Intelligent meat block size adjusting method based on image recognition and meat mincing device

Abstract

The invention belongs to the technical field of image recognition, and discloses an intelligent meat block size adjusting method based on image recognition and a meat mincing device.The method comprises the steps that image collection is conducted on meat products input into a target meat mincing device, and a corresponding spectral image sequence is obtained; performing image optimization processing on the obtained spectral image sequence to obtain a corresponding optimized meat block image sequence; performing image recognition on the obtained optimized meat block image sequence to obtain corresponding meat block recognition information, and setting a corresponding meat mincing control strategy based on the meat block recognition information; performing parameter adjustment on the corresponding meat mincing device based on the meat mincing control strategy, and performing information feedback based on a parameter adjustment result; according to the meat mincing device, the size of meat blocks of the meat mincing device is accurately controlled, and the processing efficiency and quality of meat products are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and more specifically, to an intelligent method for adjusting the size of meat blocks based on image recognition and a meat grinding device. Background Art

[0002] In the modern food processing industry, the meat grinding process is a key step in the production of many meat products; with the continuous improvement of market requirements for the quality and production efficiency of meat products, traditional meat grinding methods face many challenges;

[0003] In the traditional meat grinding process, the treatment of meat products often lacks accuracy: on the one hand, it is difficult to treat meats of different sizes, freshness and categories in a targeted manner; due to the lack of detailed analysis of meat products, meat grinding devices usually operate with fixed parameters, which leads to poor meat grinding effects when facing meats with different characteristics, such as uneven meat block sizes, excessive grinding affecting the texture of the meat, etc., thus affecting product quality and production efficiency;

[0004] On the other hand, manual participation in the meat grinding operation is not only inefficient, but also prone to human errors, making it difficult to ensure the stability and consistency of meat grinding quality. Especially in large-scale production, the limitations of manual operation become more obvious;

[0005] With the development of image recognition technology, it shows great application potential in the food processing field; image recognition technology can obtain the characteristic information of an object, such as shape, size, color, etc. by analyzing the image of the object. However, in the meat grinding process, effectively applying image recognition technology still faces many difficulties; the diversity of meat products makes the image features complex and variable, with significant differences in color, texture, etc. among different meats, and the image quality collected under different lighting conditions varies greatly, all of these factors increase the difficulty of accurately identifying the characteristics of meat products; in addition, how to effectively combine the information obtained by image recognition with the parameter adjustment of the meat grinding device to achieve precise control of the meat grinding process is also an urgent problem to be solved;

[0006] In view of this, the present invention proposes an intelligent method for adjusting the size of meat blocks based on image recognition and a meat grinding device to solve the above problems. Summary of the Invention

[0007] In order to overcome the above-mentioned defects of the prior art and to achieve the above object, the present invention provides the following technical solutions:

[0008] An intelligent method for adjusting the size of meat blocks based on image recognition, comprising:

[0009] Step 1: Based on the spectral bands of different preset wavelengths, image acquisition is performed on the meat products input into the target meat mincing device to obtain the corresponding spectral image sequence; the spectral image sequence consists of several spectral meat block images;

[0010] Step 2: Calibrate the obtained spectral meat block images, obtain the gray distribution map corresponding to the calibrated spectral meat block images, perform image gray calibration on the calibrated spectral meat block images based on the gray distribution map to obtain the corresponding initial meat block images, and perform adaptive image optimization processing on the initial meat block images to obtain the corresponding optimized meat block images;

[0011] Step 3: Based on the pre-constructed meat block recognition model, perform image recognition on the obtained optimized meat block images to obtain the corresponding meat block recognition information, and set the corresponding meat mincing control strategy based on the meat block recognition information;

[0012] Step 4: Adjust the parameters of the corresponding meat mincing device based on the meat mincing control strategy, and perform information feedback based on the parameter adjustment results.

[0013] Furthermore, the acquisition process of the corresponding spectral image sequence includes:

[0014] Set image acquisition nodes, and the image acquisition nodes include a spectral imaging terminal and an illumination terminal;

[0015] Based on the spectral imaging terminal in the image acquisition node, perform product scanning on the meat products on the conveyor belt at the feeding port of the corresponding meat mincing device in different bands, and obtain the corresponding spectral image sequence based on the scanning results; the spectral image sequence consists of spectral meat images corresponding to different spectral bands at different times;

[0016] Furthermore, the acquisition process of the optimized meat block images includes:

[0017] Obtain the corresponding spectral meat block images based on the collected spectral image sequence; perform image preprocessing on the obtained spectral meat block images to obtain the corresponding initial meat block images;

[0018] After the image preprocessing is completed, perform discrete wavelet decomposition on the obtained initial meat block images to obtain the corresponding low-frequency components and high-frequency components. At the same time, obtain the wavelet coefficients at different scales during the corresponding discrete wavelet decomposition process;

[0019] Perform threshold processing on the obtained wavelet coefficients, and perform image reconstruction based on the wavelet coefficients after threshold processing to obtain the corresponding first reconstructed image;

[0020] Perform logarithmic processing on the obtained low-frequency components to obtain corresponding enhanced low-frequency components; meanwhile, perform optimization processing on the high-frequency components based on a pre-constructed adaptive image optimization function to obtain corresponding enhanced high-frequency components; and perform image reconstruction based on the enhanced low-frequency components and enhanced high-frequency components to obtain corresponding second reconstructed images;

[0021] Furthermore, perform image fusion on the obtained first reconstructed image and the second reconstructed image to obtain a corresponding optimized meat block image;

[0022] Based on the above process of obtaining the optimized meat block image, perform image optimization processing on other spectral meat block images in the corresponding spectral image sequence to obtain a corresponding optimized image sequence;

