Food quality control method and system based on intelligent image analysis

Abstract

The invention provides a food quality control method and system based on intelligent image analysis. The method comprises the following steps: collecting multi-angle multispectral real-time image streams at different positions of a food production line through a multi-sensor system to form a comprehensive image data set; based on the comprehensive image data set, using a convolution automatic encoder to carry out image denoising and contrast enhancement processing, combining a super-resolution reconstruction technology to improve image definition and detail performance, and generating a high-quality image data set; based on the high-quality image data set, performing anomaly detection processing by using a deep residual network, identifying an abnormal region, and performing dynamic analysis in combination with a time series data analysis technology to generate an anomaly detection result; and based on the anomaly detection result, sending the anomaly detection result to a food production line control system, adjusting production parameters, and generating a food quality control strategy. According to the technical scheme, the product quality is improved, the operation cost is reduced, and the market competitiveness is enhanced.

Description

Technical Field

[0001] The embodiments of the present application relate to the technical field of food quality control, and in particular, to a food quality control method and system based on intelligent image analysis. Background Art

[0002] With the rapid development of the food industry and the increasing requirements of consumers for food safety and quality, it is necessary to monitor the appearance, color, shape, etc. of food in real time to detect whether there are foreign objects, deterioration, damage, etc. The traditional manual inspection method is not only inefficient, but also easily affected by human factors, resulting in inconsistent detection results. In order to achieve efficient and accurate quality control, a method is needed that can automatically collect multi-angle and multi-spectral image streams and process them in real time through intelligent image analysis technology. This method should have functions such as image denoising, contrast enhancement, and super-resolution reconstruction to improve image quality; at the same time, it is also necessary to combine deep learning models for anomaly detection, and be able to feedback detection results in real time to support the adjustment of production line parameters.

[0003] At present, some food production companies use quality control systems based on traditional image processing technology and simple machine learning algorithms. They use a single camera or sensor to collect images at a fixed position, and then perform preliminary analysis through image processing algorithms; use traditional image processing techniques (such as filtering, edge detection, etc.) to denoise and enhance images; and apply some basic machine learning models (such as support vector machines, random forests, etc.) for anomaly detection.

[0004] Image data collected by a single sensor cannot fully reflect all situations on the production line, especially in complex environments with multiple angles and spectra, where the integrity and accuracy of the data are difficult to guarantee; traditional image processing technology has limited effects in denoising and enhancement, and low-quality images will directly affect the accuracy of anomaly detection, thereby reducing the effect of product quality control and making product quality stability low; existing basic machine learning models are weak in feature extraction and pattern recognition, especially when processing complex and changeable image data, which are prone to false detection and missed detection; existing solutions often only focus on the analysis of static images, ignoring the factors of dynamic changes in the production process, and are unable to provide comprehensive anomaly detection results; due to the speed limitations of data processing and analysis, it is difficult for existing solutions to achieve true real-time feedback, resulting in delayed adjustment of production parameters, affecting production efficiency and product quality. Summary of the invention

[0005] The embodiments of the present application provide a food quality control method and system based on intelligent image analysis to solve the problem of poor food quality control in the prior art.

[0006] In a first aspect, an embodiment of the present application provides a food quality control method based on intelligent image analysis, comprising:

[0007] The multi-sensor system collects multi-angle and multi-spectral real-time image streams at different locations of the food production line to form a comprehensive image data set;

[0008] Based on the comprehensive image dataset, a convolutional autoencoder is used to perform image denoising and contrast enhancement processing, and super-resolution reconstruction technology is combined to improve image clarity and detail expression, thereby generating a high-quality image dataset;

[0009] Based on the high-quality image data set, a deep residual network is used to perform anomaly detection processing, abnormal areas of the high-quality image data set are identified, and dynamic analysis is performed in combination with time series data analysis technology to generate anomaly detection results;

[0010] Based on the abnormal detection results, they are sent to the food production line control system in real time through an encrypted interface to support timely adjustment of production parameters and generate food quality control strategies.

[0011] Optionally, based on the comprehensive image dataset, using a convolutional autoencoder to perform image denoising and contrast enhancement processing, combined with super-resolution reconstruction technology to improve image clarity and detail expression, and generate a high-quality image dataset, including:

[0012] Based on the comprehensive image data set, performing a convolutional autoencoder random noise removal process to improve image purity and generate a denoised image data set;

[0013] Based on the denoised image dataset, the nonlinear mapping capability of the convolutional autoencoder is utilized to analyze and adjust the image histogram, perform contrast enhancement processing, and generate a contrast enhanced image dataset;

[0014] Based on the contrast enhanced image data set, combined with super-resolution reconstruction technology, high-resolution image detail prediction is performed to improve image resolution and clarity and generate high-resolution images;

[0015] Based on the high-resolution image, the local contrast is adjusted, edge sharpening is performed, and the microscopic level is enhanced to improve the image detail expression and generate a high-quality image data set.

[0016] Optionally, based on the denoised image dataset, using the nonlinear mapping capability of a convolutional autoencoder, analyzing and adjusting the image histogram, performing contrast enhancement processing, and generating a contrast enhanced image dataset includes:

[0017] Based on the denoised image dataset, a multi-scale key feature extraction process is performed by pre-training a convolutional autoencoder to generate image dataset features;

[0018] Based on the features of the image dataset, the nonlinear mapping capability of the convolutional autoencoder is used to transform the features of the image dataset to adapt to the image histogram adjustment requirements, thereby generating a feature-transformed image dataset;

[0019] Based on the feature transformation image data set, analyzing the specific distribution of the image histogram, identifying the contrast enhancement optimization space, and generating an image histogram analysis result;

[0020] Based on the image histogram analysis result, the brightness of the low-light area is enhanced and the contrast of the high-contrast area is adjusted to improve the overall visual effect and generate a contrast-enhanced image data set.

[0021] Optionally, based on the image dataset features, using the nonlinear mapping capability of a convolutional autoencoder, the image dataset features are transformed to adapt to image histogram adjustment requirements, and a feature-transformed image dataset is generated, including:

[0022] Based on the features of the image dataset, convolution operations are performed on the features of the image dataset using convolution kernels of different sizes to extract key features of different scales;

[0023] The key features of different scales are integrated, noise is suppressed by a smoothing filter, and high-frequency noise is removed to generate an intermediate result of feature change;

[0024] The intermediate result of feature transformation is calculated by the following formula:

[0025]

[0026] Where T(x) is the intermediate result of feature transformation; i is the index of the feature map, from 1 to N; N is the number of feature maps; W 1i is the convolution kernel corresponding to the i-th feature map; x i is the i-th feature map; b 1 is the bias term; tanh is the hyperbolic tangent activation function; α is the scaling factor; γ is the parameter that controls the exponential decay speed; j is the feature index involved in the exponential decay calculation, from 1 to M; M is the number of features involved in the exponential decay calculation; c is the center point value; x j is the jth feature map;

[0027] Based on the intermediate result of the feature transformation, further nonlinear transformation is performed to enhance the feature expression capability, and the most representative subset is obtained through a feature selection method to reduce redundant information to generate an enhanced feature representation;

[0028] The enhanced feature representation is calculated using the following formula:

[0029]

[0030] Where EFR(T(x)) is the enhanced feature representation; k is the processed feature map index, from 1 to K; K is the number of processed feature maps; W 2k is the convolution kernel corresponding to the kth feature map; T k (x) is the kth processed feature map; b 2 is the bias term; σ is the activation function; β is the scaling factor; ω is the parameter that controls the sine wave oscillation frequency; l is the feature index involved in the sine wave calculation, from 1 to L; L is the number of features involved in the sine wave calculation; T l (x) is the lth processed feature map; d is the center point value of the sine wave; η is the weight coefficient of the logarithmic term; m is the feature index for logarithmic calculation, from 1 to P; P is the number of features involved in logarithmic calculation; |T m (x)| is the absolute value of the mth feature map; δ is the scaling factor of the sigmoid term; θ is the parameter that controls the slope of the sigmoid function; n is the feature index involved in the sigmoid calculation, from 1 to Q; Q is the number of features involved in the sigmoid calculation; T n (x) is the nth processed feature map; φ is the scaling factor of the square root term; o is the feature index involved in the square root calculation, from 1 to R; R is the number of features involved in the square root calculation; e is the center point value of the square root calculation; T o (x) is the oth processed feature map;

[0031] Based on the enhanced feature representation, the image space is restored through inverse transformation, and histogram adjustment is performed to ensure image quality and visual effect, thereby generating a feature transformation image dataset.

