Image Denoising, Segmentation and Recognition Method for the Monitoring Information System of Meat Processing Production Line

Abstract

The invention discloses an image denoising, segmentation and recognition method for a monitoring information system of a meat product processing production line. It mainly includes: acquiring video image information of the meat product processing production line, directly storing and transmitting the collected images through a vision processing system; performing wavelet decomposition on the collected image signals by an image processing system, and removing noise wavelet coefficients through analysis and appropriate thresholding to achieve the purpose of retaining signals and filtering out noise; automatically selecting the initial edge of the image by a feedback strategy, and extracting the actual edge features of the image by iteratively solving the Markov transition probability and Gaussian parameters; segmenting the image according to the image edge features, judging the regions of pixel points through a global threshold, and recognizing the background and target in the image to complete the processing of image information. This method has high flexibility and accuracy, and can perform denoising, segmentation and recognition according to the features of the current image, and stably and reliably complete the image processing task.

Description

[0001] This invention is a divisional application of the invention patent with the title of "An Image Processing Method for the Monitoring Information System of a Meat Product Processing Production Line" and the application number of "201811142508X". Technical Field

[0002] This invention relates to an image processing method for a monitoring information system and belongs to the fields of computer vision and digital image processing. Background Art

[0003] China is a major producer and consumer of meat products in the world, but the development level of China's meat food industry lags far behind that of other developed countries. The lack of a large-scale meat deep processing industry and the imperfection of the existing image processing technology limit the development of intelligent production monitoring and detection of meat products through image recognition; the use of blurred images in the monitoring system cannot perform intelligent on-line detection, resulting in more waste of human and material resources and higher production costs. Summary of the Invention

[0004] To solve the above problems, the purpose of this invention is to provide an image processing method with flexibility and good accuracy.

[0005] The technical solution adopted by this invention to solve its problems includes the following steps:

[0006] A. Obtain the video image information of the meat product processing production line, and directly store and transmit the collected images through a vision processing system;

[0007] B. Use an image processing system to perform wavelet decomposition on the collected image signals, and remove the noise wavelet coefficients by analysis and appropriate thresholding to achieve the purpose of retaining the signal and filtering out the noise;

[0008] C. Recalculate the obtained edges repeatedly through a feedback method. According to the continuous increase of the transition probability and the Gaussian function, iterate multiple times to obtain an edge path closer to the actual edge. Extract the actual edge features of the image by iteratively solving the Markov transition probability and Gaussian parameters;

[0009] D. Segment the image according to the image edge features, determine the regions of pixel points through global thresholding, identify the background and target in the image, and complete the processing of the image information.

[0010] Further, the step A includes:

[0011] Set a diffuse reflection shadowless light source, and irradiate the meat product processing process with the refracted light of the LED light through a refraction plate. Collect the monitoring image information of the processing process through a camera;

[0012] The acquired images are directly stored by the visual processing system, and using digital transmission technology and large-scale integrated circuits, multiple signals of multiple images are transmitted to the image processing system through a single optical fiber, achieving real-time transmission while improving the stability of image transmission.

[0013] Further, the step B includes:

[0014] (1) By the scaling and translation transformation of the basic wavelet function J(x), the wavelet transform of the image signal to be analyzed at different scales is constructed;

[0015] ① The basic wavelet function J(x) undergoes translation transformation at different scales to construct a wavelet sequence:

[0016]

[0017] where s represents the scale scaling factor, s≠0, p is the translation change factor, and s, p∈R, R is the set of real numbers, and x represents the image information;

[0018] ② The wavelet transform formula of any image f(x) to be analyzed at scale s can be expressed as:

[0019]

[0020] where s, p∈Z represent any possible scaling and translation transformations;

[0021] (2) Use the soft threshold function to perform thresholding on the wavelet transform coefficients, retain the signal wavelet coefficients and remove the noise wavelet coefficients to achieve image denoising;

[0022] ① Assume that the video image is a two-dimensional matrix. Then, after each wavelet transform in step B(1), the image is decomposed into 4 sub-block frequency band regions of the same size;

[0023] ② According to the statistical characteristics of a set of wavelet coefficients of the image, select an appropriate threshold ω to perform thresholding on the decomposed sub-band regions. According to the formula

[0024]

[0025] calculate the threshold for image denoising, where Num i represents the number of wavelet coefficients in the i-th layer frequency band, represents the variance of the noise in the i-th layer frequency band, and i represents the number of frequency bands of image decomposition;

[0026] ③ Use the soft threshold function to remove the wavelet coefficients less than the threshold ω and perform a reduction transformation on the wavelet coefficients greater than the threshold ω. The soft threshold function is:

[0027]

[0028] Among them, X represents the image wavelet coefficient. If the error between the wavelet coefficient without interference noise and the denoised wavelet coefficient in the i-th layer frequency band reaches the minimum, then the threshold ω reaches the optimum, and the optimum denoising process is achieved; otherwise, according to the formula

[0029]

[0030] Perform the thresholding process for the next layer.

