, but also send the alarm information to the terminal device of the security personnel in real time so that timely countermeasures can be taken. When the number of target objects exceeds a certain threshold, the decision module will automatically adjust the focal length, angle and other parameters of the camera according to the camera parameters and the current shooting range through calculation and analysis, expand the shooting range, and ensure that all target objects in the scene can be fully monitored to avoid monitoring blind spots.

[0036] Feedback control module: Based on the decision instruction D output by the decision module, the various parameters of the camera are precisely adjusted to ensure that the captured image can continue to meet the actual needs of monitoring. In actual operation, the parameter adjustment of the camera involves many important aspects. Focal length is a key parameter that determines the distance and size of the camera's shooting picture. When the decision instruction requires capturing a distant target, the feedback control module will calculate the focal length adjustment amount ΔP based on a specific algorithm. For example, the difference between the actual size of the target in the image and the preset size is used, combined with the optical characteristic formula of the camera, to calculate the focal length value that needs to be increased. Then follow the adjustment formula P new =P old +ΔP, the original focal length value P old Update to the new focal length value P new , so that the camera can clearly capture distant objects. The aperture mainly affects the amount of light entering the camera and the depth of field effect. Under different lighting conditions and monitoring requirements, the aperture size needs to be adjusted accordingly. If the decision instruction is to obtain a clear image in a low-light environment, the feedback control module will calculate the adjustment amount ΔP to increase the aperture. This calculation process will comprehensively consider factors such as the light intensity of the current environment and the sensitivity of the camera sensor. By increasing the aperture, more light enters the camera, thereby improving the brightness and clarity of the image. At the same time, a larger aperture can also produce a shallow depth of field effect, making the target subject more prominent. Adjustment of the shooting angle is crucial for comprehensive coverage of the monitoring area. The feedback control module will calculate the adjustment amount ΔP of the shooting angle based on the position information of the target object in the decision instruction. For example, when the target enters the edge of the monitoring area, the module will calculate the angle value that needs to be rotated for the camera, and accurately adjust the horizontal and vertical shooting angles of the camera by controlling actuators such as motors to ensure that the target is always within the picture range. When calculating the parameter adjustment amount ΔP, the feedback control module will use a series of complex mathematical models and algorithms. These algorithms will fully consider the hardware characteristics of the camera, the current shooting parameter status, and the specific needs conveyed by the decision-making instructions. At the same time, in order to ensure the accuracy and stability of the adjustment, the module will also introduce a feedback mechanism to monitor the effect of the adjusted parameters in real time and compare it with the expected effect. If it is found that the actual effect deviates from the expected effect, the parameter adjustment amount will be calculated and fine-tuned again until the captured image fully meets the monitoring needs.

[0037] The present invention also includes the following modules: Light compensation module: Real-time detection of ambient light intensity L. The light compensation module uses a light sensor to measure the ambient light intensity. When the light intensity does not meet the preset range [L min ,L max ], through the formula Perform illumination compensation on the image, I is the original image pixel value, L ref is the reference light intensity.

[0038] The present invention also includes the following modules: Anomaly detection module: detects abnormal areas in the image by calculating the entropy value H of the image. The anomaly detection module divides the image into blocks and calculates the entropy value of each image block. The entropy value calculation formula is: , p(i) is the probability of the pixel with gray value i appearing in the image; when the entropy value H exceeds the preset threshold H threshold When the image is abnormal, the system marks the abnormal area, zooms in for observation, and triggers an alarm mechanism.

[0039] In the present invention, the convolutional neural network (CNN) in the target recognition module uses a pre-trained model and is fine-tuned through transfer learning. The pre-trained model is trained on a large-scale image dataset and learns a wealth of image features and patterns. During the transfer learning process, some parameters of the pre-trained model are used as initial values ​​and fine-tuned on a specific target dataset to adapt to the specific target recognition task. The cross entropy loss function is used during the training process. , is the true label, To predict the label, the model parameters are optimized by minimizing the loss function.