[0032] Optionally, the step of predicting high-resolution image details based on the contrast-enhanced image dataset in combination with super-resolution reconstruction technology, improving image resolution and clarity, and generating a high-resolution image includes:

[0033] Based on the contrast enhanced image data set, image enlargement processing is performed by an image upsampling method to generate a preliminary image frame;

[0034] Based on the preliminary image frame and in combination with the detail information of the contrast enhanced image data set, an interpolation technique is used to fill in the missing pixels of the image to generate an optimized image frame;

[0035] Based on the optimized image frame, the optimized image frame is optimized by using the deep learning mechanism of super-resolution reconstruction technology, and potential subtle structures and texture features are predicted and supplemented to generate a reconstructed image frame;

[0036] Based on the reconstructed image framework, the overall clarity and resolution are adjusted to ensure that the image is not distorted during the image enlargement process and generate a high-resolution image.

[0037] Optionally, based on the high-quality image data set, using a deep residual network to perform anomaly detection processing, identifying abnormal areas of the high-quality image data set, combining time series data analysis technology to perform dynamic analysis, and generating anomaly detection results, including:

[0038] Based on the high-quality image dataset, a residual block is introduced to solve the gradient vanishing problem in deep learning, enhance the model learning ability, and generate a preprocessed image dataset;

[0039] Based on the preprocessed image data set, a deep residual network is used to perform anomaly detection processing, anomaly probability values ​​are calculated through forward propagation, and preliminary anomaly detection results are generated;

[0040] Based on the preliminary anomaly detection results, identify the time series pattern of abnormal events, perform dynamic analysis, and generate dynamic analysis results;

[0041] Based on the dynamic analysis results, a comprehensive evaluation is performed in combination with the multi-dimensional key features of the abnormal event to generate anomaly detection results.

[0042] Optionally, the method of performing anomaly detection processing based on the preprocessed image data set using a deep residual network, calculating anomaly probability values ​​through forward propagation, and generating preliminary anomaly detection results includes:

[0043] Based on the preprocessed image data set, a key feature extraction process is performed using a directional gradient histogram method to generate a feature representation data set;

[0044] Based on the feature representation data set, a deep residual network is used to calculate the feature representation through forward propagation to generate a deep feature representation data set;

[0045] Based on the deep feature representation data set, combined with the deep residual network classification layer, anomaly probability values ​​are obtained to generate anomaly probability value data set;

[0046] Based on the abnormal probability value data set, a reasonable threshold interval is preset for threshold judgment, and the abnormal probability value is marked and recorded to generate a preliminary abnormality detection result.

[0047] Optionally, based on the feature representation data set, a deep residual network is used to calculate feature representations through forward propagation to generate a deep feature representation data set, including:

[0048] Based on the feature representation data set, preprocess the input feature map, reduce input noise through filtering technology, highlight key information of the input feature map, and generate an intermediate output vector;

[0049] The intermediate output vector is calculated using the following formula:

[0050]

[0051] Among them, A i is the intermediate output vector; N is the number of input feature maps; W ij is the convolution kernel between the i-th output vector and the j-th input feature map; X j is the jth feature map in the input feature vector; b i is the bias term of the i-th output vector; σ is the ReLU activation function; α and β are parameters that control the strength of the nonlinear term; M is the number of input feature maps involved in the nonlinear transformation; V ik is the weight for nonlinear transformation between the i-th output vector and the k-th input feature map; j is the index of the input feature map, from 1 to N; k is the index of the input feature map, from 1 to M; X k Represents the kth input feature map participating in the nonlinear transformation;

[0052] Based on the intermediate output vector, multiple levels of nonlinear transformations are used to capture complex patterns, and additional convolutional layers and pooling layers are introduced to increase expression power and generalization performance to generate a final output vector;

[0053] The final output vector is calculated using the following formula:

[0054]

[0055] Among them, D i is the final output vector; N is the number of input feature maps; W' il is the convolution kernel between the i-th output vector and the l-th intermediate vector; A l is the lth intermediate vector; A m is the mth intermediate vector; A n is the nth intermediate vector; A o is the oth intermediate vector; b' i is the bias term of the i-th output vector; σ is the ReLU activation function; γ, δ and θ are parameters that control the strength of the nonlinear term; O is the number of intermediate vectors involved in the nonlinear transformation; U im is the weight used for nonlinear transformation between the ith output vector and the mth intermediate vector; ∈ is the parameter controlling the strength of the fractional term; P and Q are the number of intermediate vectors involved in the operation in the numerator and denominator respectively; Z in and Y ioare the weights between the i-th output vector and the n-th or o-th intermediate vector in the numerator and denominator respectively; l is the index of the input feature map, from 1 to N; m is the index of the intermediate vector, from 1 to O; n is the index of the intermediate vector, from 1 to P; o is the index of the intermediate vector, from 1 to Q;

[0056] Based on the final output vector, standardization processing is performed, dimension reduction is performed through principal component analysis, the high-dimensional feature vector is mapped into a low-dimensional space, and visualization analysis is performed to generate a deep feature representation data set.

[0057] Optionally, the abnormal detection result is sent to the food production line control system in real time through an encrypted interface to support timely adjustment of production parameters and generate a food quality control strategy, including:

[0058] Based on the abnormal detection result, the abnormal detection report is transmitted to the food production line control system in real time using an encrypted interface to generate an encrypted transmission result;

[0059] Based on the encrypted transmission results, combined with the historical data of the production line and the current production status, a comprehensive analysis is performed to generate production line adjustment suggestions;

[0060] Based on the production line adjustment suggestion, automatically adjusting production parameters through the food production line control system to generate adjusted production parameter settings;

[0061] Based on the adjustment of production parameter settings, actual operating performance indicators are recorded, combined with the abnormal detection results, continuous adjustment and optimization are carried out through a feedback mechanism to generate a food quality control strategy.

[0062] In a second aspect, the present application provides a food quality control system based on intelligent image analysis, including:

[0063] A collection module, used to collect multi-angle and multi-spectral real-time image streams at different positions of the food production line through a multi-sensor system to form a comprehensive image data set;

[0064] A processing module, for performing image denoising and contrast enhancement processing based on the comprehensive image dataset using a convolutional autoencoder, combining super-resolution reconstruction technology to improve image clarity and detail expression, and generating a high-quality image dataset;

[0065] An analysis module is used to perform anomaly detection processing based on the high-quality image data set using a deep residual network, identify abnormal areas of the high-quality image data set, perform dynamic analysis in combination with time series data analysis technology, and generate anomaly detection results;

[0066] The sending module is used to send the abnormal detection results to the food production line control system in real time through an encrypted interface to support timely adjustment of production parameters and generate food quality control strategies.

[0067] In an embodiment of the present application, a multi-angle and multi-spectral real-time image stream is collected from different positions of a food production line through a multi-sensor system to form a comprehensive image data set; based on the comprehensive image data set, a convolutional autoencoder is used to perform image denoising and contrast enhancement processing, and combined with super-resolution reconstruction technology, image clarity and detail expression are improved to generate a high-quality image data set; based on the high-quality image data set, a deep residual network is used to perform anomaly detection processing to identify abnormal areas of the high-quality image data set, and dynamic analysis is performed in combination with time series data analysis technology to generate anomaly detection results; based on the anomaly detection results, they are sent to the food production line control system in real time through an encrypted interface to support timely adjustment of production parameters and generate food quality control strategies. The collected multi-angle and multi-spectral image streams are de-noised and contrast enhanced by using a convolutional autoencoder. Combined with super-resolution reconstruction technology, the image clarity and detail can be significantly improved, providing more accurate basic data for subsequent analysis. The deep residual network model is used for anomaly detection. The model has a strong feature extraction capability and can effectively identify potential problem areas in the food production process. Combined with time series data analysis technology, the system can not only detect anomalies at the static level, but also capture dynamic change trends, further improving the comprehensiveness and accuracy of anomaly detection. The anomaly detection results are quickly fed back to the production line control system through an encrypted interface, supporting immediate adjustment of relevant production parameters, thereby achieving a rapid response to quality issues. The use of an encrypted interface to transmit anomaly detection information ensures the security of sensitive data, prevents unauthorized access or tampering, and maintains the information security of the enterprise. The entire solution embodies a highly integrated intelligent control concept. Through effective monitoring and management of each link in the production process, an integrated solution from data collection, processing to decision support is realized, which promotes the digital transformation of the food industry and improves the overall management level.

