Baijiu picking method and system based on deep learning

Abstract

The application discloses a liquor picking method based on deep learning, comprising the following steps: S1, model training: inputing the liquor flower image data set into the lightweight YOLOv5 model to learn the relationship between the liquor flower characteristics and different grades of liquor; S2, image acquisition; S3, image input: inputing the liquor flower image into the trained lightweight YOLOv5 model; S4, model detection: the lightweight YOLOv5 model classifies the liquor flower image according to the liquor flower characteristics of the input liquor flower image and outputs the classification result; S5, quantity and quality picking: inputting the classification result of step S4 into the DCS system, and the DCS system controls the valve to segmentally pack the liquor flowing out of the liquor receiving device into the liquor storage tank. The liquor picking method based on deep learning is adopted, the obtained liquor flower image is analyzed by the lightweight YOLOv5, the liquor can be graded, the automatic quantity and quality picking is realized, the cost is low, and the stability is high.

Description

Technical Field

[0001] This invention relates to a target detection technology, and more particularly to a method and system for extracting liquor based on deep learning. Background Technology

[0002] In the distillation process of baijiu (Chinese liquor), most distilleries still use the traditional method of "observing the blooms during distillation." This method involves workers observing the size, shape, and duration of the blooms formed by the droplets on the distillation cloth or container to determine the alcohol content of the flowing liquor. Different base liquors are then stored separately by manually switching valves. This method demands a high level of experience from the workers. Different workers have varying degrees of accuracy in determining the timing of distillation, and even the same person may make different judgments at different times (e.g., fatigue, excitement, illness). Therefore, it is highly subjective, making it difficult to precisely control the quality of the base liquor, which in turn affects the quality of the baijiu and significantly limits its production volume.

[0003] In addition, instruments for measuring alcohol concentration are used for wine extraction, such as installing alcohol concentration detectors and online flavor compound detectors on the wine flow tubes to detect relevant substances and perform quantitative and qualitative wine extraction. However, these instruments are expensive, making large-scale automated wine extraction difficult. Furthermore, the quality of the flowing wine has a significant impact on the installation and long-term stable use of the instruments, and it is difficult to maintain real-time consistency between detection and wine extraction.

[0004] The method of analyzing the chemical composition of wine by combining visible light and near-infrared spectroscopy, and classifying the base wine by grade (i.e., wine extraction) by combining the established classification model, has the following drawbacks: (1) poor universality and generalizability, and the model must be rebuilt once the instrument status and the test sample change. (2) high cost of required instruments and modeling. (3) slow detection speed, making it difficult to achieve real-time consistency between detection and wine extraction.

[0005] Studies have shown that the hops patterns of different grades of liquor exhibit variations and are closely related to alcohol concentration. Therefore, computer vision-based image processing technology can be used to classify hop images, thereby automating the classification of liquors of different grades and achieving quantitative and qualitative liquor extraction. This method is low-cost and highly stable. However, current methods are performed in multiple independent steps, seeking the optimal method for each step. Therefore, the final result cannot guarantee the optimal solution for the entire problem. Furthermore, it requires a high level of technical expertise from employees, presenting challenges for widespread adoption, and most methods do not yet meet the real-time requirements for detection. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based method for extracting spirits from baijiu (Chinese liquor). By using the lightweight YOLOv5 to analyze the acquired foam images, the spirits can be graded, achieving automated extraction based on both quantity and quality. This method is low-cost and highly stable.

[0007] To achieve the above objectives, this invention provides a deep learning-based method for extracting spirits from baijiu (Chinese liquor), comprising the following steps:

[0008] S1, Model Training

[0009] Input a hop image dataset into a lightweight YOLOv5 model to learn the relationship between hop features and different grades of beer;

[0010] S2, Image Acquisition

[0011] The image acquisition device captures images of the foam formed when the wine falls from the still into the receiving device, thus obtaining images of the foam.

[0012] S3, Image Input

[0013] Input the hop image obtained in step S2 into the trained lightweight YOLOv5 model;

[0014] S4, Model Detection

[0015] The lightweight YOLOv5 model classifies hop images based on their hop features and outputs the classification results.