[0040] In the present invention, the decision module is a key link, and its internal preset rules and strategies adopt fuzzy logic algorithms, which brings unique advantages to system decision-making. The processing flow of fuzzy logic algorithms is relatively sophisticated. First, the fuzzification of factors related to the target object is performed. Taking the category C of the target object as an example, in actual scenarios, the object category may not be an absolutely clear boundary. For example, in security monitoring, for an animal whose appearance is between a small dog and a large cat, the simple classification of "dog" or "cat" is not accurate enough. The fuzzy logic algorithm will map it to a fuzzy set described by membership such as "small dog degree" and "large cat degree" based on its appearance characteristics, behavior patterns and other multi-dimensional information. For the size S of the target object, it is no longer a simple specific size value, but a fuzzy interval is divided according to actual monitoring needs. For example, in warehouse cargo monitoring, the size of the cargo is fuzzified into fuzzy sets such as "extremely small", "smaller", "medium", "larger" and "extremely large". According to factors such as the actual proportion of the cargo occupying the screen and the comparison with surrounding reference objects, its membership in different fuzzy sets is determined. The same is true for the position P of the target object. In a large monitoring area, the position is not an exact coordinate point, but is divided according to the regional function. For example, in shopping mall monitoring, the position is fuzzified into fuzzy sets such as "near the entrance", "cashier area", "inside the store", and "channel". By calculating the degree of association between the target object coordinates and each area, its membership is determined. After the fuzzification process is completed, the fuzzy logic algorithm will generate a decision instruction D based on the preset fuzzy reasoning rules. These reasoning rules are based on a large number of practical experiences and application scenarios. For example, when a target object is detected to have a high "intrusion target membership" in the fuzzy set, and the location is also high in the fuzzy set membership of "near important areas", according to the preset rules, the system will generate a higher priority alarm decision instruction. At the same time, if the size of the target object is in the "large" fuzzy set with a high membership, it may also trigger additional instructions such as additional security forces. The power of the fuzzy logic algorithm lies in its ability to effectively handle the uncertainty and fuzzy information that are prevalent in the real world. In the application scenario of the smart camera system, factors such as lighting changes and partial occlusion of the target object will lead to inaccuracy in the information obtained. The fuzzy logic algorithm can make relatively reasonable and flexible decisions based on this imprecision through the fuzzification and reasoning process, so that the smart camera system can make smarter decisions when facing complex and changing environments and can better meet actual monitoring needs.

[0041] In the present invention, the feedback control module not only undertakes the task of adjusting the camera parameters according to the decision instructions, but also has a strong adaptive adjustment capability to ensure that high-quality images are taken. The feedback control module continuously monitors the adjusted image effect, and its monitoring basis is a series of scientific image quality evaluation indicators. Among them, clarity is a key indicator to measure the degree of detail presentation of the image. In order to accurately evaluate the clarity, the module adopts the Laplace variance method. The principle of this method is to filter the image using the Laplace operator, which can enhance the edge and detail information in the image. After filtering, the variance of the image is calculated. The larger the variance value, the richer the details in the image and the higher the clarity. For example, when monitoring road scenes, if the details such as the license plate of the vehicle and the texture of the road surface show a higher variance value in the Laplace variance calculation, it means that the image clarity is good; on the contrary, if the variance value is low, it means that the image is relatively blurred and the camera parameters may need to be further adjusted. Contrast is another important evaluation indicator, which reflects the degree of light and dark differences between different areas in the image. The feedback control module uses a special contrast evaluation function to measure. This function usually analyzes the distribution range and differences of the grayscale values ​​of pixels in the image. For example, in a landscape image, if the grayscale difference between the bright part of the sky and the dark part of the ground is obvious, and the grayscale value distribution range is wide, then the result calculated by the contrast evaluation function will show that the image has a high contrast and a strong sense of layering; if the image is dark or bright as a whole, and the grayscale values ​​are concentrated in a smaller range, the contrast evaluation function will give a lower score, indicating that the camera parameters need to be adjusted to improve the contrast. In actual operation, the feedback control module will call these evaluation functions in real time to evaluate the image quality. Once it is detected that the image quality does not meet the preset requirements, such as the clarity is below a certain threshold or the contrast is insufficient, the module will automatically start the parameter adjustment mechanism. It will recalculate the parameter adjustment amount ΔP based on the current evaluation results. For example, if it is found that the image blur is caused by inaccurate focal length, the module will calculate the focal length value that needs to be adjusted based on the feedback information of the clarity evaluation through a specific algorithm, and then press P new =P old +ΔP formula to adjust the focal length parameters of the camera. During the adjustment process, the module will continuously monitor the changes in image quality and fine-tune the parameters until the image clarity, contrast and other indicators reach the best state, thereby achieving the best shooting effect and providing high-quality image data for subsequent image processing and analysis.