[0068] Furthermore, random noise removal processing is performed through a convolutional autoencoder, which can effectively reduce the noise in the image and improve the purity of the image, which is helpful for subsequent image processing steps and ensures the basis of image quality; the nonlinear mapping capability of the convolutional autoencoder is utilized to adjust the image histogram and perform contrast enhancement processing, so that the details in the image are more obvious, improving the visual effect and recognizability; combined with super-resolution reconstruction technology, high-resolution image detail prediction is performed on the denoised and contrast-enhanced images, which significantly improves the resolution and clarity of the images, and helps to more accurately capture and analyze minor defects or anomalies in the food production process; by performing local contrast adjustment and edge sharpening processing on the high-resolution images, the microscopic level of the image is further enhanced, making the subtle structure in the image clearer, which helps to improve the accuracy and reliability of anomaly detection; the high-quality image dataset finally generated not only has high clarity and rich details, but also has good contrast and purity, providing a high-quality data foundation for subsequent deep learning models, thereby improving the performance of the overall system.

[0069] Furthermore, by introducing residual blocks, the common gradient vanishing problem in deep learning is solved, the learning ability of the model is enhanced, and the deep residual network is able to better learn complex features and improve the accuracy of anomaly detection; the deep residual network is used for forward propagation to calculate the anomaly probability value and generate preliminary anomaly detection results, which can quickly identify potential abnormal areas and provide a basis for subsequent dynamic analysis; based on the preliminary anomaly detection results, the time series pattern of abnormal events is identified and dynamic analysis is performed, which can capture the development trend and change law of abnormal events and improve the comprehensiveness and accuracy of anomaly detection; a comprehensive evaluation is performed based on the multi-dimensional key features of abnormal events to generate the final anomaly detection results, which can comprehensively evaluate abnormal events from multiple angles and improve the reliability and robustness of anomaly detection; the generated anomaly detection results are sent to the food production line control system in real time through an encrypted interface, which supports timely adjustment of production parameters, helps to quickly respond to quality problems in the production process, improve product quality stability, reduce losses, and optimize production processes.

[0070] These and other aspects of the present application will become more clearly understood in the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0072] Figure 1A flowchart of a food quality control method based on intelligent image analysis provided in an embodiment of the present application;

[0073] Figure 2 A schematic diagram of the structure of a food quality control system based on intelligent image analysis provided in an embodiment of the present application;

[0074] Figure 3 A schematic diagram of the structure of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0075] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0076] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to be different types.

[0077] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0078] Figure 1 A flowchart of a food quality control method based on intelligent image analysis is provided for an embodiment of the present application, such as Figure 1 As shown, the method includes:

[0079] 101. Collect multi-angle and multi-spectral real-time image streams at different locations of the food production line through a multi-sensor system to form a comprehensive image data set;

[0080] The multi-sensor system includes multiple cameras and other sensors (such as infrared sensors, ultraviolet sensors, etc.), which are used to collect real-time image data of the food production line from different positions and angles. These data include visible light images, infrared images, ultraviolet images and other types of images, which are used to comprehensively monitor various situations in the production process.

[0081] A comprehensive image dataset refers to the integration of these multi-angle and multi-spectral image data to form a complete dataset for subsequent image processing and analysis.

[0082] In the embodiments of the present application, firstly, a multi-sensor system installs multiple cameras and other sensors at different positions of the food production line to ensure that image data can be collected from multiple angles and in multiple spectral ranges; secondly, these sensors collect images in real time and transmit the data to a central processing system; thirdly, the central processing system performs preliminary processing on the collected image data, including timestamp marking, location information recording, etc., to ensure the integrity and traceability of the data; finally, the processed image data is integrated into a comprehensive image data set to provide a high-quality data foundation for subsequent steps such as image denoising, contrast enhancement, and anomaly detection.

[0083] Suppose there is a food production line that needs to perform quality control on the packaging stage;

[0084] First, the above-mentioned cameras and other sensors are installed at key locations of the production line and calibrated to ensure that they can accurately capture the required image data; second, each sensor begins to collect image data in real time, such as 30 frames per second for visible light cameras, 10 frames per second for infrared cameras, and 5 frames per second for ultraviolet cameras; third, the collected image data is transmitted to the central processing system via a wired or wireless network, and the central processing system performs preliminary processing on the image data, including timestamp marking, location information recording, and basic format conversion; finally, the central processing system integrates all collected image data into a comprehensive image data set. This data set contains all images collected by each sensor at different time points, and each image is accompanied by a timestamp and location information for subsequent analysis and processing.

[0085] Through the above steps, a comprehensive image dataset containing multi-angle, multi-spectral real-time images is formed, which provides a solid foundation for subsequent image processing and quality control.

[0086] 102. Based on the comprehensive image dataset, use a convolutional autoencoder to perform image denoising and contrast enhancement processing, combine super-resolution reconstruction technology to improve image clarity and detail expression, and generate a high-quality image dataset;

[0087] Convolutional autoencoder is a deep learning model used for unsupervised learning to learn representations of images through encoding and decoding processes.

[0088] Image denoising refers to removing noise from an image and improving its purity.

[0089] Contrast enhancement refers to adjusting the brightness distribution of an image to make the details in the image more obvious.

[0090] Super-resolution reconstruction technology is an image processing technology that converts low-resolution images into high-resolution images through algorithms to improve image clarity and detail.

[0091] A high-quality image dataset refers to an image dataset that has high clarity, rich details, and good contrast after the above processing.

[0092] In this step, first, the convolutional autoencoder is used to remove random noise from the comprehensive image dataset to generate a denoised image dataset; second, the nonlinear mapping capability of the convolutional autoencoder is used to analyze and adjust the image histogram, perform contrast enhancement processing, and generate a contrast enhanced image dataset; third, combined with super-resolution reconstruction technology, high-resolution image details are predicted for the contrast-enhanced image to improve the resolution and clarity of the image and generate a high-resolution image; finally, the high-resolution image is subjected to local contrast adjustment and edge sharpening processing to further enhance the microscopic level of the image and generate the final high-quality image dataset.

[0093] Optionally, the step 102 uses a convolutional autoencoder to perform image denoising and contrast enhancement processing based on the comprehensive image dataset, and combines super-resolution reconstruction technology to improve image clarity and detail expression to generate a high-quality image dataset, including: based on the comprehensive image dataset, performing a convolutional autoencoder random noise removal processing to improve image purity and generate a denoised image dataset; based on the denoised image dataset, using the nonlinear mapping capability of the convolutional autoencoder, analyzing and adjusting the image histogram, performing contrast enhancement processing, and generating a contrast enhanced image dataset; based on the contrast enhanced image dataset, combining super-resolution reconstruction technology, performing high-resolution image detail prediction, improving image resolution and clarity, and generating a high-resolution image; based on the high-resolution image, adjusting the local contrast, performing edge sharpening processing, and enhancing the microscopic level to improve image detail expression and generate a high-quality image dataset.

[0094] Among them, based on the denoised image dataset, the nonlinear mapping capability of the convolutional autoencoder is utilized to analyze and adjust the image histogram, perform contrast enhancement processing, and generate a contrast enhanced image dataset, including: based on the denoised image dataset, multi-scale key feature extraction processing is performed by pre-training the convolutional autoencoder to generate image dataset features; based on the image dataset features, the nonlinear mapping capability of the convolutional autoencoder is utilized to transform the image dataset features to adapt to the image histogram adjustment requirements, and generate a feature transformed image dataset; based on the feature transformed image dataset, the specific distribution of the image histogram is analyzed, the contrast enhancement optimization space is identified, and the image histogram analysis result is generated; based on the image histogram analysis result, the brightness of the low-light area is enhanced, and the contrast of the high-contrast area is adjusted to improve the overall visual effect, and a contrast enhanced image dataset is generated.

[0095] Convolutional autoencoder is a deep learning model used for unsupervised learning to learn representations of images through encoding and decoding processes.

[0096] Image denoising refers to removing noise from an image to improve its purity. Contrast enhancement refers to adjusting the brightness distribution of an image to make the details in the image more obvious.

[0097] Super-resolution reconstruction technology is an image processing technology that converts low-resolution images into high-resolution images through algorithms to improve image clarity and detail.

[0098] A high-quality image dataset refers to an image dataset that has high clarity, rich details, and good contrast after the above processing.