[0016] S5, Quality-based Distillation

[0017] The classification results from step S4 are input into the DCS system, and the DCS system controls the valves to transfer the wine flowing out of the wine receiving device into the storage tank in segments.

[0018] Preferably, the lightweight YOLOv5 model is an improvement on the YOLOv5s model as the initial model, and the improvement methods include:

[0019] Improved anchor boxes: K-means clustered anchor boxes replace the default anchor boxes to improve the model's detection accuracy and stability;

[0020] Improved backbone network: The ShuffleNetV2 network replaces the YOLOv5s model backbone network for feature extraction, which is used for lightweight improvement, reducing network parameters, reducing computational load, and improving the speed and accuracy of the detection network;

[0021] Added attention mechanism: Added CBAM attention mechanism to improve the attention of hop features and detection accuracy.

[0022] Preferably, the input channel of the ShuffleNetV2 network is split into two equal branches by a separation operation. One branch follows an identity mapping, and the other branch is passed through two 1×1 ordinary convolutions (Conv) and one 3×3 depthwise convolution (DWConv). The two branches are then connected (ConCat), and Channel Shuffle is used to enable the information channel between the two branches.

[0023] Preferably, the CBAM attention mechanism includes a channel attention module and a spatial attention module, which perform attention on the channel and space, respectively.

[0024] Preferably, for a feature map, the CBAM attention mechanism generates a channel attention map first and then a two-dimensional spatial attention map in the order of channel first and then spatial. The generation process is as follows:

[0025]

[0026]

[0027] In the formula, F is the input feature map of CBAM, and M is the input feature map of CBAM. c (F) — Feature map generated by the channel attention module, F c — Input feature map for the spatial attention module, M s (F c — Spatial attention module generates feature maps, F s —CBAM output feature map.

[0028] Preferably, step S5 specifically includes the following steps:

[0029] The classification result of step S4 is output to the display device. At the same time, the host computer in the DCS system sends an instruction to the slave computer according to the classification result. The slave computer controls the valve to remove the segments according to the instruction.

[0030] Preferably, the hop features in step S4 include hop texture, hop shape, and hop outline.

[0031] A system for a deep learning-based method for extracting baijiu (Chinese liquor) includes an image acquisition device positioned above the receiving device, a YOLOv5 model connected to the image acquisition device, and a DCS system connected to the YOLOv5 model.

[0032] The wine receiving device has multiple wine outlet branch pipes connected in parallel to each other on the wine outlet pipe. Each wine outlet branch pipe is equipped with a valve, and the valve is connected to the DCS system.

[0033] The DCS system is also connected to a display device.

[0034] Preferably, the image acquisition device is an industrial camera.

[0035] The beneficial effects of this invention are as follows:

[0036] 1. Compared with the traditional "flower-viewing and wine-picking" method, this invention is based on computer vision, which can maintain high-precision wine picking stably, while avoiding the pollution that may be introduced by manual operation during the liquor brewing process, and avoiding damage to the health of wine-picking workers in high temperature and high humidity environments.

[0037] 2. Compared with the segmented wine extraction method using testing instruments, the present invention has lower usage and maintenance costs.

[0038] 3. Compared with existing computer vision-based wine extraction methods, the end-to-end solution of this invention has lower technical requirements for operators and is more conducive to promotion.

[0039] 4. The high-efficiency testing speed further improves the quality of the base liquor and increases the company's related profits.

[0040] 5. Low memory footprint facilitates the deployment of embedded devices, further saving enterprise costs.

[0041] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the system structure of a deep learning-based method for extracting spirits from baijiu (Chinese liquor) according to an embodiment of the present invention.

[0043] Figure 2 The basic unit structure diagram of ShuffleNetV2, a system for a deep learning-based method for extracting spirits from baijiu (Chinese liquor), is shown in the embodiment of the present invention.

[0044] Figure 3 This is a spatial downsampling structure diagram of ShuffleNetV2, a system for a deep learning-based method for extracting spirits from baijiu (Chinese liquor), as an embodiment of the present invention.

[0045] Figure 4 This is a structural diagram of the attention mechanism of a deep learning-based method for extracting spirits from baijiu (Chinese liquor) according to an embodiment of the present invention.

