Image recognition-based quality evaluation method for liquor pit mud

Abstract

The present application provides a kind of liquor pit mud quality evaluation method based on image recognition, it is related to the technical field of brewing liquor, first by obtaining pit mud surface image, and combining physical and chemical indexes and expert experience are marked with quality grade, the pit mud image obtained is preprocessed, obtain the pit mud image after preprocessing, construct high-quality training data set, by training data set the image after preprocessing is input into improved YOLO11 model and is trained and tested, obtain pit mud quality evaluation model, obtain the pit mud image of pit mud to be evaluated, after preprocessing, pit mud quality class and confidence of pit mud to be evaluated are obtained using pit mud quality evaluation model, the existing problem of low efficiency of liquor pit mud quality evaluation detection is solved, the present application is applicable to pit mud quality evaluation.

Description

Technical Field

[0001] This invention relates to the field of brewing technology, and in particular to a method for evaluating the quality of baijiu cellar mud based on image recognition. Background Technology

[0002] In recent years, the rapid development of artificial intelligence technology has driven the intelligent upgrading of traditional industries. The widespread application of computer vision and image recognition technologies in fields such as industrial inspection and agricultural monitoring has provided new means for the automation and quality control of production processes. As a traditional Chinese fermented liquor, the quality of baijiu is closely related to the condition of the fermentation pits. The pit mud is a key factor affecting the fermentation microbial community and the generation of aroma substances, and its quality directly determines the flavor and stability of baijiu.

[0003] Currently, the assessment of the quality of baijiu cellar mud relies heavily on manual sensory judgment or physicochemical testing. Manual judgment is highly subjective and inconsistent in standards; physicochemical testing is costly, time-consuming, and inefficient. Summary of the Invention

[0004] The technical problem solved by this invention: This invention provides a method for quality assessment of baijiu cellar mud based on image recognition, which solves the problem of low detection efficiency in existing baijiu cellar mud quality assessment methods.

[0005] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a method for evaluating the quality of baijiu cellar mud based on image recognition, comprising the following steps:

[0006] S1. Obtain images of the pit mud and label the quality grade of the pit mud;

[0007] S2. Preprocess the acquired pit mud images to obtain preprocessed pit mud images and obtain the training dataset.

[0008] S3. The preprocessed images are input into the improved YOLO11 model for training and testing using the training dataset to obtain the pit mud quality assessment model. The improved YOLO11 model includes adding a confidence branch to the detection head and adopts a multi-task joint optimization loss function, which is: ,in, Indicates the overall loss. Indicates location loss. Represents classification loss. Indicates confidence loss. Indicates the positioning loss weights. Represents the classification loss weights. Indicates the confidence loss weight. This indicates that the confidence level dynamically adjusts the weights, and the weights are automatically updated based on the confidence level.

[0009] S4. Obtain images of the pit mud to be evaluated, perform preprocessing, and then use the pit mud quality assessment model to obtain the pit mud quality category and confidence level.

[0010] Furthermore, in S1, acquiring the cellar mud image includes taking pictures of a designated area on the surface of the cellar mud in the liquor fermentation tank using a camera at a preset shooting angle and shooting distance to obtain the cellar mud image. The cellar mud quality grade labeling includes labeling the cellar mud quality grade on the image based on physicochemical indicators and expert experience.

[0011] Furthermore, in S2, the preprocessing includes denoising, contrast enhancement, image size normalization, and multi-scale texture enhancement. The denoising employs bilateral filtering to preserve the texture features of the pit mud image. The contrast enhancement includes using an adaptive histogram equalization method to self-adjust the brightness and contrast of pit mud images acquired under different lighting conditions, enhancing the visibility of surface details and highlighting the texture features of the pit mud. The image size normalization includes dynamically cropping and adjusting the pit mud image to a standard size based on texture direction and local gradient distribution features. The standard size is 640×640 or 416×416. The multi-scale texture enhancement includes performing multi-scale Gaussian pyramid decomposition on the pit mud image, extracting pit mud texture information at different spatial frequencies, and weighted fusion of features at each scale to enhance the high-frequency texture features of micro-cracks and pores on the pit mud surface.

[0012] Furthermore, S2 also includes one or more of the following composite data augmentation operations on the mud images in the training dataset: random rotation, horizontal flipping, vertical flipping, brightness perturbation, contrast perturbation, random cropping and mixing, and color space jittering.