[0099] In an embodiment of the present application, first, a convolutional autoencoder is used to remove random noise from a comprehensive image dataset to improve the purity of the image and generate a denoised image dataset. The nonlinear mapping capability of the convolutional autoencoder is used to perform multi-scale key feature extraction and generate image dataset features. Secondly, based on these features, the nonlinear mapping capability of the convolutional autoencoder is used to transform the image dataset features to adapt to the image histogram adjustment requirements, generate a feature-transformed image dataset, analyze the specific distribution of the image histogram, identify the contrast enhancement optimization space, and generate an image histogram analysis result. Thirdly, based on the image histogram analysis result, the brightness of low-light areas is enhanced, and the contrast of high-contrast areas is adjusted to improve the overall visual effect, generate a contrast-enhanced image dataset, and combine super-resolution reconstruction technology to predict high-resolution image details, improve image resolution and clarity, and generate a high-resolution image. Finally, based on the high-resolution image, the local contrast is adjusted, the edge sharpening is performed, and the microscopic level is enhanced to improve the image detail performance and generate a high-quality image dataset.

[0100] Assume that an organic food production line needs to perform quality control on the packaging process. The production line is equipped with multiple cameras and other sensors to form a comprehensive image dataset.

[0101] First, the collected image data is processed using a pre-trained convolutional autoencoder to remove random noise from the image and generate a relatively pure image dataset. The key features of the image are extracted through the convolutional autoencoder, and these features are transformed using its nonlinear mapping capability to adapt to the subsequent histogram adjustment needs; secondly, based on the transformed features, the system analyzes the specific distribution of the image histogram and identifies the space where the contrast can be optimized. According to the analysis results, the system enhances the brightness of low-light areas in the image and adjusts the contrast of high-contrast areas, thereby improving the overall visual effect and generating a contrast-enhanced image dataset; thirdly, based on the contrast-enhanced image dataset, the system uses super-resolution reconstruction technology to predict and generate higher-resolution images to further improve the clarity and detail of the image; finally, the system performs local contrast adjustment and edge sharpening on the high-resolution image to enhance the subtle structure in the image and ensure that every detail is clearly visible, ultimately generating a high-quality image dataset.

[0102] Through the above steps, the generated high-quality image dataset not only has high clarity and rich details, but also has good contrast and purity, providing a high-quality data foundation for subsequent deep learning models, thereby improving the performance of the overall system.

[0103] This application takes into account that in image processing and feature extraction, the convolutional autoencoder transforms the features of the image dataset through nonlinear mapping capabilities to adapt to the image histogram adjustment requirements, and finally generates a feature transformed image dataset, improves the feature expression capability through complex nonlinear transformations, and reduces redundant information to ensure image quality and visual effects.

[0104] Optionally, based on the image dataset features, using the nonlinear mapping capability of a convolutional autoencoder, the image dataset features are transformed to adapt to image histogram adjustment requirements, and a feature-transformed image dataset is generated, including:

[0105] Based on the features of the image dataset, convolution operations are performed on the features of the image dataset using convolution kernels of different sizes to extract key features of different scales;

[0106] The key features of different scales are integrated, noise is suppressed by a smoothing filter, and high-frequency noise is removed to generate an intermediate result of feature change;

[0107] The intermediate result of feature transformation is calculated by the following formula:

[0108]

[0109] Where T(x) is the intermediate result of feature transformation; i is the index of the feature map, from 1 to N; N is the number of feature maps; W 1i is the convolution kernel corresponding to the i-th feature map; x i is the i-th feature map; b 1 is the bias term; tanh is the hyperbolic tangent activation function; α is the scaling factor; γ is the parameter that controls the exponential decay speed; j is the feature index involved in the exponential decay calculation, from 1 to M; M is the number of features involved in the exponential decay calculation; c is the center point value; x j is the jth feature map;

[0110] Based on the intermediate result of the feature transformation, further nonlinear transformation is performed to enhance the feature expression capability, and the most representative subset is obtained through a feature selection method to reduce redundant information to generate an enhanced feature representation;

[0111] The enhanced feature representation is calculated using the following formula:

[0112]

[0113] Where EFR(T(x)) is the enhanced feature representation; k is the processed feature map index, from 1 to K; K is the number of processed feature maps; W 2k is the convolution kernel corresponding to the kth feature map; T k (x) is the kth processed feature map; b 2 is the bias term; σ is the activation function; β is the scaling factor; ω is the parameter that controls the sine wave oscillation frequency; l is the feature index involved in the sine wave calculation, from 1 to L; L is the number of features involved in the sine wave calculation; T l (x) is the lth processed feature map; d is the center point value of the sine wave; η is the weight coefficient of the logarithmic term; m is the feature index for logarithmic calculation, from 1 to P; P is the number of features involved in logarithmic calculation; |T m (x)| is the absolute value of the mth feature map; δ is the scaling factor of the sigmoid term; θ is the parameter that controls the slope of the sigmoid function; n is the feature index involved in the sigmoid calculation, from 1 to Q; Q is the number of features involved in the sigmoid calculation; T n (x) is the nth processed feature map; φ is the scaling factor of the square root term; o is the feature index involved in the square root calculation, from 1 to R; R is the number of features involved in the square root calculation; e is the center point value of the square root calculation; To ( x) is the oth processed feature map;

[0114] Based on the enhanced feature representation, the image space is restored through inverse transformation, and histogram adjustment is performed to ensure image quality and visual effect, thereby generating a feature transformation image dataset.

[0115] This method aims to extract more representative and discriminative features from the original image data through a series of mathematical transformations, so as to capture the key information in the image, reduce the impact of noise and redundant data, and thus improve the accuracy of subsequent processing tasks (such as classification, recognition, etc.).

[0116] In the intermediate result of feature transformation, the convolution operation W 1i *x i +b 1 : Extract key features of different scales through convolution kernels of different sizes; exponential decay term Control the weight distribution of features so that the weight of features near the center point is larger;

[0117] Where N is the number of feature maps, which can be pre-set by the model architecture; W 1i is the convolution kernel corresponding to the i-th feature map, which is optimized by the back propagation algorithm during training; x i is the i-th feature map, obtained directly from the input image or the output of the previous network layer; b 1 is the bias term, which is also optimized by the back propagation algorithm during training; α is the scaling factor, which can be adjusted based on experiments to achieve the best performance; γ is the parameter that controls the exponential decay rate, which also needs to be determined through experiments; c is the center point value, which represents the expected feature center position, usually set to 0.5 or adjusted according to the actual problem; x k is the jth feature map, obtained directly from the input image or the output of the previous network layer.

[0118] In enhanced feature representation, nonlinear transformation The new convolution kernel further enhances the feature expression ability; the sine wave oscillation term Introducing periodic changes to enhance the diversity of features; logarithmic terms Increase the dynamic range of the feature through the logarithmic function; S-type term Smooth nonlinear transformation introduced by sigmoid function; square root term The variance of the feature is controlled by the square root function;

[0119] Among them, K is the number of feature maps after processing, which is determined by the model architecture; W 2k is the convolution kernel corresponding to the kth feature map, which is learned during the training process; T k (x) is the kth processed feature map, which comes from the output of the previous step; b 2is the bias term, which is optimized during training; β,ω,d,η,P,m,δ,θ,Q,n,φ,R,o,e are various parameters, including scaling factor, frequency, center point, etc., which need to be tuned according to specific tasks and experimental results;

[0120] Assume that a green food production line needs to perform quality control on food packaging and collects a large number of images as image datasets;

[0121] Assume that the parameter N = 5, W 1i Assume randomly initialized 3×3 and 5×5 convolution kernels, x i is the i-th feature map, obtained directly from the input image or the output of the previous network layer, b 1 =0.2, α=0.5, γ=0.1, c=0.5; assuming x i =[0.1,0.2,0.3,0.4,0.5]; Assume K = 4; W 2k is the convolution kernel corresponding to the kth feature map, assuming it is a randomly initialized 3×3 convolution kernel; T k (x) is the kth processed feature map, which comes from the output of the previous step; b 2 =0.3, β=0.4, ω=2, d=0.7:, η=0.3, P=2, δ=0.2, θ=1, Q=3, φ=0.1, R=2, e=0.6, T k (x) = [0.7, 0.8, -0.6, 0.9];

[0122]

[0123] Assume that the threshold is set to 0.8. Since most of the feature values ​​are higher than the preset threshold of 0.8, it indicates that the currently inspected package has obvious quality problems, such as cracks, contamination, or label errors. The high-value feature representation indicates that the features contained in the input image are very consistent with the expected defect pattern. Therefore, it can be inferred that there are significant abnormalities in the packaging material. Through the above steps, the production line can more accurately detect minor defects in packaging materials, such as cracks, contamination, or label errors, to ensure the consistency and stability of product quality.