[0046] The components include: 1. Display device; 2. Image acquisition device; 3. Distillation still; 4. Wine receiving device; 5. Wine storage tank; and 6. Valve. Detailed Implementation

[0047] The present invention will be further described below with reference to the accompanying drawings. It should be noted that this embodiment is based on the present technical solution and provides detailed implementation methods and specific operation processes, but the protection scope of the present invention is not limited to this embodiment.

[0048] Figure 1 This is a schematic diagram of the system structure of a deep learning-based method for extracting spirits from baijiu (Chinese liquor) according to an embodiment of the present invention. Figure 2 The basic unit structure diagram of ShuffleNetV2, a system for a deep learning-based method for extracting spirits from baijiu (Chinese liquor), is shown in the embodiment of the present invention. Figure 3 This is a spatial downsampling structure diagram of ShuffleNetV2, a system for a deep learning-based method for extracting spirits from baijiu (Chinese liquor), as an embodiment of the present invention. Figure 4 This is a structural diagram of the attention mechanism of a deep learning-based baijiu (Chinese liquor) distillation method according to an embodiment of the present invention. Figures 1-4 As shown, the present invention includes the following steps:

[0049] S1, Model Training

[0050] Input a hop image dataset into a lightweight YOLOv5 model to learn the relationship between hop features and different grades of beer;

[0051] S2, Image Acquisition

[0052] The image acquisition device 1 captures the wine foam formed by the wine dropper 3 falling into the wine receiving device 4, and obtains an image of the wine foam.

[0053] S3, Image Input

[0054] Input the hop image obtained in step S2 into the trained lightweight YOLOv5 model;

[0055] S4, Model Detection

[0056] The lightweight YOLOv5 model classifies hop images based on their hop features and outputs the classification results.

[0057] Preferably, the hop features in step S4 include hop texture, hop shape, and hop outline.

[0058] S5, Quality-based Distillation

[0059] The classification results of step S4 are input into the DCS system, and the DCS system controls valve 6 to transfer the wine flowing out of the wine receiving device 4 into the wine storage tank 5 in segments.

[0060] Preferably, step S5 specifically includes the following steps:

[0061] The classification result of step S4 is output to the display device 1 for display. At the same time, the host computer in the DCS system sends an instruction to the slave computer according to the classification result. The slave computer controls the valve 6 to remove the segments according to the instruction.

[0062] In this embodiment, the lightweight YOLOv5 model is an improvement on the YOLOv5s model as the initial model. The improvement method includes:

[0063] Improved anchor boxes: K-means clustered anchor boxes replace the default anchor boxes to improve the model's detection accuracy and stability;

[0064] Improved backbone network: The ShuffleNetV2 network replaces the YOLOv5s model backbone network for feature extraction, which is used for lightweight improvement, reducing network parameters, reducing computational load, and improving the speed and accuracy of the detection network;

[0065] Compared with the initial YOLOv5s model, the lightweight YOLOv5 model proposed in this paper reduces memory usage by 92.5%, parameter count by 93.7%, computational cost by 63.4%, detection accuracy by 2.8%, and FPS by 526, which can effectively achieve real-time detection and classification of hops.

[0066] Preferably, the input channel of the ShuffleNetV2 network is split into two equal branches by a separation operation. One branch follows an identity mapping, and the other branch is passed through two 1×1 ordinary convolutions (Conv) and one 3×3 depthwise convolution (DWConv). The two branches are then connected (ConCat), and Channel Shuffle is used to enable the information channel between the two branches.

[0067] Added attention mechanism: Added CBAM attention mechanism to improve the attention of hop features and detection accuracy.

[0068] Preferably, the CBAM attention mechanism includes a channel attention module and a spatial attention module, which perform attention on the channel and space, respectively.

[0069] Preferably, for a feature map, the CBAM attention mechanism generates a channel attention map first and then a two-dimensional spatial attention map in the order of channel first and then spatial. The generation process is as follows:

[0070]

[0071]

[0072] In the formula, F is the input feature map of CBAM, and M is the input feature map of CBAM. c (F) — Feature map generated by the channel attention module, F c — Input feature map for the spatial attention module, M s (F c — Spatial attention module generates feature maps, F s —CBAM output feature map.