[0013] Furthermore, the method also includes generating a visualized quality heat map based on attention weights to show the distribution trends of the areas and features that the pit mud quality assessment model focuses on.

[0014] Furthermore, the method also includes extracting the surface feature analysis results of the pit mud, which include color uniformity, texture roughness, degree of mottling, and crack density.

[0015] Furthermore, the method also includes comparing the quality grades of the same batch of pit mud to obtain high-quality pit mud of that batch, and analyzing the fermentation environment that improves the quality of that batch of pit mud.

[0016] The beneficial effects of this invention are as follows: This invention provides a method for evaluating the quality of baijiu (Chinese liquor) cellar mud based on image recognition. First, images of the cellar mud surface are acquired through a standardized image acquisition process, and quality grades are labeled using physicochemical indicators and expert experience, constructing a high-quality training dataset. In the preprocessing stage, bilateral filtering for noise reduction, adaptive histogram equalization to enhance contrast, dynamic cropping and size standardization, and multi-scale texture enhancement are employed to effectively preserve and highlight key texture features of the cellar mud surface, laying a solid data foundation for model training. The improved YOLO11 model, by adding a confidence branch and designing a loss function for multi-task joint optimization, not only improves the classification accuracy of cellar mud quality categories but also outputs the confidence level of the evaluation results, enhancing the model's reliability. Furthermore, by generating a quality heat map, the key areas of interest to the model are visually displayed, providing a visual basis for cellar mud quality analysis; extracted surface feature parameters such as color uniformity and texture roughness further enrich the evaluation dimensions. Comparative analysis of the quality of cellar mud from the same batch can screen out high-quality cellar mud samples and infer key parameters for optimizing the fermentation environment, providing data support for cellar mud management and quality improvement in the baijiu production process. Compared with traditional manual sensory judgment and physicochemical testing, this method has advantages such as high detection efficiency, strong objectivity and low cost. It helps to promote the intelligent upgrading of the liquor brewing industry, realize the precise control of cellar mud quality, and solve the problem of low detection efficiency in the existing liquor cellar mud quality assessment. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a method for evaluating the quality of baijiu cellar mud based on image recognition, provided by the present invention. Detailed Implementation

[0018] This invention addresses the problem of low efficiency in existing methods for assessing and detecting the quality of baijiu (Chinese liquor) cellar mud by providing an image recognition-based method for assessing the quality of baijiu cellar mud. Figure 1 As shown, it includes the following steps:

[0019] S1. Obtain images of the pit mud and label the quality grade of the pit mud.

[0020] Specifically, acquiring images of the fermentation pit mud involves photographing a designated area on the surface of the mud in the baijiu fermentation tank using a camera at a preset shooting angle and distance. The mud quality grade labeling includes assigning a quality grade to the image based on physicochemical indicators and expert experience. An industrial camera or a high-resolution camera can be used. The preset shooting angle is defined as follows: the angle between the camera's optical axis and the normal to the mud surface is between 0 and 45 degrees; the shooting distance is set between 30 and 100 centimeters; and the designated area is a complete and clear section of the mud surface. The mud quality grade labeling covers four quality levels: excellent, good, medium, and poor. Excellent indicates the surface is covered with white crystals and has excellent sensory characteristics; good indicates the surface has a small amount of crystals and a uniform texture; medium indicates no obvious crystals and a generally average texture; and poor indicates hardening, mold, or degradation.

[0021] S2. Preprocess the acquired pit mud images to obtain preprocessed pit mud images and obtain the training dataset.

[0022] Specifically, the preprocessing includes denoising, contrast enhancement, image size normalization, and multi-scale texture enhancement. The denoising employs bilateral filtering to preserve the texture features of the pit mud image. The contrast enhancement involves using an adaptive histogram equalization method to self-adjust the brightness and contrast of pit mud images acquired under different lighting conditions, enhancing the visibility of surface details and highlighting the texture features of the pit mud. The image size normalization involves dynamically cropping and adjusting the pit mud image to a standard size based on texture direction and local gradient distribution features; the standard size is 640×640 or 416×416. The multi-scale texture enhancement involves performing multi-scale Gaussian pyramid decomposition on the pit mud image, extracting pit mud texture information at different spatial frequencies, and weighted fusion of features at each scale to enhance the high-frequency texture features of micro-cracks and pores on the pit mud surface.