[0124] Optionally, based on the contrast enhanced image data set, combined with super-resolution reconstruction technology, high-resolution image detail prediction is performed to improve image resolution and clarity, and generate a high-resolution image, including: based on the contrast enhanced image data set, image enlargement processing is performed by an image upsampling method to generate a preliminary image frame; based on the preliminary image frame, combined with the detail information of the contrast enhanced image data set, interpolation technology is used to fill the missing pixels of the image to generate an optimized image frame; based on the optimized image frame, the super-resolution reconstruction technology deep learning mechanism is used to optimize the optimized image frame, predict and supplement potential subtle structures and texture features, and generate a reconstructed image frame; based on the reconstructed image frame, the overall clarity and resolution are adjusted to ensure that there is no distortion during the image enlargement process to generate a high-resolution image.

[0125] The contrast enhanced image dataset refers to the image dataset that has been processed with contrast enhancement to improve the visual effect of the image.

[0126] Super-resolution reconstruction technology is an image processing technology that converts low-resolution images into high-resolution images through algorithms to improve image clarity and detail.

[0127] Image upsampling methods refer to techniques for upscaling low-resolution images to higher resolutions through interpolation or other methods.

[0128] Interpolation technology refers to the method of estimating unknown data points between known data points, and is often used for pixel filling during image enlargement.

[0129] Deep learning mechanism refers to the technology of using deep neural network for feature learning and prediction.

[0130] In the embodiments of the present application, firstly, based on the contrast enhanced image data set, an image enlargement process is performed by an image upsampling method to generate a preliminary image frame; secondly, based on the preliminary image frame, in combination with the detailed information of the contrast enhanced image data set, an interpolation technique is used to fill in the missing pixels of the image to generate an optimized image frame; thirdly, based on the optimized image frame, a deep learning mechanism of super-resolution reconstruction technology is used to further optimize the optimized image frame, predict and supplement potential subtle structures and texture features, and generate a reconstructed image frame; finally, based on the reconstructed image frame, the overall clarity and resolution are adjusted to ensure that the image is not distorted during the enlargement process, and a final high-resolution image is generated.

[0131] Suppose a food processing company needs to perform quality control on meat products on the production line to ensure product safety and consistency;

[0132] First, the contrast-enhanced image dataset is enlarged using an image upsampling method (such as bilinear interpolation) to generate a preliminary image frame. Secondly, based on the preliminary image frame and combined with the detail information in the contrast-enhanced image dataset, a more advanced interpolation technique (such as bicubic interpolation) is used to fill in the missing pixels in the image to generate an optimized image frame. Thirdly, based on the optimized image frame, a pre-trained super-resolution reconstruction model is used to further optimize the optimized image frame through a deep learning mechanism to predict and supplement potential subtle structures and texture features to generate a reconstructed image frame. Finally, based on the reconstructed image frame, the overall clarity and resolution are adjusted to ensure that the image is not distorted during enlargement and generate a final high-resolution image.

[0133] Through the above steps, the generated high-resolution images not only have higher clarity and rich details, but also maintain the quality of the original images, providing a high-quality data foundation for subsequent anomaly detection and quality control.

[0134] 103. Based on the high-quality image data set, use a deep residual network to perform anomaly detection processing, identify abnormal areas of the high-quality image data set, perform dynamic analysis in combination with time series data analysis technology, and generate anomaly detection results;

[0135] The deep residual network is a deep learning model that solves the gradient vanishing problem by introducing residual blocks and enhances the learning ability of the model.

[0136] Anomaly detection refers to identifying abnormal areas or defects in images.

[0137] Time series data analysis technology refers to the analysis of time series data to capture the patterns and trends of data changes over time.

[0138] The anomaly detection result refers to the abnormal area and its related information identified after the above processing.

[0139] In this step, first, the gradient vanishing problem in deep learning is solved by introducing residual blocks, the learning ability of the model is enhanced, and a preprocessed image data set is generated; secondly, a deep residual network is used to perform forward propagation calculations on the preprocessed image data set to generate preliminary anomaly detection results; thirdly, based on the preliminary anomaly detection results, the time series patterns of abnormal events are identified, dynamic analysis is performed, and dynamic analysis results are generated; finally, a comprehensive evaluation is performed based on the multi-dimensional key features of the abnormal events to generate the final anomaly detection results.

[0140] Optionally, in step 103, based on the high-quality image dataset, a deep residual network is used to perform anomaly detection processing, abnormal areas of the high-quality image dataset are identified, dynamic analysis is performed in combination with time series data analysis technology, and anomaly detection results are generated, including: based on the high-quality image dataset, a residual block is introduced to solve the gradient vanishing problem of deep learning, the model learning ability is enhanced, and a preprocessed image dataset is generated; based on the preprocessed image dataset, a deep residual network is used to perform anomaly detection processing, anomaly probability values ​​are calculated through forward propagation, and preliminary anomaly detection results are generated; based on the preliminary anomaly detection results, the time series pattern of abnormal events is identified, dynamic analysis is performed, and dynamic analysis results are generated; based on the dynamic analysis results, a comprehensive evaluation is performed in combination with multi-dimensional key features of abnormal events to generate anomaly detection results.

[0141] Among them, based on the preprocessed image data set, a deep residual network is used to perform anomaly detection processing, and anomaly probability values ​​are calculated through forward propagation to generate preliminary anomaly detection results, including: based on the preprocessed image data set, a directional gradient histogram method is used to perform key feature extraction processing to generate a feature representation data set; based on the feature representation data set, a deep residual network is used to calculate feature representation through forward propagation traversal to generate a deep feature representation data set; based on the deep feature representation data set, in combination with the deep residual network classification layer, anomaly probability values ​​are obtained to generate anomaly probability value data set; based on the anomaly probability value data set, a reasonable threshold interval is preset for threshold judgment, the anomaly probability value is marked and recorded, and a preliminary anomaly detection result is generated.

[0142] A high-quality image dataset refers to a collection of images with clear details and high resolution after a series of processing (such as denoising, contrast enhancement, etc.).

[0143] The deep residual network is a neural network architecture that introduces residual blocks to alleviate the gradient vanishing problem that occurs during deep network training, thereby improving the learning ability of the model.

[0144] Anomaly detection is the process of identifying data points or areas that are significantly different from normal patterns based on machine learning or deep learning methods.

[0145] Time series data analysis technology is a technology that analyzes data that changes over time to discover trends, periodicities, or other patterns.

[0146] The histogram of oriented gradients is a feature descriptor commonly used in object detection. It describes the shape of an object by calculating the statistical information of the gradient direction in a local area of ​​the image.

[0147] In an embodiment of the present application, first, a high-quality image data set is preprocessed, and a residual block is used to solve the gradient vanishing problem, so as to generate a preprocessed image that is easier to process in subsequent steps; secondly, key features are extracted from the preprocessed image, using a histogram of oriented gradients method, and further converted into a deep feature representation through a deep residual network; thirdly, based on these deep features, the probability value of each sample becoming an anomaly is calculated, and a preliminary anomaly detection result is determined according to a set threshold; finally, combined with timing analysis technology, the time series pattern of abnormal events is dynamically analyzed, the abnormal situation is comprehensively evaluated, and a final anomaly detection report is formed.

[0148] A frozen food company wants to improve product quality by monitoring the integrity of product packaging on its production line. The company collects a large number of product packaging images as a high-quality image dataset;

[0149] Firstly, the residual block structure in the deep residual network is used to preprocess the dataset, which enhances the model's learning ability for complex patterns and generates a preprocessed image dataset; secondly, the oriented gradient histogram method is used to extract key visual features from the preprocessed images and construct a feature representation dataset; thirdly, this feature representation is input into the deep residual network, and the deep feature representation of each image and its corresponding abnormal probability value are obtained through the forward propagation process. When the probability exceeds the preset threshold, it is marked as a potential anomaly; finally, combined with the time series data analysis technology, the occurrence time distribution characteristics of the marked anomalies are analyzed, and factors such as the frequency, duration and interval of abnormal events are considered to conduct a comprehensive evaluation and generate a detailed anomaly detection report.