[0073] The system for a deep learning-based method for extracting baijiu (Chinese liquor) includes an image acquisition device 1 positioned above the receiving device 4, a YOLOv5 model connected to the image acquisition device 1, and a DCS system connected to the YOLOv5 model.

[0074] The wine receiving device 4 has multiple wine outlet branch pipes connected in parallel on its wine outlet pipe. Each wine outlet branch pipe is equipped with a valve 6, which is connected to the DCS system.

[0075] The DCS system is also connected to the display device 1.

[0076] Preferably, the image acquisition device 1 is an industrial camera.

[0077] Therefore, the present invention adopts the above-mentioned deep learning-based method for extracting liquor, and uses the lightweight YOLOv5 to analyze the acquired foam images to classify the liquor, realize automated quantitative and qualitative liquor extraction, and achieve low cost and high stability.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A deep learning-based Baijiu (liquor) picking method, characterized in that: The method comprises the following steps: S1, model training Input the hop image dataset into the lightweight YOLOv5 model, and learn the relationship between the hop features and different grades of liquor; The lightweight YOLOv5 model is an improvement based on the YOLOv5s model as the initial model, and the improvement method comprises: Improved anchor box: the anchor box clustered by k-mean replaces the default anchor box to improve the detection accuracy and stability of the model; Improved backbone network: the ShuffleNetV2 network replaces the backbone network of the YOLOv5s model for feature extraction, which is used for lightweight improvement, reduces network parameters and calculation amount, and improves the speed and accuracy of the detection network; Increase attention mechanism: increase CBAM attention mechanism to improve hop feature attention and detection accuracy; The input channels of the ShuffleNetV2 network are separated into two equal branches by a separation operation, one branch follows an identity mapping, and the other branch passes through 2 1×1 ordinary convolution Conv and 1 3×3 deep convolution DWConv, and then the two branches are ConCat, and the information channels between the two branches are enabled using Channel Shuffle; S2, image acquisition The hop image is obtained by shooting the hop formed by the wine still falling into the wine receiving device through the image acquisition device; S3, image input The hop image obtained in step S2 is input into the trained lightweight YOLOv5 model; S4, model detection The lightweight YOLOv5 model classifies the hop image according to the hop features of the input hop image and outputs the classification result; S5, quality and quantity wine picking The classification result of step S4 is input into the DCS system, and the DCS system controls the valve to segmentally fill the wine liquid flowing out of the wine receiving device into the wine storage tank.

2. The Baijiu distillation method based on deep learning according to claim 1, characterized in that: The CBAM attention mechanism includes a channel attention module and a spatial attention module, which perform Attention in the channel and space respectively.

3. The Baijiu distillation method based on deep learning according to claim 2, characterized in that: For a feature map, the CBAM attention mechanism first generates a channel attention map and then generates a two-dimensional spatial attention map in the order of channel first and then space, and the generation process is as follows: In the formula, F——CBAM input feature map, M c (F)——Channel attention module generated feature map, F c ——Spatial attention module input feature map, M s (F c )——Spatial attention module generated feature map, F s ——CBAM output feature map.

4. The Baijiu distillation method based on deep learning according to claim 1, characterized in that: Step S5 specifically comprises the following steps: The classification result of step S4 is output to the display device for display, and the upper computer in the DCS system issues an instruction to the lower computer according to the classification result, and the lower computer controls the valve according to the instruction to segmentally pick.

5. The Baijiu distillation method based on deep learning according to claim 1, characterized in that: The hop features in step S4 include hop texture, hop shape and hop contour.

6. A system based on the deep learning-based Baijiu liquor picking method according to any one of claims 1-5, characterized in that: The image acquisition device is arranged above the wine receiving device, the YOLOv5 model is connected with the image acquisition device, and the DCS system is connected with the YOLOv5 model; A plurality of wine outlet branch pipes are connected in parallel on the wine outlet pipe of the wine receiving device, and a valve is arranged on the wine outlet branch pipe, and the valve is connected with the DCS system; The DCS system is also connected with a display device.

7. The system of a deep learning-based Baijiu liquor picking method according to claim 6, characterized in that: The image acquisition device is an industrial camera.

Images