[0023] Specifically, one or more of the following composite data augmentation operations can be applied to the mud images in the training dataset: random rotation, horizontal flipping, vertical flipping, brightness and contrast perturbation, random cropping and mixing, and color space jittering, thereby increasing the amount of data, improving data diversity, and enhancing generalization ability.

[0024] The random rotation angle range is set from -15° to 15° to simulate slight angular deviations that may occur in the actual acquisition of pit mud images. Horizontal and vertical flipping operations expand the diversity of sample orientations through mirror transformation. Brightness and contrast perturbations randomly adjust the brightness component by ±20% and dynamically change the contrast component by ±15% in the HSV color space of the original image to adapt to image features under different lighting conditions. Random cropping and blending involves randomly selecting 20% ​​to 50% of the image for cropping and then swapping and blending to enhance the diversity of local texture combinations. Color space dithering simulates color deviations caused by different shooting devices or ambient light by adding random pixel value offsets of -10 to 10 in the RGB color channels, further enriching the visual variation dimensions of the training samples. The synergistic effect of these composite enhancement strategies effectively avoids overfitting during model training, ensuring more stable recognition performance when facing real pit mud images.

[0025] S3. The preprocessed images are input into the improved YOLO11 model for training and testing using the training dataset to obtain the pit mud quality assessment model. The improved YOLO11 model includes adding a confidence branch to the detection head and adopts a multi-task joint optimization loss function, which is: ,in, Indicates the overall loss. Indicates location loss. Represents classification loss. Indicates confidence loss. Indicates the positioning loss weights. Represents the classification loss weights. Indicates the confidence loss weight. This indicates that the confidence level dynamically adjusts the weights, and the weights are automatically updated based on the confidence level.

[0026] The classification loss is obtained using the Focal Loss function to address the imbalance in the number of samples at different quality levels. The localization loss uses the Complete IoU (CIoU) loss function to optimize the matching accuracy between the detection box and the real defect region. An adaptive learning rate scheduling mechanism (Cosine Annealing or OneCycle strategy) is introduced during training to prevent overfitting and improve the robustness and stability of the model under different lighting and texture conditions. The confidence branch outputs the confidence score, which reflects the reliability of the model's judgment on the quality level of the kiln mud, providing a reference for subsequent automated quality assessment and manual review.

[0027] In particular, the confidence scores of the model output are calibrated by temperature scaling to ensure that the confidence score distributions between different quality levels have better interpretability and consistency. The calibrated confidence scores can be directly used for quality level decisions and statistical analysis.

[0028] S4. Obtain images of the pit mud to be evaluated, perform preprocessing, and then use the pit mud quality assessment model to obtain the pit mud quality category and confidence level.

[0029] Through the above steps, the quality score and confidence level of the pit mud to be evaluated can be obtained. To achieve a more comprehensive evaluation of the pit mud quality, the following work can also be carried out: generating a visualized quality heat map based on attention weights to present the distribution of features and areas of interest in the pit mud quality evaluation model; extracting the surface feature analysis results of the pit mud, specifically covering color uniformity, texture roughness, degree of blooming, and crack density; conducting a comparative analysis of the quality grades of pit mud from the same batch, screening out the high-quality pit mud of the batch, and analyzing the fermentation environment that helps improve the quality of the batch of pit mud.

[0030] During the generation of the quality heat map, the Grad-CAM (Gradient-weighted Class Activation Mapping) algorithm is used to backpropagate the gradient weights of neurons in each layer of the model to the input image. This results in a heat map that visually displays the areas the model focuses on when judging the quality grade of the pit mud, such as dense white crystal areas in high-quality pit mud and compacted or moldy areas in low-quality pit mud. This helps technicians quickly locate key visual features affecting the quality of the pit mud. The surface feature analysis of the pit mud is achieved through image segmentation and feature extraction algorithms: color uniformity is quantified by calculating the standard deviation of pixel brightness and hue in the CIELAB color space; texture roughness is extracted based on the gray-level co-occurrence matrix (GLCM) to obtain texture parameters such as contrast and entropy; the degree of bacterial bloom is determined by identifying and statistically analyzing the area ratio of white or yellow bacterial clusters on the pit mud surface; and crack density is calculated by using the Canny edge detection algorithm to extract crack outlines and then calculating the total crack length per unit area. These feature parameters not only serve as auxiliary indicators for model evaluation but also provide data support for analyzing the trend of pit mud quality changes. When comparing the quality of pit mud from the same batch, the system automatically calculates the proportion of samples at each grade and constructs a correlation model based on fermentation environment parameters (such as temperature, humidity, pH, and microbial species). It then uses multiple linear regression or random forest algorithms to identify environmental factors that significantly affect pit mud quality. For example, if a batch of high-quality pit mud is found to have a stable fermentation temperature of 28-32℃ and humidity of 65%-70%, this range can be used as a reference standard for subsequent fermentation environment optimization, thereby achieving precise control and continuous improvement of pit mud quality. This enables multi-dimensional and interpretable automated quality assessment of pit mud, providing not only quality grade conclusions but also traceable characteristic evidence and optimization directions, significantly improving the scientific rigor and operability of the assessment.