[0031] Furthermore, the step C includes:

[0032] (1) Use the path and path measure method to obtain the boundary of the denoised image or the boundary of an object in the image, and perform sequential search of the image;

[0033] ① The path in a two-dimensional image can be represented as an ordered set, including the starting node, the starting direction, and the path direction:

[0034] path = <(n1, n2), dir, [s1, …, s n >

[0035] Among them, (n1, n2) represents the starting node coordinates, dir represents the starting direction, and [s1, …, s n belongs to the direction set S = [Left, Mediate, Right];

[0036] ② According to the Markov transition probability and the Gaussian function, calculate the possible occurrence probability of the path to complete the search of the path. The Markov transition probability is:

[0037] P trans (path) = P trans (z m z m-1 )P trans (z m-1 z m-2 )…P trans (z1z0)

[0038] Among them, Z = (z0, z1, … z m ) represents the space of all possible state sequence spaces, and m represents the number of state sequences;

[0039] The value of the Gaussian function is determined by the value at the position of the starting node. When the node is on the edge, the Gaussian function is p b = exp(-(path - μ b ) 2 / 2σ b 2 )), where path represents the ordered set of paths in the two-dimensional image, μb , σ b represents the mean and standard deviation of the edge nodes; when the node is at any other position, the Gaussian function is p r = exp(-(path - μ r ) 2 / 2σ r 2 ), where μ r , σ br represent the mean and standard deviation of any node at other positions;

[0040] ③ The path measure method based on which the image is sequentially searched can be expressed as:

[0041]

[0042] where, represents the pixel value of the starting node;

[0043] (2) Adopt a feedback strategy to select the initial edge of the image, improving the automation degree of sequentially searching for the image edge and the accuracy of the initial edge.

[0044] Furthermore, the step D includes:

[0045] Adopt an approximation method to select an appropriate threshold to segment the target and background in the image according to the edge of the image;

[0046] ① Assume that the image contains two types of pixels, background and target. First, calculate the gray value H within the same edge region according to the edge of the image. The maximum gray value in the image is denoted as H max , and the minimum gray value is H min , then the initial threshold can be expressed as

[0047]

[0048] ② Assume that the background in the video image is darker. Then, according to the threshold O, the pixels with gray values less than O in the image are marked as background pixels, and similarly, the other pixels are marked as target pixels. Then, calculate the average gray values H back and H aim , then the new partitioning threshold is

[0049]

[0050] If O = O' + 1, then the image is segmented into background and target according to the magnitude of the gray value by this threshold. The pixels corresponding to the gray values greater than this threshold are target pixels, and the pixels corresponding to the gray values less than this threshold are background pixels; otherwise, repeat the partitioning calculation until O = O' + 1 holds to obtain the threshold.

[0051] The beneficial effects of the present invention are as follows:

[0052] In image processing with high complexity, the present invention can flexibly and accurately complete the preprocessing of images, and can perform denoising, segmentation, and recognition according to the characteristics of the current image, having the beneficial effects of practicability and stability. Description of the Drawings

[0053] Figure 1 It is the overall flowchart of an image denoising, segmentation, and recognition method for a monitoring information system of a meat product processing production line;

[0054] Figure 2 It is a schematic diagram of a sequential connection method based on feedback;

[0055] Figure 3 It is the algorithm flowchart for seeking the optimal edge path. Detailed Embodiment

[0056] Referring to Figure 1 , the method of the present invention includes the following steps:

[0057] A. Obtain the video image information of the meat product processing production line, directly store the collected images through digital transmission technology, and transmit them to the image processing system in real time;

[0058] (1) Set a diffuse reflection shadowless light source, irradiate the meat product processing process with the refracted light of the LED light through a refraction plate, and collect the monitoring image information of the processing process through a camera;

[0059] (2) Directly store the collected images with a visual processing system, and use digital transmission technology to transmit the image signals in real time;

[0060] ① During the transmission of the collected image information, various links such as video cables, encoders, and decoders are passed through, and delays are generated during the data exchange process, thus affecting the real-time performance of image transmission;

[0061] ② Using digital transmission technology and large-scale integrated circuits, multiple signals of multiple images are transmitted to the image processing system through a single optical fiber, realizing real-time transmission while improving the stability of image transmission;

[0062] B. Perform wavelet decomposition on the collected image signals with an image processing system, and remove the noise wavelet coefficients through analysis and appropriate thresholding to achieve the purpose of retaining the signal and filtering out the noise;