[0023] Further, the formula for obtaining wavelet coefficients is: In the formula, I(X) represents the initial meat block image; u and v respectively represent the horizontal and vertical coordinates of pixel points in the initial meat block image; respectively represent wavelet functions in the horizontal and vertical directions; a and s respectively represent the scale factor and translation factor in the discrete wavelet decomposition process; M and N respectively represent the length and width of the initial meat block image; xb (b,k) represents the wavelet coefficient at the decomposition level b and translation position k;

[0024] The formula for threshold processing is:

[0025] In the formula, sgn() represents the sign function; T represents the adaptive threshold; xb` (b,k) represents the wavelet coefficient after threshold processing; T represents the adaptive threshold; ω is the Gaussian standard deviation at the corresponding scale, N represents the mean of wavelet coefficients; ∫() represents the integral operation;

[0026] The formula for logarithmic processing is: R`(u,v) = exp(log 10 (S(u,v)) - log 10 (S(u,v) * F(u,v))); R`(u,v) represents the enhanced low-frequency component obtained after logarithmic processing; S(u,v) represents the low-frequency component corresponding to the initial meat block image; F(u,v) represents the Gaussian surround function; τ represents the scale parameter;

[0027] The formula for the adaptive image optimization function is:

[0028]

[0029] R(m); where R`(m) represents the brightness value of the m-th pixel point in the corresponding reconstructed meat block image after image enhancement processing; Ω represents a sliding window of size Q1×Q2 centered on pixel point m; both Q1 and Q2 are constants; n represents the pixel point index within the sliding window; P (m,n) represents the probability of selecting pixel point n within the sliding window centered on pixel point m; f() represents a preset adaptive weight adjustment function based on the brightness ratio;

[0030] sign + (ξ) is a constant; where, where R(m) and R(n) represent the brightness values of pixel point m and pixel point n respectively; sign - (ξ) = 1 - sign + (ξ).

[0031] Furthermore, the process of image preprocessing for the obtained spectral meat block images includes:

[0032] Performing image correction on each spectral meat block image respectively to obtain the corresponding corrected image;

[0033] Respectively obtaining the gray values of each pixel point within the corresponding corrected image and constructing the gray distribution map of the corresponding corrected image based on them; where the formula for obtaining the gray distribution map is: where I(i_0,j_0) represents the gray value at pixel point (i_0,j_0) within the corrected image; j_0 = 1, 2, ……, W, W is a natural number, W represents the length of the corrected image; H represents the height of the corrected image; μi_0 represents the average gray value of the pixel points in the i_0-th column within the corrected image; where the gray distribution map is the change curve of the average gray value of the pixel points in the corresponding columns of the corrected image at different times;

[0034] Obtaining the gray distribution map corresponding to the corresponding corrected image and obtaining the first peak set and the first valley set corresponding to the corrected image within the corresponding time curve based on it; (i.e., the local maximum and local minimum sets);

[0035] Respectively performing set update on the corresponding first peak set and first valley set based on a preset set update rule; and performing difference calculation between the updated first peak set and first valley set and the original first peak set and first valley set respectively to obtain the corresponding initial peak set and initial valley set; where the set update rule includes that if the average gray value of the column pixel points within the corresponding first peak or valley set is outside the preset peak or valley threshold range, then it is excluded, and when the Euclidean distance between the average gray values of adjacent column pixel points is less than the preset distance threshold, the average gray value of the smaller column pixel point is excluded from the corresponding first peak or valley set;

[0036] Furthermore, perform smoothing filtering on the obtained initial peak set and initial valley set to obtain the corresponding mean gray distribution map U = (μ`1, μ`2, ……, μ`i_0);

[0037] Furthermore, perform image gray correction on the original corrected image based on the mean gray distribution map to obtain the corresponding initial meat block image;

[0038] The formula for image gray correction is:

[0039] Y i_0 [j_0] represents the gray value at the pixel point (i_0, j_0) after image gray correction; X i_0 [j_0] represents the gray value at the pixel point (i_0, j_0) before image gray correction, and μ`i_0 represents the gray mean of the i_0 - th column of pixel points after smoothing filtering; σ i_0 represents the mean gray value of the i_0 - th column of pixel points in the original corrected image; σ` i_0 represents the gray mean of the i_0 - th column of pixel points after image gray correction;

[0040] Further, the process of performing image recognition on the obtained optimized meat block image sequence to obtain the corresponding meat block recognition information and setting the corresponding meat grinding control strategy based on the meat block recognition information includes:

[0041] Input the obtained optimized image sequence into a pre - constructed meat block recognition model to obtain the corresponding model data results, and obtain the corresponding meat block recognition information based on it. The meat block recognition information includes information such as meat category, meat contour size, meat freshness, etc.;

[0042] Furthermore, construct the corresponding meat grinding control strategy based on the obtained meat block recognition information. The meat grinding control strategy includes various operating parameters required by the meat grinding device; the operating parameters include blade speed, torque, and other related parameters;

[0043] Furthermore, perform parameter control on the target meat grinding device based on the meat grinding control strategy; among them, the parameter control is a delay control strategy; that is, obtain the time node when the corresponding meat product passes through the corresponding scanning area of the image acquisition node and the current transmission speed of the conveyor belt. Then, based on these, predict the expected time node when the corresponding meat block reaches the meat grinding device, and perform parameter regulation on the operating parameters corresponding to the corresponding meat grinding device based on the expected time node.