Claims

1. A method for evaluating the quality of baijiu cellar mud based on image recognition, characterized in that, Includes the following steps: S1. Obtain images of the pit mud and label the quality grade of the pit mud; S2. Preprocess the acquired pit mud images to obtain preprocessed pit mud images and obtain the training dataset. S3. The preprocessed images are input into the improved YOLO11 model for training and testing using the training dataset to obtain the pit mud quality assessment model. The improved YOLO11 model includes adding a confidence branch to the detection head and adopts a multi-task joint optimization loss function, which is: ,in, Indicates the overall loss. Indicates location loss. Represents classification loss. Indicates confidence loss. Indicates the positioning loss weights. Represents the classification loss weights. Indicates the confidence loss weight. This indicates that the confidence level dynamically adjusts the weights, and the weights are automatically updated based on the confidence level. S4. Obtain images of the pit mud to be evaluated, perform preprocessing, and then use the pit mud quality assessment model to obtain the pit mud quality category and confidence level.

2. The method for evaluating the quality of baijiu cellar mud based on image recognition according to claim 1, characterized in that, In S1, obtaining the cellar mud image includes taking pictures of a designated area on the surface of the cellar mud in the liquor fermentation tank using a camera at a preset shooting angle and shooting distance to obtain the cellar mud image. The cellar mud quality grade labeling includes labeling the cellar mud quality grade on the image based on physicochemical indicators and expert experience.

3. The method for evaluating the quality of baijiu cellar mud based on image recognition according to claim 1, characterized in that, In S2, the preprocessing includes denoising, contrast enhancement, image size normalization, and multi-scale texture enhancement. The denoising uses bilateral filtering to preserve the texture features of the pit mud image. The contrast enhancement includes using an adaptive histogram equalization method to self-adjust the brightness and contrast of pit mud images acquired under different lighting conditions, enhancing the visibility of surface details and highlighting the texture features of the pit mud. The image size normalization includes dynamically cropping and adjusting the pit mud image to a standard size based on texture direction and local gradient distribution features. The standard size is 640×640 or 416×416. The multi-scale texture enhancement includes performing multi-scale Gaussian pyramid decomposition on the pit mud image, extracting pit mud texture information at different spatial frequencies, and weighted fusion of features at each scale to enhance the high-frequency texture features of micro-cracks and pores on the pit mud surface.

4. The method for evaluating the quality of baijiu cellar mud based on image recognition according to claim 1, characterized in that, S2 also includes one or more of the following composite data augmentation operations on the pit mud images in the training dataset: random rotation, horizontal flipping, vertical flipping, brightness perturbation, contrast perturbation, random cropping and mixing, and color space jittering.

5. The method for evaluating the quality of baijiu cellar mud based on image recognition according to claim 1, characterized in that, The method also includes generating a visualized quality heat map based on attention weights to show the distribution trends of the areas and features that the pit mud quality assessment model focuses on.

6. The method for evaluating the quality of baijiu cellar mud based on image recognition according to claim 1, characterized in that, The method also includes extracting the surface feature analysis results of the pit mud, which include color uniformity, texture roughness, degree of mottling, and crack density.

7. The method for evaluating the quality of baijiu cellar mud based on image recognition according to claim 1, characterized in that, The method also includes comparing the quality grades of the same batch of pit mud to obtain high-quality pit mud of that batch, and analyzing the fermentation environment that improves the quality of that batch of pit mud.

Images