[0150] Through the above steps, the company can promptly detect and repair product packaging defects that may occur on the production line, ensuring product quality standards.

[0151] This application takes into account that in the feature representation calculation based on the deep residual network, the feature representation is calculated by forward propagation traversal to generate a deep feature representation dataset. This process includes preprocessing the input feature map to reduce noise and highlight key information, capturing complex patterns through complex nonlinear transformations, and generating the final deep feature representation dataset through standardization and dimensionality reduction techniques.

[0152] Optionally, based on the feature representation data set, a deep residual network is used to calculate feature representations through forward propagation to generate a deep feature representation data set, including:

[0153] Based on the feature representation data set, preprocess the input feature map, reduce input noise through filtering technology, highlight key information of the input feature map, and generate an intermediate output vector;

[0154] The intermediate output vector is calculated using the following formula:

[0155]

[0156] Among them, A i is the intermediate output vector; N is the number of input feature maps; W ij is the convolution kernel between the i-th output vector and the j-th input feature map; X j is the jth feature map in the input feature vector; b i is the bias term of the i-th output vector; σ is the ReLU activation function; α and β are parameters that control the strength of the nonlinear term; M is the number of input feature maps involved in the nonlinear transformation; V ik is the weight for nonlinear transformation between the i-th output vector and the k-th input feature map; j is the index of the input feature map, from 1 to N; k is the index of the input feature map, from 1 to M; X j Represents the kth input feature map participating in the nonlinear transformation;

[0157] Based on the intermediate output vector, multiple levels of nonlinear transformations are used to capture complex patterns, and additional convolutional layers and pooling layers are introduced to increase expression power and generalization performance to generate a final output vector;

[0158] The final output vector is calculated using the following formula:

[0159]

[0160] Among them, D i is the final output vector; N is the number of input feature maps; W' il is the convolution kernel between the i-th output vector and the l-th intermediate vector; A l is the lth intermediate vector; A m is the mth intermediate vector; A n is the nth intermediate vector; A o is the oth intermediate vector; b' i is the bias term of the i-th output vector; σ is the ReLU activation function; γ, δ and θ are parameters that control the strength of the nonlinear term; O is the number of intermediate vectors involved in the nonlinear transformation; U im is the weight used for nonlinear transformation between the ith output vector and the mth intermediate vector; ∈ is the parameter controlling the strength of the fractional term; P and Q are the number of intermediate vectors involved in the operation in the numerator and denominator respectively; Z in and Y io are the weights between the i-th output vector and the n-th or o-th intermediate vector in the numerator and denominator respectively; l is the index of the input feature map, from 1 to N; m is the index of the intermediate vector, from 1 to O; n is the index of the intermediate vector, from 1 to P; o is the index of the intermediate vector, from 1 to Q;

[0161] Based on the final output vector, standardization processing is performed, dimension reduction is performed through principal component analysis, the high-dimensional feature vector is mapped into a low-dimensional space, and visualization analysis is performed to generate a deep feature representation data set.

[0162] This method aims to generate intermediate output vectors through complex nonlinear transformations (such as logarithmic terms and sine terms), introduce additional convolutional layers and pooling layers for multi-level nonlinear transformations, enhance the expressiveness and generalization performance of the model, and map high-dimensional features to low-dimensional space through standardization and principal component analysis dimensionality reduction for easy visualization and further analysis.

[0163] In the intermediate output vector, the convolution operation W ij ×X j : Extract key information of input feature map through convolution kernel; bias term b i : Introduce an offset for each output vector to increase the flexibility of the model; nonlinear terms Enhance feature expression capabilities through nonlinear transformation and control feature weight distribution;

[0164] Where N is the number of input feature maps, which is determined by the model architecture; W ij is the convolution kernel between the i-th output vector and the j-th input feature map, which is optimized by the back propagation algorithm during the training process; X j is the jth feature map in the input feature vector, obtained directly from the input image or the output of the previous network layer; b i is the bias term of the i-th output vector, which is also optimized by the back propagation algorithm in the training process; α, β are parameters that control the strength of the nonlinear term, which are usually determined by experimental tuning; M is the number of input feature maps involved in the nonlinear transformation, which is set according to the actual task requirements; V ik is the weight used for nonlinear transformation between the i-th output vector and the k-th input feature map, which is optimized by the back propagation algorithm during training;

[0165] In the final output vector, the additional convolutional layer and the pooling layer W' il ×A l +b' i :Capture more complex patterns through multiple levels of nonlinear transformations, increase the model's expressiveness and generalization performance; the nonlinear term γ·sin(δ·∑U im ·A m +θ): Introducing periodic changes to enhance the diversity of features; fractional terms Adjust the importance of features through fractional terms and control the dynamic range of features;

[0166] Where N is the number of input feature maps, which is determined by the model architecture; W' il is the convolution kernel between the ith output vector and the lth intermediate vector, which is optimized by the back propagation algorithm during the training process; A l is the lth intermediate vector, which comes from the output of the previous step; b' i is the bias term of the i-th output vector, which is also optimized by the back propagation algorithm during the training process; γ, δ, θ are parameters that control the strength of the nonlinear term, which are usually determined by experimental tuning; O is the number of intermediate vectors involved in the nonlinear transformation, which is set according to the actual task requirements; U im is the weight used for nonlinear transformation between the ith output vector and the mth intermediate vector, which is optimized by the back propagation algorithm during training; ∈ is the parameter controlling the strength of the fractional term, which is usually determined by experimental tuning; P, Q are the number of intermediate vectors involved in the operation in the numerator and denominator, which are set according to the actual task requirements; Z in ,Y io is the weight between the ith output vector and the nth or oth intermediate vector in the numerator and denominator, which is optimized by the back propagation algorithm during the training process;

[0167] Assume that a fresh food processing company needs to monitor the product quality on its production line in real time, especially the quality control of the packaging process;

[0168] Assume parameter N = 3; W ij is a randomly initialized 3×3 convolution kernel; X j =[0.1,0.2,0.3]; b i =0.2; α=0.5; β=0.2; M=2; V ik are randomly initialized weights;

[0169]

[0170] Assume parameter N = 3; W' il A is a randomly initialized 3×3 convolution kernel; l =[1.02,1.02,1.02]; b' i =0.3; γ = 0.4; δ = 0.3; θ = 0.1; O = 2; U im are randomly initialized weights; ∈ = 0.3; P = 2; Q = 2; Z in ,Y io are randomly initialized weights;

[0171] Due to D iThe result of 4.4229 is significantly higher than the preset threshold of 4.0, which indicates that the features in the input image are highly matched with the known defect patterns, which means that there are obvious quality problems in the packaging materials, such as breakage, impurity contamination or inaccurate label information. Through the above steps, the production line can achieve high-precision detection of tiny defects in packaging materials, specifically covering typical defect types such as cracks, contamination and label errors, to ensure the uniformity and long-term stability of product quality.

[0172] 104. Based on the abnormal detection results, the abnormal detection results are sent to the food production line control system in real time through an encrypted interface to support timely adjustment of production parameters and generate food quality control strategies.

[0173] An encrypted interface is a secure data transmission protocol that ensures that data is not tampered with or leaked during transmission.

[0174] Food production line control system refers to a system used to monitor and control food production lines, capable of adjusting production parameters based on the data received.

[0175] Food quality control strategy refers to a series of measures formulated based on abnormal detection results to ensure stable product quality and optimize the production process.

[0176] In this step, first, the anomaly detection results are securely transmitted through an encrypted interface to ensure the security of data during transmission; second, the food production line control system receives and parses these anomaly detection results; third, the control system automatically generates corresponding production parameter adjustment instructions based on the anomaly detection results; finally, the control system executes these instructions, adjusts production parameters in a timely manner, and generates food quality control strategies to ensure the stability of product quality and optimize the production process.

[0177] Optionally, the abnormal detection result in step 104 is sent to the food production line control system in real time through an encrypted interface to support timely adjustment of production parameters and generation of food quality control strategies, including: based on the abnormal detection result, the abnormal detection report is transmitted to the food production line control system in real time using an encrypted interface to generate an encrypted transmission result; based on the encrypted transmission result, a comprehensive analysis is performed in combination with the production line historical data and the current production status to generate production line adjustment suggestions; based on the production line adjustment suggestions, the production parameters are automatically adjusted by the food production line control system to generate adjusted production parameter settings; based on the adjusted production parameter settings, actual operating performance indicators are recorded, combined with the abnormal detection results, and continuous adjustment and optimization are performed through a feedback mechanism to generate a food quality control strategy.