[0044] Further, the construction process of the corresponding meat block recognition model includes:

[0045] The backbone network of the meat block recognition model is an improved convolutional neural network, and the basic framework of the improved convolutional neural network is an input layer, a feature layer, and an output layer;

[0046] The output layer is used to receive the input vector and perform scaling processing on the corresponding image to meet the input requirements of the model;

[0047] The feature layer is composed of two feature branches. The feature branches include a spectral feature branch and an image feature branch. The spectral feature branch is used to extract the spectral feature data of the corresponding input vector to obtain the corresponding spectral feature map; the image feature branch is used to perform three-dimensional convolution operations on the input vector to obtain the corresponding convolution feature map;

[0048] The input layer fuses the spectral feature map and the convolution feature map received to obtain the corresponding spectral-image feature map, and maps it to the required meat block recognition information;

[0049] Construct a training data set, and the training data set is composed of several training samples; among them, the construction process of the training data set includes:

[0050] Obtain several historical spectral meat block images with sample labels. The sample labels include meat category, meat size, meat freshness, etc.; respectively extract the spectral features and image features of the corresponding historical spectral meat block images, and then fuse the spectral features and image features to obtain the corresponding spectral-image feature data; construct the mapping relationship between the spectral-image feature data and the corresponding historical spectral meat block images and sample labels, and construct the corresponding training data set based on it;

[0051] Define AdaGrad as the optimizer to continuously optimize the parameters of the meat block recognition model during the training process. Input the corresponding training sample batch into the meat block recognition model, record the value of the corresponding loss function. When the values of the loss function for L consecutive batches no longer decrease or change, save the parameters of the meat block recognition model at this time, that is, complete the training of the meat block recognition model.

[0052] Furthermore, the formula for obtaining the corresponding spectral feature map is:

[0053] where P i represents the spectral band depth; c represents the index of the spectral band; θ _i,j is the bias term in the two-dimensional convolution operation process; is the weight parameter; represents the pixel value at the pixel point (x, y) in the image corresponding to the c-th spectral band in the j-th feature map in the (i - 1)-th feature layer;

[0054] The formula for performing a three-dimensional convolution operation is as follows:

[0055] In the formula, J (i-1) represents the total number of feature maps included in the image feature branch within the feature layer of the (i - 1)-th layer; H i , W i and R i respectively represent the length, width, and height of the convolution kernel within the same feature layer; represents the value corresponding to the three-dimensional convolution operation of the region with the starting three-dimensional spatial coordinates (x, y, z) on the j-th feature map within the image feature branch of the i-th feature layer; represents the weight value of the three-dimensional convolution kernel; θ i,j represents the bias term during the three-dimensional convolution operation; h, w, and r are the indices of the three-dimensional convolution kernel in the three dimensions during the three-dimensional convolution operation; * represents the convolution operation; softmax[] represents the activation function; represents the pixel value at the three-dimensional spatial coordinates [(x + h), (y + w), (z + r)] within the j-th feature map of the (i - 1)-th feature layer;

[0056] The formula for feature fusion is as follows: V fusion represents the fused image, that is, the spectral-image feature map;

[0057] Define the loss function of the meat block recognition model In the formula, d t represents the true sample label of the input training sample; represents the probability that the prediction result is a certain sample label; A t represents the linear combination of the input training sample with the weight parameters and their corresponding features; β is the regularization coefficient; represents the total number of weight coefficients; the weight coefficients refer to the weights between the bias term within the corresponding meat block recognition model and multiple spectral-image features within the corresponding training sample; t is the number of input training samples; N` represents the total number of training samples.

[0058] Furthermore, the process of parameter adjustment for the corresponding meat grinding device based on the meat grinding control strategy and information feedback based on the parameter adjustment results includes:

[0059] Transmit the obtained meat grinding control strategy to the control terminal of the corresponding meat grinding device. Furthermore, the control terminal performs parameter control on the target meat grinding device based on the meat grinding control strategy;

[0060] After the parameter control is completed, based on the corresponding image acquisition node, image acquisition is performed on the meat products at the conveyor belt corresponding to the discharge port of the corresponding meat mincing device, image optimization processing is performed on them, and they are input into the meat block recognition model to obtain the corresponding minced meat block information; based on the minced meat block information, the size of the meat block outline of the meat products at the corresponding discharge port is obtained, and it is matched with the expected outline size. If the size of the meat block outline of the meat products at the discharge port meets the expected outline size, no other operations are performed; if not, the outline deviation between the size of the meat block outline of the meat products at the discharge port and the expected outline size is obtained and fed back into the corresponding control terminal. Then, after receiving the corresponding outline deviation, the control terminal makes a feedback adjustment to the existing meat mincing control strategy based on it.

[0061] A meat mincing device, comprising a conveyor belt, a meat mincer, a corresponding control terminal, and a computer program stored on the control terminal and capable of realizing parameter data regulation.

[0062] The technical effects and advantages of a method for intelligently adjusting the size of meat blocks based on image recognition and a meat mincing device according to the present invention:

[0063] 1. In the image acquisition stage, spectral image sequences are obtained using different wavelength spectral bands, providing rich data for subsequent analysis; after optimization processing, high-quality optimized meat block images can be obtained, improving the recognition accuracy; information such as the meat category, outline size, and freshness obtained in the recognition and analysis link provides a strong basis for setting the meat mincing control strategy, thereby achieving precise control of each link in the meat mincing process.

[0064] 2. By performing image acquisition, optimization processing, recognition, and strategy formulation on the meat products input into the meat mincing device, the operating parameters of the meat mincing device can be accurately regulated according to the actual situation of the meat blocks, ensuring that the meat mincing quality is stable and meets expectations; and according to information such as different meat categories, outline sizes, and freshness, the meat mincing strategy can be adjusted in real time to adapt to various meat processing requirements, effectively solving the problem of unstable meat mincing quality caused by meat block differences in traditional meat mincing methods. Description of the Drawings

[0065] Figure 1 It is a schematic diagram of a method for intelligently adjusting the size of meat blocks based on image recognition according to the present invention;

[0066] Figure 2 It is a schematic diagram of a system for intelligently recognizing and adjusting the size of meat blocks according to the present invention. Detailed Embodiments

[0067] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0068] Embodiment 1

[0069] Please refer to Figure 1 As shown in the figure, an intelligent adjustment method for the size of meat blocks based on image recognition in this embodiment includes:

[0070] Step 1: Perform image acquisition on the meat products input into the target meat mincing device based on the spectral bands of different wavelengths set in advance to obtain corresponding spectral image sequences; the spectral image sequences are composed of several spectral meat block images;

[0071] Step 2: Correct the obtained spectral meat block images, and obtain the gray distribution map corresponding to the corrected spectral meat block images. Based on the gray distribution map, perform image gray correction on the corrected spectral meat block images to obtain corresponding initial meat block images, and perform adaptive image optimization processing on the initial meat block images to obtain corresponding optimized meat block images;

[0072] Step 3: Perform image recognition on the obtained optimized meat block images based on the pre-constructed meat block recognition model to obtain corresponding meat block recognition information, and set corresponding meat mincing control strategies based on the meat block recognition information;

[0073] Step 4: Adjust the parameters of the corresponding meat mincing device based on the meat mincing control strategy, and perform information feedback based on the parameter adjustment results.