[0063] (1) Through the scaling and translation transformation of the basic wavelet function J(x), construct the wavelet transform of the image signal to be analyzed at different scales;

[0064] ① The basic wavelet function J(x) performs translation transformation at different scales to construct a wavelet sequence:

[0065]

[0066] Among them, s represents the scale scaling factor (when s, as a kind of scale, changes, it generates the characteristics of multi-resolution analysis), s≠0, p is the translation change factor, and s, p∈R, where R is the set of real numbers, and x represents the image information;

[0067] ② The wavelet transform formula of any image f(x) to be analyzed at scale s can be expressed as:

[0068]

[0069] Among them, s, p∈Z represents any possible scaling and translation transformation;

[0070] (2) Use the soft threshold function to perform thresholding on the wavelet transform coefficients, retain the signal wavelet coefficients and remove the noise wavelet coefficients to achieve image denoising;

[0071] ① Assume that the video image is a two-dimensional matrix. Then, after each wavelet transform in step B(1), the image is decomposed into 4 sub-block frequency band regions of the same size;

[0072] ② According to the statistical characteristics of a group of wavelet coefficients of the image, select an appropriate threshold ω to perform thresholding on the decomposed sub-frequency band regions. According to the formula

[0073]

[0074] calculate the threshold for image denoising, where Num i represents the number of wavelet coefficients in the i-th layer frequency band, represents the variance of the noise in the i-th layer frequency band, and i represents the number of frequency bands of image decomposition;

[0075] ③ Use the soft threshold function to remove the wavelet coefficients smaller than the threshold ω and perform a reduction transformation on the wavelet coefficients larger than the threshold ω. The soft threshold function is:

[0076]

[0077] Among them, X represents the image wavelet coefficients. If the error between the wavelet coefficients without interference noise and the denoised wavelet coefficients in the i-th layer frequency band reaches the minimum, then the threshold ω is optimal to achieve the optimal denoising process; otherwise, according to the formula

[0078]

[0079] perform the thresholding process for the next layer.

[0080] C. Automatically select the initial edges of the denoised image using a feedback strategy, and extract the actual edge features of the image by iteratively solving the Markov transition probability and Gaussian parameters;

[0081] (1) Use the path and path measure method to obtain the boundary of the denoised image or the boundary of an object in the image, and perform sequential search of the image;

[0082] ① The path in a two-dimensional image can be represented as an ordered set, including the starting node, starting direction, and path direction:

[0083] path = <(n1,n2),dir,[s1,…,s n >

[0084] where (n1,n2) represents the starting node coordinates, dir represents the starting direction, and [s1,…,s n belongs to the direction set S = [Left, Mediate, Right];

[0085] ② Calculate the possible occurrence probability of the path according to the Markov transition probability and Gaussian function to complete the search of the path. The Markov transition probability is:

[0086] P trans (path) = P trans (z m z m-1 )P trans (z m-1 z m-2 )…P trans (z1z0)

[0087] where Z = (z0,z1,…z m ) represents the space of all possible state sequence spaces, and m represents the number of state sequences;

[0088] The value of the Gaussian function is determined by the value at the position of the starting node. When the node is on the edge, the Gaussian function is p b = exp(-(path - μ b ) 2 / 2σ b 2 ), where path represents the ordered set of paths in the two-dimensional image, and μ b , σ b represent the mean and standard deviation of the edge nodes; when the node is at any other position, the Gaussian function is p r = exp(-(path - μ r ) 2 / 2σ r 2 ), where μ r , σbr represent the mean and standard deviation of any node at other positions;

[0089] ③ The path measure method based on which the image is searched sequentially can be expressed as:

[0090]

[0091] where represents the pixel value of the starting node;

[0092] (2) Adopt a feedback strategy to select the initial edge of the image, improving the automation degree of sequentially searching for the image edge and the accuracy of the initial edge;

[0093] ① Repeatedly recalculate the obtained edge through feedback. According to the increasing transition probability and Gaussian function, iterate about 8 times to obtain an edge path closer to the actual edge;

[0094] ② The algorithm flow of iterative operation using the feedback method is as Figure 3 shown.

[0095] D. Segment the image according to the image edge features, judge the region of pixel points through the global threshold, identify the background and target in the image, and complete the processing of the image information.