[0178] The encryption interface is a secure way of data transmission, ensuring that data is not tampered with or leaked during transmission.

[0179] Food production line control systems are systems used to monitor and control food production lines, and are able to adjust production parameters based on the data received.

[0180] Production line historical data includes past production records, equipment status, quality inspection results and other information.

[0181] The current production status refers to the current operating status of the production line, including equipment operating status, production speed, product quality, etc.

[0182] Comprehensive analysis combines multiple data sources for multi-dimensional analysis to generate more accurate recommendations.

[0183] The feedback mechanism is a continuous optimization process that monitors actual results and continuously adjusts strategies to achieve optimal performance.

[0184] In the embodiment of the present application, first, the anomaly detection report is transmitted to the food production line control system in real time using an encrypted interface to ensure data security; secondly, based on the results of the encrypted transmission, combined with the historical data of the production line and the current production status, a comprehensive analysis is performed to generate specific production line adjustment suggestions; thirdly, the production parameters are automatically adjusted through the food production line control system to respond to these adjustment suggestions; finally, the actual operating performance indicators are recorded, and combined with the anomaly detection results, continuous adjustment and optimization are performed through a feedback mechanism to generate the final food quality control strategy.

[0185] Suppose a food processing plant needs to perform quality control on the biscuit packaging process on the production line. The plant has generated anomaly detection results through image analysis technology;

[0186] First, the anomaly detection report is transmitted to the food production line control system in real time using an encrypted interface to ensure the security of data during transmission and generate encrypted transmission results. Second, based on the encrypted transmission results, combined with the historical data of the production line (such as previous packaging defect records, equipment maintenance records) and the current production status (such as equipment operating conditions, production speed), a comprehensive analysis is conducted to generate specific production line adjustment suggestions, such as adjusting the speed of the packaging machine, increasing the frequency of quality inspections, etc. Third, the production parameters are automatically adjusted through the food production line control system, such as slowing down the speed of the packaging machine and increasing the sensitivity of the sensor in response to these adjustment suggestions. Finally, the actual operating performance indicators, such as packaging defect rate, production efficiency, etc., are recorded, and combined with the anomaly detection results, the production parameters are continuously adjusted and optimized through the feedback mechanism to generate the final food quality control strategy to ensure the stability of the production line and the consistency of product quality.

[0187] Through the above steps, the factory can promptly identify and solve problems in the packaging process, and improve product quality and production efficiency.

[0188] Figure 2 A structural schematic diagram of a food quality control system based on intelligent image analysis is provided for an embodiment of the present application, such as Figure 2 As shown, the device comprises:

[0189] A collection module 21, used to collect multi-angle and multi-spectral real-time image streams at different positions of the food production line through a multi-sensor system to form a comprehensive image data set;

[0190] A processing module 22 is used to perform image denoising and contrast enhancement processing based on the comprehensive image data set by using a convolutional autoencoder, and to improve image clarity and detail performance by combining super-resolution reconstruction technology to generate a high-quality image data set;

[0191] An analysis module 23 is used to perform anomaly detection processing based on the high-quality image data set using a deep residual network, identify abnormal areas of the high-quality image data set, perform dynamic analysis in combination with time series data analysis technology, and generate anomaly detection results;

[0192] The sending module 24 is used to send the abnormal detection results to the food production line control system in real time through an encrypted interface to support timely adjustment of production parameters and generate food quality control strategies.

[0193] Figure 2 The food quality control system based on intelligent image analysis can be implemented Figure 1 The implementation principle and technical effect of the food quality control method based on intelligent image analysis described in the embodiment shown will not be repeated. The specific way in which each module and unit performs operations in the food quality control system based on intelligent image analysis in the above embodiment has been described in detail in the embodiment of the method, and will not be elaborated here.

[0194] In one possible design, Figure 2 A food quality control system based on intelligent image analysis of the embodiment shown can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32;

[0195] The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32 .

[0196] The processing component 32 is used to: collect multi-angle and multi-spectral real-time image streams at different positions of the food production line through a multi-sensor system to form a comprehensive image data set; based on the comprehensive image data set, use a convolutional autoencoder to perform image denoising and contrast enhancement processing, combine super-resolution reconstruction technology to improve image clarity and detail expression, and generate a high-quality image data set; based on the high-quality image data set, use a deep residual network to perform anomaly detection processing, identify abnormal areas of the high-quality image data set, combine time series data analysis technology to perform dynamic analysis, and generate anomaly detection results; based on the anomaly detection results, send them to the food production line control system in real time through an encrypted interface to support timely adjustment of production parameters and generate food quality control strategies.

[0197] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to perform the above method.

[0198] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0199] Of course, the computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0200] The input / output interface provides an interface between the processing component and the peripheral interface module, which may be an output device, an input device, etc.

[0201] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0202] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0203] The present application also provides a computer storage medium storing a computer program, wherein the computer program can achieve the above-mentioned Figure 1 The illustrated embodiment is a food quality control method based on intelligent image analysis.

[0204] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0205] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0206] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0207] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A food quality control method based on intelligent image analysis, characterized in that: include: The multi-sensor system collects multi-angle and multi-spectral real-time image streams at different locations of the food production line to form a comprehensive image data set; Based on the comprehensive image dataset, a convolutional autoencoder is used to perform image denoising and contrast enhancement processing, and super-resolution reconstruction technology is combined to improve image clarity and detail expression, thereby generating a high-quality image dataset; Based on the high-quality image data set, a deep residual network is used to perform anomaly detection processing, abnormal areas of the high-quality image data set are identified, and dynamic analysis is performed in combination with time series data analysis technology to generate anomaly detection results; Based on the abnormal detection results, they are sent to the food production line control system in real time through an encrypted interface to support timely adjustment of production parameters and generate food quality control strategies.

2. The method according to claim 1, characterized in that Based on the comprehensive image dataset, the convolutional autoencoder is used to perform image denoising and contrast enhancement processing, and the super-resolution reconstruction technology is combined to improve the image clarity and detail expression, and generate a high-quality image dataset, including: Based on the comprehensive image data set, performing a convolutional autoencoder random noise removal process to improve image purity and generate a denoised image data set; Based on the denoised image dataset, the nonlinear mapping capability of the convolutional autoencoder is utilized to analyze and adjust the image histogram, perform contrast enhancement processing, and generate a contrast enhanced image dataset; Based on the contrast enhanced image data set, combined with super-resolution reconstruction technology, high-resolution image detail prediction is performed to improve image resolution and clarity and generate high-resolution images; Based on the high-resolution image, the local contrast is adjusted, edge sharpening is performed, and the microscopic level is enhanced to improve the image detail expression and generate a high-quality image data set.

3. The method according to claim 2, characterized in that The method of analyzing and adjusting the image histogram based on the denoised image data set and performing contrast enhancement processing by utilizing the nonlinear mapping capability of the convolutional autoencoder to generate a contrast enhanced image data set includes: Based on the denoised image dataset, a multi-scale key feature extraction process is performed by pre-training a convolutional autoencoder to generate image dataset features; Based on the features of the image dataset, the nonlinear mapping capability of the convolutional autoencoder is used to transform the features of the image dataset to adapt to the image histogram adjustment requirements, thereby generating a feature-transformed image dataset; Based on the feature transformation image data set, analyzing the specific distribution of the image histogram, identifying the contrast enhancement optimization space, and generating an image histogram analysis result; Based on the image histogram analysis result, the brightness of the low-light area is enhanced and the contrast of the high-contrast area is adjusted to improve the overall visual effect and generate a contrast-enhanced image data set.