[0042] The present invention also includes the following modules: Data storage and analysis module: stores various types of information such as collected image data, feature vectors, target recognition results, and decision instructions, and mines data value through data analysis algorithms. In terms of data storage, in order to meet the large amount of data storage needs continuously generated by the intelligent camera system, the data storage and analysis module uses large-capacity hard disks or cloud storage devices. For local hard disk storage, enterprise-level hard disks will be selected, which have the characteristics of high reliability and high read and write speed. For example, hard disks with SAS interfaces have faster data transmission rates than ordinary SATA hard disks, and can quickly store high-definition image data collected by the camera in real time. At the same time, in order to prevent data loss, disk array technology (such as RAID5 or RAID6) will be used. Through data redundancy storage, even if some hard disks fail, the integrity and recoverability of the data can be guaranteed. If cloud storage devices are selected, they will cooperate with well-known cloud service providers, such as Amazon AWS and Alibaba Cloud. These cloud platforms have powerful distributed storage architectures and perfect data backup mechanisms. Data will be stored in servers in multiple geographical locations, which not only ensures the safe storage of data, but also allows convenient access and call of data on different devices and locations. In addition, cloud storage also has elastic expansion capabilities, which can automatically increase storage capacity according to the growth of data volume, without worrying about the problem of insufficient capacity of local storage devices. In the data analysis stage, this module uses algorithms such as cluster analysis and association rule mining. Cluster analysis algorithms can classify different types of target objects. Taking the security monitoring scene as an example, the system will collect a large amount of image data containing different target objects such as pedestrians, vehicles, and animals. The cluster analysis algorithm will automatically divide the target objects into different categories based on their feature vectors, such as shape, texture, color and other information. Specifically, when using the K-Means clustering algorithm, K cluster centers are first randomly determined, and then the distance between each target object feature vector and these cluster centers is calculated, and the target object is assigned to the cluster with the closest distance. By continuously iterating and updating the cluster center, a more accurate classification result is finally obtained. The association rule mining algorithm is used to discover the association between the appearance of the target object and environmental factors. For example, in shopping mall monitoring, the relationship between the passenger flow (target object) in different time periods and the promotion activities (environmental factors) in the mall is analyzed. Using the Apriori algorithm, we first find all frequent item sets, that is, combinations of target objects and environmental factors that appear frequently, and then generate association rules based on these frequent item sets. For example, if we find that when a mall is holding a discount promotion, the customer flow in a certain area will increase significantly, this association information can provide a basis for adjusting the mall's operating strategy and setting the monitoring focus of smart cameras, and also help to further optimize the functions and performance of the entire smart camera system.

[0043] The present invention also includes the following steps: Image acquisition steps: Use image sensors to collect image data in real time. The image sensors have a resolution of 4K and a frame rate of 60fps. Image preprocessing steps: denoising, enhancing and normalizing the collected images. The normalization formula is: , denoising uses bilateral filtering algorithm, and enhancement uses histogram equalization algorithm; Feature extraction step: Use the local binary pattern (LBP) and histogram of oriented gradients (HOG) algorithms to extract the texture and shape features of the image and generate the feature vector F; Target recognition step: Based on a deep learning model, specifically a convolutional neural network (CNN), the feature vector F is classified and recognized to determine the target object and its category C in the image. The CNN model is trained using pre-training and transfer learning methods; Decision-making step: Generate decision instruction D based on target recognition results and combined with preset rules and strategies. The preset rules and strategies use fuzzy logic algorithm; Feedback control step: According to the decision instruction D, the parameters of the camera are adjusted, specifically the focal length, aperture, and shooting angle. The adjustment formula is P new =P old +ΔP.

[0044] In the present invention, before the image preprocessing step, it also includes real-time detection of the ambient light intensity L. A high-precision light sensor is used for detection. When the light intensity does not meet the preset range [L min ,L max ], through the formula Perform illumination compensation on images to improve image quality under different lighting conditions.

[0045] In the present invention, after the target recognition step, the entropy value H of the image is calculated to detect abnormal areas in the image. The image is divided into blocks, and the entropy value of each image block is calculated respectively. The entropy value calculation formula is When the entropy value H exceeds the preset threshold H threshold When an abnormality is detected in the image, the image is judged to be abnormal and the corresponding processing mechanism is triggered, such as marking, zooming in or alarming.

[0046] The above are only preferred specific implementation modes of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and inventive concepts of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An intelligent camera system based on image processing algorithm, characterized in that: Includes the following modules: Image acquisition module: uses image sensors to collect image data in real time and collect images under different lighting conditions; Image preprocessing module: denoise, enhance and normalize the collected images. The denoising method uses the bilateral filtering algorithm, and the enhancement method uses the histogram equalization algorithm. The normalization formula is: , I is the original image pixel value, I min and I max are the minimum and maximum values ​​of image pixels respectively; Feature extraction module: Use the local binary pattern LBP and directional gradient histogram HOG algorithm to extract the texture and shape features of the image and generate the feature vector F; Target recognition module: classifies and recognizes the feature vector F to determine the target object in the image and its category C; the model includes convolutional layer, pooling layer and fully connected layer. The convolutional layer extracts the deep features of the image through the convolution kernel, the pooling layer performs dimensionality reduction processing on the feature map, and the fully connected layer maps the extracted features to the target category; Decision module: Generates decision instruction D based on target recognition results combined with preset rules and strategies. The preset rules and strategies take into account the category, quantity, location, and motion state of the target objects. When the number of target objects exceeds the threshold, adjusts the camera's shooting range. Feedback control module: adjust the camera parameters according to the decision instruction D, and the adjustment formula is P new =P old +ΔP,P old is the original parameter value, and ΔP is the parameter adjustment amount calculated according to the decision instruction.