[0074] It should be further noted that in the specific implementation process, the process of obtaining the corresponding spectral image sequences includes:

[0075] Set image acquisition nodes, and the image acquisition nodes are deployed at the conveyor belts corresponding to the inlet and outlet of the target meat mincing device; the image acquisition nodes include spectral imaging terminals and lighting terminals; among them, during the deployment process of the image acquisition nodes, it is necessary to consider reducing the shadows and reflections caused by light factors during the acquisition process of the corresponding spectral meat block images to improve the image quality of the acquired images;

[0076] When the meat mincing device is in the working state, based on the image acquisition nodes, scan the meat products on the conveyor belt at the corresponding inlet based on different spectral bands, and obtain corresponding spectral image sequences X(t0) based on the scan results; where X(t0) ∈ R W×H×B, where W and H respectively represent the width and height of the image; B represents the number of bands; X(t0) represents the spectral image sequence at time t0 within the corresponding time interval when the corresponding meat product enters the scanning area of the spectral imaging terminal; wherein, the time interval is the entire time period for scanning the corresponding meat product, and the spectral image sequence is composed of a number of spectral meat images.

[0077] It should be further noted that in the specific implementation process, the process of optimizing the acquisition of meat images includes:

[0078] Obtaining the corresponding spectral meat image X based on the collected spectral image sequence (B,t0) ; where X (B,t0) ∈R W×H ; X (B,t0) represents the spectral meat image at the B-th wavelength at time t0; t0 ∈ QZ; QZ represents the time interval of product scanning corresponding to the corresponding meat product;

[0079] Performing image preprocessing on the obtained spectral meat image X (B,t0) to obtain the corresponding initial meat image; wherein, the process of performing image preprocessing on the obtained spectral meat image includes:

[0080] Performing image correction on each spectral meat image respectively to obtain the corresponding corrected image; Image correction is used to correct the visual deviation caused by the deployment position of the image acquisition node;

[0081] Obtaining the gray value of each pixel point in the corresponding corrected image and constructing the gray distribution map of the corresponding corrected image based on it; wherein, the acquisition formula of the gray distribution map is: In the formula, I(i_0,j_0) represents the gray value at the pixel point (i_0,j_0) in the corrected image; j_0 = 1, 2,..., W, W is a natural number, W represents the length of the corrected image; H represents the height of the corrected image; μi_0 represents the average gray value of the pixel points in the i_0-th column of the corrected image J (B,t0) ; wherein, the gray distribution map is the change curve of the average gray value of the pixel points in the columns of the corresponding corrected image at different times;

[0082] Obtaining the gray distribution map corresponding to the corresponding corrected image and obtaining the first peak set and the first valley set corresponding to the corrected image in the corresponding time curve based on it; (i.e., the set of local maximum and local minimum values);

[0083] Update the corresponding first peak set and first valley set respectively based on a preset set update rule; and perform a difference calculation between the updated first peak set and first valley set and the original first peak set and first valley set respectively to obtain the corresponding initial peak set and initial valley set; wherein, the set update rule includes removing column pixel points whose gray-scale mean value in the corresponding first peak or valley set is outside a preset peak or valley threshold range, and removing the gray-scale mean value of the smaller column pixel point from the corresponding first peak or valley set when the Euclidean distance between the gray-scale mean values of adjacent column pixel points is less than a preset distance threshold;

[0084] Furthermore, perform a smoothing filtering process on the obtained initial peak set and initial valley set to obtain the corresponding mean gray-scale distribution map U = (μ`1, μ`2, ……, μ`i_0);

[0085] Furthermore, perform image gray-scale correction on the original corrected image based on the mean gray-scale distribution map to obtain the corresponding initial meat block image;

[0086] The formula for performing image gray-scale correction is:

[0087] Y i_0 [j_0] represents the gray-scale value at the pixel point (i_0, j_0) after image gray-scale correction; X i_0 [j_0] represents the gray-scale value at the pixel point (i_0, j_0) before image gray-scale correction, and μ`i_0 represents the gray-scale mean value of the i_0th column pixel points after the smoothing filtering process; σ i_0 represents the gray-scale value mean of the i_0th column pixel points in the original corrected image; σ` i_0 represents the gray-scale mean value of the i_0th column pixel points after image gray-scale correction;

[0088] After the image preprocessing is completed, perform a discrete wavelet decomposition on the obtained initial meat block image to obtain the corresponding low-frequency component and high-frequency component, and at the same time, obtain the wavelet coefficients at different scales during the corresponding discrete wavelet decomposition process In the formula, I(X) represents the initial meat block image; u and v respectively represent the horizontal coordinate and vertical coordinate of the pixel points in the initial meat block image; respectively represent the wavelet functions in the horizontal direction and vertical direction; a and s respectively represent the scale factor and translation factor during the discrete wavelet decomposition process; M and N respectively represent the length and width of the initial meat block image; xb (b,k) represents the wavelet coefficient at the decomposition layer b and translation position k;

[0089] Furthermore, threshold processing is performed on the obtained wavelet coefficients, and image reconstruction is carried out based on the wavelet coefficients after threshold processing to obtain a corresponding first reconstructed image. The formula for threshold processing is as follows: In the formula, sgn() represents the sign function; xb` (b,k) represents the wavelet coefficient after threshold processing; T represents the adaptive threshold; ω is the Gaussian standard deviation at the corresponding scale, N represents the mean of the wavelet coefficients; ∫() represents the integral operation;

[0090] Further, logarithmic processing is performed on the obtained low-frequency components to obtain corresponding enhanced low-frequency components; meanwhile, optimization processing is carried out on the high-frequency components based on a pre-constructed adaptive image optimization function to obtain corresponding enhanced high-frequency components; and image reconstruction is carried out based on the enhanced low-frequency components and the enhanced high-frequency components to obtain a corresponding second reconstructed image.