[0096] Adopt an approximation method to select an appropriate threshold to segment the target and background in the image according to the edge of the image;

[0097] ① Assume that the image contains two types of pixels, background and target. First, calculate the gray value H within the same edge region according to the edge of the image. The maximum gray value in the image is denoted as H max , and the minimum gray value is H min , then the initial threshold can be expressed as

[0098]

[0099] ② Assume that the background in the video image is darker. Then, according to the threshold O, the pixels with gray values less than O in the image are marked as background pixels, and similarly, the others are marked as target pixels, and the average gray values H back and H aim are calculated respectively. Then the new division threshold is

[0100]

[0101] If O = O'+1, then segment the background and target of the image according to the size of the gray value by this threshold. The pixels corresponding to the gray values greater than this threshold are target pixels, and the pixels corresponding to the gray values less than this threshold are background pixels; otherwise, repeat the division calculation until O = O'+1 holds to find the threshold;

[0102] In summary, an image denoising, segmentation and recognition method for a monitoring information system of a meat product processing production line is realized. In image processing with high complexity, the present invention can flexibly and accurately complete the preprocessing of images, and can perform denoising, segmentation and recognition according to the characteristics of the current image, having the beneficial effects of practicability and stability.

Claims

1. An image denoising, segmentation and recognition method for a monitoring information system of a meat product processing production line, characterized in that: The method includes the following steps: A. Obtain the video image information of the meat product processing production line, and directly store and transmit the collected images through a vision processing system; B. Perform wavelet decomposition on the collected image signals with an image processing system, and remove the noise wavelet coefficients by analysis and appropriate thresholding to achieve the purpose of retaining the signals and filtering out the noise; C. Recalculate the obtained edges repeatedly by a feedback method. According to the increasing transition probability and Gaussian function, iterate multiple times to obtain an edge path closer to the actual edge. Extract the actual edge features of the image by iteratively solving the Markov transition probability and Gaussian parameters; D. Segment the image according to the image edge features, determine the regions of pixel points through global thresholding, identify the background and target in the image, and complete the processing of the image information. The step D includes: Adopt an approximation method, select an appropriate threshold to segment the target and background in the image according to the image edge; ① Assume that the image contains two types of pixels, background and target. First, calculate the gray value H within the same edge region based on the edges of the image. The maximum gray value in the image is denoted as H max , and the minimum gray value is H min . Then the initial threshold can be expressed as ② Assume that the background in the video image is darker. Then, according to the threshold O, the pixels in the image with gray values less than O are marked as background pixels, and the other pixels are marked as target pixels. The average gray values H back and H aim are calculated. Then, the new partitioning threshold is E. If O = O'+1, then segment the background and target of the image according to the magnitude of the gray value by this threshold. The pixels corresponding to the gray values greater than this threshold are target pixels, and the pixels corresponding to the gray values less than this threshold are background pixels; otherwise, repeat the partitioning calculation until O = O'+1 holds to obtain the threshold.

2. The image denoising, segmentation and recognition method of the monitoring information system for the meat product processing production line according to claim 1, characterized in that: The step A includes: Set a diffuse reflection shadowless light source, irradiate the refracted light of the LED light through a refraction plate during the meat product processing process, and collect the monitoring image information of the processing process through a camera; Directly store the collected images with a vision processing system, and use digital transmission technology and large-scale integrated circuits to transmit multiple signals of multiple images through a single optical fiber to the image processing system, realizing real-time transmission while improving the image transmission stability.

3. The image denoising, segmentation and recognition method of the monitoring information system for the meat product processing production line according to claim 1 or 2, characterized in that: The step B includes: (1) Construct the wavelet transform of the image signal to be analyzed at different scales through the scaling and translation transformation of the basic wavelet function J(x); ① The basic wavelet function J(x) performs translation transformation at different scales to construct a wavelet sequence: where s represents the scale scaling factor, s≠0, p is the translation change factor, and s,p∈R, R is the real number, and x represents the image information; ② The wavelet transform formula of any image f(x) to be analyzed at scale s is expressed as: where s,p∈Z represents any possible scaling and translation transformation; (2) Use a soft threshold function to perform thresholding on the wavelet transform coefficients, retain the signal wavelet coefficients and remove the noise wavelet coefficients to achieve image denoising; ① Assume that the video image is a two-dimensional matrix. Then, after each wavelet transform in step B(1), the image is decomposed into 4 sub-block frequency band regions of the same size; ② According to the statistical characteristics of a group of wavelet coefficients of the image, select an appropriate threshold ω to perform thresholding on the decomposed sub-frequency band regions according to the formula Calculate the threshold for image denoising, where Num i represents the number of wavelet coefficients in the i-th layer frequency band, represents the variance of the noise in the i-th layer frequency band, and i represents the number of frequency bands for image decomposition; ③ Use the soft threshold function to remove the wavelet coefficients smaller than the threshold ω, and perform a reduction transformation on the wavelet coefficients greater than the threshold ω. The soft threshold function is: Among them, X represents the image wavelet coefficient. If the error between the wavelet coefficient without interference noise and the denoised wavelet coefficient on the i-th layer frequency band reaches the minimum, then the threshold ω reaches the optimum, achieving the optimal denoising process; otherwise, perform the thresholding process for the next layer according to the formula to perform the thresholding process for the next layer.

Images