4. The method according to claim 3, characterized in that Based on the features of the image dataset, the nonlinear mapping capability of the convolutional autoencoder is used to transform the features of the image dataset to adapt to the image histogram adjustment requirements, and generate a feature-transformed image dataset, including: Based on the features of the image dataset, convolution operations are performed on the features of the image dataset using convolution kernels of different sizes to extract key features of different scales; The key features of different scales are integrated, noise is suppressed by a smoothing filter, and high-frequency noise is removed to generate an intermediate result of feature change; The intermediate result of feature transformation is calculated by the following formula: Where T(x) is the intermediate result of feature transformation; i is the index of the feature map, from 1 to N; N is the number of feature maps; W 1i is the convolution kernel corresponding to the i-th feature map; x j is the i-th feature map; b1 is the bias term; tanh is the hyperbolic tangent activation function; Δ is the scaling factor; γ is the parameter that controls the exponential decay speed; j is the feature index participating in the exponential decay calculation, from 1 to M; M is the number of features participating in the exponential decay calculation; c is the center point value; x j is the jth feature map; Based on the intermediate result of the feature transformation, further nonlinear transformation is performed to enhance the feature expression capability, and the most representative subset is obtained through a feature selection method to reduce redundant information to generate an enhanced feature representation; The enhanced feature representation is calculated using the following formula: Where EFR(T(x)) is the enhanced feature representation; k is the processed feature map index, from 1 to K; K is the number of processed feature maps; W 2k is the convolution kernel corresponding to the kth feature map; T k (x) is the kth processed feature map; b2 is the bias term; σ is the activation function; β is the scaling factor; ω is the parameter that controls the sine wave oscillation frequency; l is the feature index involved in the sine wave calculation, from 1 to L; L is the number of features involved in the sine wave calculation; T l (x) is the lth processed feature map; d is the center point value of the sine wave; η is the weight coefficient of the logarithmic term; m is the feature index for logarithmic calculation, from 1 to P; P is the number of features involved in logarithmic calculation; |T m (x)| is the absolute value of the mth feature map; δ is the scaling factor of the sigmoid term; θ is the parameter that controls the slope of the sigmoid function; n is the feature index involved in the sigmoid calculation, from 1 to Q; Q is the number of features involved in the sigmoid calculation; T n (x) is the nth processed feature map; φ is the scaling factor of the square root term; o is the feature index involved in the square root calculation, from 1 to R; R is the number of features involved in the square root calculation; e is the center point value of the square root calculation; T o (x) is the oth processed feature map; Based on the enhanced feature representation, the image space is restored through inverse transformation, and histogram adjustment is performed to ensure image quality and visual effect, thereby generating a feature transformation image dataset.

5. The method according to claim 2, characterized in that: The method of predicting high-resolution image details based on the contrast-enhanced image data set and combining super-resolution reconstruction technology to improve image resolution and clarity and generate a high-resolution image includes: Based on the contrast enhanced image data set, image enlargement processing is performed by an image upsampling method to generate a preliminary image frame; Based on the preliminary image frame and in combination with the detail information of the contrast enhanced image data set, an interpolation technique is used to fill in the missing pixels of the image to generate an optimized image frame; Based on the optimized image frame, the optimized image frame is optimized by using the deep learning mechanism of super-resolution reconstruction technology, and potential subtle structures and texture features are predicted and supplemented to generate a reconstructed image frame; Based on the reconstructed image framework, the overall clarity and resolution are adjusted to ensure that the image is not distorted during the image enlargement process and generate a high-resolution image.

6. The method according to claim 1, characterized in that Based on the high-quality image data set, the deep residual network is used to perform anomaly detection processing, the abnormal area of ​​the high-quality image data set is identified, and the time series data analysis technology is combined to perform dynamic analysis to generate anomaly detection results, including: Based on the high-quality image dataset, a residual block is introduced to solve the gradient vanishing problem in deep learning, enhance the model learning ability, and generate a preprocessed image dataset; Based on the preprocessed image data set, a deep residual network is used to perform anomaly detection processing, anomaly probability values ​​are calculated through forward propagation, and preliminary anomaly detection results are generated; Based on the preliminary anomaly detection results, identify the time series pattern of abnormal events, perform dynamic analysis, and generate dynamic analysis results; Based on the dynamic analysis results, a comprehensive evaluation is performed in combination with the multi-dimensional key features of the abnormal event to generate anomaly detection results.

7. The method according to claim 6, characterized in that Based on the pre-processed image data set, the deep residual network is used to perform anomaly detection processing, and the anomaly probability value is calculated by forward propagation to generate a preliminary anomaly detection result, including: Based on the preprocessed image data set, a key feature extraction process is performed using a directional gradient histogram method to generate a feature representation data set; Based on the feature representation data set, a deep residual network is used to calculate the feature representation through forward propagation to generate a deep feature representation data set; Based on the deep feature representation data set, combined with the deep residual network classification layer, anomaly probability values ​​are obtained to generate anomaly probability value data set; Based on the abnormal probability value data set, a reasonable threshold interval is preset for threshold judgment, and the abnormal probability value is marked and recorded to generate a preliminary abnormality detection result.

8. The method according to claim 7, characterized in that Based on the feature representation data set, a deep residual network is used to calculate the feature representation through forward propagation traversal to generate a deep feature representation data set, including: Based on the feature representation data set, preprocess the input feature map, reduce input noise through filtering technology, highlight key information of the input feature map, and generate an intermediate output vector; The intermediate output vector is calculated using the following formula: Among them, A i is the intermediate output vector; N is the number of input feature maps; W ij is the convolution kernel between the i-th output vector and the j-th input feature map; X j is the jth feature map in the input feature vector; b i is the bias term of the i-th output vector; σ is the ReLU activation function; α and β are parameters that control the strength of the nonlinear term; M is the number of input feature maps involved in the nonlinear transformation; V ik is the weight for nonlinear transformation between the i-th output vector and the k-th input feature map; j is the index of the input feature map, from 1 to N; k is the index of the input feature map, from 1 to M; X k Represents the kth input feature map participating in the nonlinear transformation; Based on the intermediate output vector, multiple levels of nonlinear transformations are used to capture complex patterns, and additional convolutional layers and pooling layers are introduced to increase expression power and generalization performance to generate a final output vector; The final output vector is calculated using the following formula: Among them, D i is the final output vector; N is the number of input feature maps; W' il is the convolution kernel between the i-th output vector and the l-th intermediate vector; A l is the lth intermediate vector; A m is the mth intermediate vector; A n is the nth intermediate vector; A o is the oth intermediate vector; b' i is the bias term of the i-th output vector; σ is the ReLU activation function; γ, δ and θ are parameters that control the strength of the nonlinear term; O is the number of intermediate vectors involved in the nonlinear transformation; U im is the weight used for nonlinear transformation between the ith output vector and the mth intermediate vector; ∈ is the parameter controlling the strength of the fractional term; P and Q are the number of intermediate vectors involved in the operation in the numerator and denominator respectively; Z in and Y io are the weights between the i-th output vector and the n-th or o-th intermediate vector in the numerator and denominator respectively; l is the index of the input feature map, from 1 to N; m is the index of the intermediate vector, from 1 to O; n is the index of the intermediate vector, from 1 to P; o is the index of the intermediate vector, from 1 to Q; Based on the final output vector, standardization processing is performed, dimension reduction is performed through principal component analysis, the high-dimensional feature vector is mapped into a low-dimensional space, and visualization analysis is performed to generate a deep feature representation data set.

9. The method according to claim 1, characterized in that: The abnormal detection result is sent to the food production line control system in real time through an encrypted interface to support timely adjustment of production parameters and generate food quality control strategies, including: Based on the abnormal detection result, the abnormal detection report is transmitted to the food production line control system in real time using an encrypted interface to generate an encrypted transmission result; Based on the encrypted transmission results, combined with the historical data of the production line and the current production status, a comprehensive analysis is performed to generate production line adjustment suggestions; Based on the production line adjustment suggestion, automatically adjusting production parameters through the food production line control system to generate adjusted production parameter settings; Based on the adjustment of production parameter settings, actual operating performance indicators are recorded, combined with the abnormal detection results, continuous adjustment and optimization are carried out through a feedback mechanism to generate a food quality control strategy.

10. A food quality control system based on intelligent image analysis, characterized in that: include: A collection module, used to collect multi-angle and multi-spectral real-time image streams at different positions of the food production line through a multi-sensor system to form a comprehensive image data set; A processing module, for performing image denoising and contrast enhancement processing based on the comprehensive image dataset using a convolutional autoencoder, combining super-resolution reconstruction technology to improve image clarity and detail expression, and generating a high-quality image dataset; An analysis module is used to perform anomaly detection processing based on the high-quality image data set using a deep residual network, identify abnormal areas of the high-quality image data set, perform dynamic analysis in combination with time series data analysis technology, and generate anomaly detection results; The sending module is used to send the abnormal detection results to the food production line control system in real time through an encrypted interface to support timely adjustment of production parameters and generate food quality control strategies.

Images