[0091] Furthermore, the obtained first reconstructed image and the second reconstructed image are subjected to image fusion to obtain a corresponding optimized meat block image.

[0092] Based on the above process of obtaining the optimized meat block image, image optimization processing is carried out on other spectral meat block images in the corresponding spectral image sequence to obtain a corresponding optimized image sequence.

[0093] The formula for logarithmic processing is: R`(u,v) = exp(log 10 (S(u,v)) - log 10 (S(u,v) * F(u,v))); R`(u,v) represents the enhanced low-frequency component obtained after logarithmic processing; S(u,v) represents the low-frequency component corresponding to the initial meat block image; F(u,v) represents the Gaussian surround function, which is used to balance the processing process of the low-frequency component; τ represents the scale parameter. If its value is too large, the detailed information of the hyperspectral image is not rich enough; if its value is too small, the color distortion of the hyperspectral image is serious, resulting in insufficient clarity of the hyperspectral image.

[0094] The formula for the adaptive image optimization function is:

[0095]

[0096] In the formula, R`(m) represents the brightness value of the m-th pixel point in the corresponding reconstructed meat block image after image enhancement processing; Ω represents a sliding window of size Q1×Q2 constructed with the pixel point m as the center; Q1 and Q2 are both constants; n represents the pixel point index in the sliding window; P (m,n)It represents the probability of selecting pixel point n within the sliding window centered on pixel point m; f() represents a preset adaptive weight adjustment function based on the brightness ratio, which is used to adjust the contribution weight of pixel point n to the current pixel point m according to the ratio between pixel point m and pixel point n;

[0097] sign + (ξ) is a constant; where, In the formula, ξ = R(n) - R(m); R(m) and R(n) respectively represent the brightness values of pixel point m and pixel point n; sign - (ξ) = 1 - sign + (ξ);

[0098] It should be further noted that in the specific implementation process, the process of performing image recognition on the obtained optimized meat block image sequence to obtain corresponding meat block recognition information and setting corresponding meat grinding control strategies based on the meat block recognition information includes:

[0099] Input the obtained optimized image sequence into a pre-constructed meat block recognition model to obtain corresponding model data results, and obtain corresponding meat block recognition information based on it. The meat block recognition information includes information such as meat category, meat contour size, and meat freshness;

[0100] Furthermore, construct corresponding meat grinding control strategies based on the obtained meat block recognition information. The meat grinding control strategies include various operating parameters required by the meat grinding device; the operating parameters include blade speed, torque, and other related parameters; among them, the construction process of the meat grinding control strategy adopts a stepping motor subdivision control algorithm. The stepping motor subdivision control algorithm can accurately adjust the reamer speed, propulsion speed, and other operating parameters of the corresponding meat grinding device based on the obtained meat block recognition information to ensure the stability and consistency of the meat grinding quality; among them, the stepping motor subdivision control algorithm is a prior art, and the present invention will not elaborate too much.

[0101] It should be further noted that in the specific implementation process, the process of adjusting the parameters of the corresponding meat grinding device based on the meat grinding control strategy and performing information feedback based on the parameter adjustment results includes:

[0102] Transmit the obtained meat grinding control strategy to the control terminal of the corresponding meat grinding device. Furthermore, the control terminal performs parameter control on the target meat grinding device based on the meat grinding control strategy; among them, the parameter control is a delay control strategy; that is, obtain the time node when the corresponding meat product passes through the scanning area corresponding to the image acquisition node and the current conveyor belt transmission speed. Furthermore, based on this, predict the expected time node when the corresponding meat block reaches the meat grinding device, and perform parameter regulation on the operating parameters corresponding to the corresponding meat grinding device based on the expected time node;

[0103] After the parameter control is completed, based on the corresponding image acquisition node, image acquisition is performed on the meat products at the conveyor belt corresponding to the discharge port of the corresponding meat grinding device, image optimization processing is performed on them, and they are input into the meat block recognition model to obtain the corresponding meat grinding meat block information; based on the meat grinding meat block information, the meat block contour size of the meat products at the discharge port is obtained, and it is matched with the expected contour size. If the meat block contour size of the meat products at the discharge port meets the expected contour size, no other operations are performed; if not, the contour deviation between the meat block contour size of the meat products at the discharge port and the expected contour size is obtained and fed back into the corresponding control terminal. Then, after receiving the corresponding contour deviation, the control terminal makes a feedback adjustment to the existing meat grinding control strategy based on it.