2. The intelligent camera system based on image processing algorithm according to claim 1, characterized in that: Also includes: Light compensation module: Real-time detection of ambient light intensity L. The light compensation module uses a light sensor to measure the ambient light intensity. When the light intensity does not meet the preset range [L min ,L max ], through the formula Perform illumination compensation on the image, I is the original image pixel value, L ref is the reference light intensity.

3. The intelligent camera system based on image processing algorithm according to claim 1, characterized in that: Also includes: Anomaly detection module: detects abnormal areas in the image by calculating the entropy value H of the image. The anomaly detection module divides the image into blocks and calculates the entropy value of each image block. The entropy value calculation formula is: , p(i) is the probability of the pixel with gray value i appearing in the image; when the entropy value H exceeds the preset threshold H threshold When the image is abnormal, the system marks the abnormal area, zooms in for observation, and triggers an alarm mechanism.

4. The intelligent camera system based on image processing algorithm according to claim 1, characterized in that: The convolutional neural network (CNN) in the object recognition module uses a pre-trained model and is fine-tuned through transfer learning; the pre-trained model is trained on image datasets to learn image features and patterns; In the process of transfer learning, the parameters of the pre-trained model are used as initial values ​​and fine-tuned on the target dataset; The cross entropy loss function is used during training. , is the true label, is the predicted label.

5. The intelligent camera system based on image processing algorithm according to claim 1, characterized in that: The preset rules and strategies in the decision-making module adopt fuzzy logic algorithm, which fuzzifies the category C, size S and position P factors of the target object and maps them to different fuzzy sets; through the preset fuzzy inference rules, the corresponding decision instructions D are generated according to the input fuzzy variables.

6. The intelligent camera system based on image processing algorithm according to claim 1, characterized in that: The feedback control module has the ability of adaptive adjustment. By monitoring the adjusted image effect, it dynamically adjusts the parameter adjustment amount ΔP according to the image quality evaluation index. It uses the image clarity evaluation function, specifically the Laplace variance method and the contrast evaluation function, to evaluate the image quality in real time. When the image quality does not meet the requirements, it automatically adjusts the camera parameters.

7. The intelligent camera system based on image processing algorithm according to claim 1, characterized in that: Also includes: Data storage and analysis module: stores collected image data, feature vectors, target recognition results and decision instruction information. The data storage and analysis module uses hard disk or cloud storage devices to discover potential patterns and rules in the data through data analysis algorithms, specifically cluster analysis and association rule mining.

8. An intelligent camera optimization method based on an image processing algorithm based on the system according to any one of claims 1 to 7, characterized in that: The following steps are involved: Image acquisition steps: Use image sensors to collect image data in real time. The image sensors have a resolution of 4K and a frame rate of 60fps. Image preprocessing steps: denoising, enhancing and normalizing the collected images. The normalization formula is: , denoising uses bilateral filtering algorithm, and enhancement uses histogram equalization algorithm; Feature extraction step: Use the local binary pattern LBP and directional gradient histogram HOG algorithm to extract the texture and shape features of the image and generate the feature vector F; Target recognition step: Based on a deep learning model, specifically a convolutional neural network (CNN), the feature vector F is classified and recognized to determine the target object in the image and its category C. The CNN model is trained using pre-training and transfer learning methods; Decision-making step: Generate decision instruction D based on target recognition results and combined with preset rules and strategies. The preset rules and strategies use fuzzy logic algorithm; Feedback control step: According to the decision instruction D, the parameters of the camera are adjusted, specifically the focal length, aperture, and shooting angle. The adjustment formula is P new =P old +ΔP.

9. The intelligent camera optimization method based on image processing algorithm according to claim 8, characterized in that: Before the image preprocessing step, it also includes real-time detection of the ambient light intensity L; using a light sensor for detection, when the light intensity does not meet the preset range [L min ,L max ], through the formula Perform lighting compensation on the image.

10. The intelligent camera optimization method based on image processing algorithm according to claim 8, characterized in that: After the target recognition step, the entropy value H of the image is calculated to detect abnormal areas in the image; the image is divided into blocks and the entropy value of each image block is calculated respectively. The entropy value calculation formula is: , when the entropy value H exceeds the preset threshold H threshold When an abnormality is detected in the image, the corresponding processing mechanism is triggered, including marking, magnification observation and alarm.

Images