[0104] It should be further noted that in the specific implementation process, the construction process of the corresponding meat block recognition model includes:

[0105] The backbone network of the meat block recognition model is an improved convolutional neural network, and the basic framework of the improved convolutional neural network is an input layer, a feature layer, and an output layer;

[0106] The output layer is used to receive the input vector and perform stretching processing on the corresponding image to make it meet the input requirements of the model;

[0107] The feature layer is composed of two feature branches. The feature branches include a spectral feature branch and an image feature branch. The spectral feature branch is used to extract the spectral feature data of the corresponding input vector to obtain the corresponding spectral feature map; the image feature branch is used to perform three-dimensional convolution operations on the input vector to obtain the corresponding convolution feature map;

[0108] The input layer fuses the received spectral feature map and convolution feature map to obtain the corresponding spectral-image feature map and maps it to the required meat block recognition information;

[0109] The formula for obtaining the corresponding spectral feature map is:

[0110] In the formula, P i represents the spectral band depth (i.e., the total number of bands); c represents the index of the spectral band; θ _i,j is the bias term in the two-dimensional convolution process; is the weight parameter; represents the spectral feature map after two-dimensional convolution processing of the image corresponding to the c-th spectral band; represents the pixel value at the pixel point (x, y) in the image corresponding to the c-th spectral band in the j-th feature map in the i-1 feature layer;

[0111] The formula for performing three-dimensional convolution operations is:

[0112] In the formula, J (i-1) represents the total number of feature maps included in the image feature branch within the (i - 1)-th layer of the feature layer; H i , W i and R i respectively represent the length, width, and height of the convolutional kernel within the same feature layer; represents the value corresponding to the region three-dimensional convolution operation on the j-th feature map in the image feature branch within the i-th layer of the feature layer with the starting three-dimensional spatial coordinates of (x, y, z); represents the weight value of the convolutional kernel; θ i,j represents the bias term during the three-dimensional convolution operation process; h, w, and r are the indices in the three dimensions during the three-dimensional convolution operation process; the three dimensions respectively refer to the three-dimensional directions of the length, width, and height of the corresponding convolutional kernel; * represents the convolution operation; softmax[] represents the activation function; represents the pixel value at the three-dimensional spatial coordinates of [(x + h), (y + w), (z + r)] within the j-th feature map in the (i - 1)-th feature layer;

[0113] Among them, the optimized image sequence can be represented as a three-dimensional stereoscopic image;

[0114] The formula for feature fusion is: V fusion represents the fused image, that is, the spectral-image feature map;

[0115] Define the loss function of the meat block recognition model In the formula, d t represents the true sample label of the input training sample; represents the probability that the prediction result is a certain sample label; A t represents the linear combination between the input training sample and the weight parameters and their corresponding features; β is the regularization coefficient; represents the total number of weight coefficients; the weight coefficient refers to the weight between the bias term within the corresponding meat block recognition model and multiple spectral-image features within the corresponding training sample; t is the number of input training samples; N` represents the total number of training samples.

[0116] Construct a training data set, and the training data set consists of several training samples; among them, the construction process of the training data set includes:

[0117] Obtain a number of historical spectral meat block images with sample labels, where the sample labels include meat category, meat size, meat freshness, etc.; respectively perform feature extraction on the corresponding historical spectral meat block images to obtain corresponding spectral features and image features. Furthermore, fuse the spectral features and image features to obtain corresponding spectral-image feature data; construct a mapping relationship between the spectral-image feature data and the corresponding historical spectral meat block images and sample labels, and construct a corresponding training data set based on it;

[0118] Define AdaGrad as the optimizer to continuously optimize the parameters of the meat block recognition model during training. Input the corresponding training sample batches into the meat block recognition model, record the values of the corresponding loss function. When the values of the loss function for L consecutive batches no longer decrease or change, save the parameters of the meat block recognition model at this time, that is, complete the training of the meat block recognition model.

[0119] Through steps such as image acquisition, optimization processing, recognition analysis, and adjustment of the parameters of the meat grinding device, the present invention realizes precise control of the meat grinding process of meat products. This method can automatically adapt to the size and freshness of different meat products, improve the efficiency and product quality of meat grinding, and has the characteristics of intelligence and automation.

[0120] Embodiment 2

[0121] Please refer to Figure 2 As shown, for the parts not described in detail in this embodiment, refer to the description content of Embodiment 1. Provide an intelligent adjustment system for meat block size based on image recognition, including:

[0122] A data acquisition module that performs image acquisition on the meat products input into the target meat grinding device based on the spectral bands of different wavelengths set in advance to obtain a corresponding spectral image sequence; the spectral image sequence consists of a number of spectral meat block images;

[0123] A data processing module for correcting the obtained spectral meat block images, obtaining the gray distribution map corresponding to the corrected spectral meat block images, performing image gray correction on the corrected spectral meat block images based on the gray distribution map to obtain corresponding initial meat block images, and performing adaptive image optimization processing on the initial meat block images to obtain corresponding optimized meat block images;

[0124] A strategy construction module that performs image recognition on the obtained optimized meat block images based on the pre-constructed meat block recognition model to obtain corresponding meat block recognition information, and sets corresponding meat grinding control strategies based on the meat block recognition information;

[0125] A strategy feedback module that adjusts the parameters of the corresponding meat grinding device based on the meat grinding control strategy and performs information feedback based on the parameter adjustment results.

[0126] Each module is connected by wired and / or wireless means to achieve data transmission between modules.

[0127] Embodiment 3

[0128] This embodiment publicly provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the operation mode of the above-provided intelligent adjustment method for meat block size based on image recognition.

[0129] Since the electronic device introduced in this embodiment is the electronic device used to implement an intelligent adjustment method for meat block size based on image recognition in the embodiments of the present application, based on the intelligent adjustment method for meat block size based on image recognition introduced in the embodiments of the present application, those skilled in the art can understand the specific implementation manners and various variations of the electronic device in this embodiment. Therefore, the specific implementation of how this electronic device implements the method in the embodiments of the present application will not be described in detail here. As long as those skilled in the art implement the electronic device used in the intelligent adjustment method for meat block size based on image recognition in the embodiments of the present application, it belongs to the scope protected by the present application.

[0130] The above formulas are all dimensionless and take their numerical values for calculation. The formula is obtained by collecting a large amount of data for software simulation to get a formula closest to the real situation. The preset parameters and threshold selection in the formula are set by those skilled in the art according to the actual situation.

[0131] The above is only the preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technical users in the technical field, several improvements and refinements made without departing from the principle of the present invention should also be regarded as within the protection scope of the present invention.

Claims

1. An intelligent adjustment method for the size of meat chunks based on image recognition, characterized in that, Including: Step 1: Based on the spectral bands of different wavelengths set in advance, image acquisition is performed on the meat products input into the target meat mincing device to obtain a corresponding spectral image sequence; the spectral image sequence consists of several spectral meat block images; Step 2: Correct the obtained spectral meat block images, and obtain the gray distribution map corresponding to the corrected spectral meat block images. Based on the gray distribution map, perform image gray correction on the corrected spectral meat block images to obtain corresponding initial meat block images, and perform adaptive image optimization processing on the initial meat block images to obtain corresponding optimized meat block images; Step 3: Based on the pre-constructed meat block recognition model, perform image recognition on the obtained optimized meat block images to obtain corresponding meat block recognition information, and set corresponding meat mincing control strategies based on the meat block recognition information; Step 4: Based on the meat mincing control strategy, adjust the parameters of the corresponding meat mincing device, and perform information feedback based on the parameter adjustment results.

2. The intelligent adjustment method for the size of meat chunks based on image recognition according to claim 1, characterized in that, The acquisition process of the corresponding spectral image sequence includes: Set image acquisition nodes, and the image acquisition nodes include spectral imaging terminals and lighting terminals; Based on the spectral imaging terminal in the image acquisition node, scan the meat products on the conveyor belt at the feeding port of the corresponding meat mincing device in different spectral bands, and obtain a corresponding spectral image sequence based on the scanning results; the spectral image sequence consists of spectral meat block images corresponding to the scanning results in different spectral bands at different times.

3. The intelligent adjustment method for meat block size based on image recognition according to claim 2, wherein The acquisition process of the optimized meat block images includes: Based on the acquired spectral image sequence, obtain corresponding spectral meat block images; perform image preprocessing on the obtained spectral meat block images to obtain corresponding initial meat block images; After the image preprocessing is completed, perform discrete wavelet decomposition on the obtained initial meat block images to obtain corresponding low-frequency components and high-frequency components. At the same time, obtain the wavelet coefficients at different scales during the corresponding discrete wavelet decomposition process; Perform threshold processing on the obtained wavelet coefficients, and perform image reconstruction based on the wavelet coefficients after threshold processing to obtain a corresponding first reconstructed image; Perform logarithmic processing on the obtained low-frequency components to obtain corresponding enhanced low-frequency components; at the same time, perform optimization processing on the high-frequency components based on the pre-constructed adaptive image optimization function to obtain corresponding enhanced high-frequency components; and perform image reconstruction based on the enhanced low-frequency components and enhanced high-frequency components to obtain a corresponding second reconstructed image; Fuse the obtained first reconstructed image and the second reconstructed image to obtain a corresponding optimized meat block image; Based on the above acquisition process of the optimized meat block images, perform image optimization processing on other spectral meat block images in the corresponding spectral image sequence to obtain a corresponding optimized image sequence.

4. The intelligent adjustment method for meat block size based on image recognition according to claim 3, characterized in that, The acquisition formula for wavelet coefficients is as follows: In the formula, I(X) represents the initial meat block image; u and v respectively represent the horizontal coordinate and vertical coordinate of the pixel points in the initial meat block image; respectively represent the wavelet functions in the horizontal direction and vertical direction; a and s respectively represent the scale factor and translation factor in the discrete wavelet decomposition process; M and N respectively represent the length and width of the initial meat block image; xb (b,k) represents the wavelet coefficient at the decomposition layer b and the translation position k; The formula for performing threshold processing is: where sgn() represents the sign function; xb` (b,k) represents the wavelet coefficient after threshold processing; T represents the adaptive threshold; ω is the Gaussian standard deviation at the corresponding scale, N represents the mean value of the wavelet coefficients; ∫() represents the integral operation; The formula for logarithmic processing is: R`(u, v) = exp(log 10 (S(u, v)) - log 10 (S(u, v) * F(u, v))); R`(u, v) represents the enhanced low-frequency component obtained after logarithmic processing; S(u, v) represents the low-frequency component corresponding to the initial meat block image; F(u, v) represents the Gaussian surround function; τ represents the scale parameter; The formula for the adaptive image optimization function is: R(m); where R`(m) represents the brightness value of the m-th pixel point in the corresponding reconstructed meat block image after image enhancement processing; Ω represents a sliding window of size Q1×Q2 centered on the pixel point m; both Q1 and Q2 are constants; n represents the pixel point index within the sliding window; P (m,n) represents the probability of selecting the pixel point n within the sliding window centered on the pixel point m; f() represents a preset adaptive weight adjustment function based on the brightness ratio; sign + (ξ) is a constant; where, In the formula, ξ = R(m) - R(n); R(m) and R(n) respectively represent the luminance values of pixel point m and pixel point n; sign - (ξ) = 1 - sign + (ξ).

5. The intelligent adjustment method for meat block size based on image recognition according to claim 3, wherein The process of performing image preprocessing on the obtained spectral meat block images includes: Perform image correction on each spectral meat block image respectively to obtain corresponding corrected images; Obtain the gray values of each pixel point in the corresponding corrected image respectively, and construct a gray distribution map of the corresponding corrected image based on them; among them, the acquisition formula of the gray distribution map is: In the formula, I(i_0,j_0) represents the gray value at the pixel point (i_0,j_0) in the corrected image; j_0 = 1, 2, ……, W, W is a natural number, W represents the length of the corrected image; H represents the height of the corrected image; μi_0 represents the average gray value of the pixel points in the i_0-th column in the corrected image; among them, the gray distribution map is the change curve of the average gray value of the pixel points in the corresponding corrected image at different times in the column; Obtain the gray distribution map corresponding to the corresponding corrected image, and based on it, obtain the first peak set and the first valley set corresponding to the corrected image in the corresponding time curve; Update the corresponding first peak set and first trough set respectively based on the preset set update rules; and perform difference calculations on the updated first peak set and first trough set with the original first peak set and first trough set respectively to obtain the corresponding initial peak set and initial trough set; Perform smoothing filtering on the obtained initial peak set and initial trough set to obtain the corresponding mean gray distribution map U = (μ`1, μ`2, ……, μ`i_0); Perform image gray correction on the original corrected image based on the mean gray distribution map to obtain the corresponding initial meat block image; The formula for performing image gray correction is: Y i_0 [j_0] represents the gray value at the pixel point (i_0, j_0) after image gray correction; X i_0 [j_0] represents the gray value at the pixel point (i_0, j_0) before image gray correction, and μ`i_0 represents the average gray value of the pixel points in the i_0-th column after smoothing filtering; σ i_0 represents the average gray value of the pixel points in the i_0-th column in the original corrected image; σ` i_0 represents the average gray value of the pixel points in the i_0-th column after image gray correction.

6. The intelligent adjustment method for meat block size based on image recognition according to claim 3, characterized in that The process of performing image recognition on the obtained optimized meat block image sequence to obtain the corresponding meat block recognition information and setting the corresponding meat grinding control strategy based on the meat block recognition information includes: Input the obtained optimized image sequence into the pre-constructed meat block recognition model to obtain the corresponding model data result, and obtain the corresponding meat block recognition information based on it; Construct the corresponding meat grinding control strategy based on the obtained meat block recognition information, and the meat grinding control strategy includes various operating parameters required by the meat grinding device.

7. The intelligent adjustment method for meat block size based on image recognition according to claim 6, characterized in that, The construction process of the corresponding meat block recognition model includes: The backbone network of the meat block recognition model is an improved convolutional neural network, and the basic framework of the improved convolutional neural network is an input layer, a feature layer, and an output layer; The output layer is used to receive the input vector and perform scaling processing on the corresponding image to make it meet the input requirements of the model; The feature layer is composed of two feature branches, and the feature branches include a spectral feature branch and an image feature branch. The spectral feature branch is used to extract the spectral feature data of the corresponding input vector to obtain the corresponding spectral feature map; the image feature branch is used to perform three-dimensional convolution operations on the input vector to obtain the corresponding convolution feature map; The input layer fuses the spectral feature map and the convolution feature map received to obtain the corresponding spectral-image feature map, and maps it to the required meat block recognition information; Construct a training data set, and the training data set is composed of several training samples; Define AdaGrad as the optimizer to continuously optimize the parameters of the meat block recognition model during training. Input the corresponding training sample batches into the meat block recognition model, record the values of the corresponding loss function. When the values of the loss function for L consecutive batches no longer decrease or change, save the parameters of the meat block recognition model at this time, that is, complete the training of the meat block recognition model; L is a constant.

8. The intelligent adjustment method for meat block size based on image recognition according to claim 7, characterized in that The formula for obtaining the corresponding spectral feature map is: In the formula, P i represents the spectral band depth; c represents the index of the spectral band; θ_ i,j is the bias term in the two-dimensional convolution operation process; is the weight parameter; represents the corresponding spectral feature map; represents the pixel value at the pixel point (x, y) in the image corresponding to the c-th spectral band in the j-th feature map in the (i - 1)-th feature layer; The formula for performing three-dimensional convolution operations is: Formula Among them, J (i-1) represents the total number of feature maps included in the image feature branch within the (i - 1)-th layer feature layer; H i , W i and R i respectively represent the length, width, and height of the convolutional kernel within the same feature layer; represents the value corresponding to the region three-dimensional convolution operation on the j-th feature map within the image feature branch of the i-th layer feature layer with the starting three-dimensional spatial coordinates of (x, y, z); represents the weight value of the three-dimensional convolutional kernel; θ i,j represents the bias term in the three-dimensional convolution operation process; h, w, and r are the indices of the three-dimensional convolution kernel in the three dimensions during the three-dimensional convolution operation process; * represents the convolution operation; softmax[] represents the activation function; represents the pixel value at the three-dimensional spatial coordinates [(x + h), (y + w), (z + r)] within the j-th feature map in the (i - 1)-th feature layer; The formula for feature fusion is as follows: V fusion represents the fused image; Define the loss function of the meat block recognition model In the formula, d t represents the true sample label of the input training sample; represents the probability that the prediction result is a certain sample label; A t represents the linear combination of the input training sample, the weight parameter, and its corresponding features; β is the regularization coefficient; represents the total number of weight coefficients; the weight coefficient refers to the weight between the bias term in the corresponding meat block recognition model and multiple spectral-image features in the corresponding training sample; t is the number of input training samples; N` represents the total number of training samples.

9. The intelligent adjustment method for meat block size based on image recognition according to claim 6, wherein The process of adjusting the parameters of the corresponding meat grinding device based on the meat grinding control strategy and performing information feedback based on the parameter adjustment results includes: Transmit the obtained meat grinding control strategy to the control terminal of the corresponding meat grinding device, and the control terminal controls the parameters of the target meat grinding device based on the meat grinding control strategy; After the parameter control is completed, based on the corresponding image acquisition node, image acquisition is performed on the meat products at the conveyor belt corresponding to the discharge port of the corresponding meat grinding device, image optimization processing is performed on it, and it is input into the meat block recognition model to obtain the corresponding meat block information of the ground meat, and based on this, the existing meat grinding control strategy is feedback-adjusted.

10. A meat mincing device, comprising a conveyor belt, a meat mincing machine, a corresponding control terminal, and a computer program stored on the control terminal and capable of realizing parameter data regulation, characterized in that, When the control terminal executes the computer program, it implements the intelligent adjustment method for the size of meat blocks based on image recognition described in claims 1 to 9.

Images