Improved OfficientNetV2-based citrus reiculata Blanco quality grading method

Abstract

The invention discloses a citrus reiculata Blanco quality grading method based on improved OfficientNetV2. The method comprises the following steps: acquiring a citrus reiculata Blanco image and performing data enhancement; preprocessing the Orah image; dividing the preprocessed Or images into a training set, a verification set and a test set, and marking different quality levels; the method comprises the following steps: constructing a convolutional neural network model OfficientNetV2; the convolutional neural network model OfficientNetV2 is trained based on the training set, and a trained convolutional neural network model is obtained; and testing the convolutional neural network model OfficientNetV2 based on the test set, and outputting the accuracy rate of Or orange grading. And obtaining predicted Or orange quality grading information based on the trained convolutional neural network model. The citrus reiculata Blanco is classified according to different quality grades, and the method has the advantages of being low in cost, good in adaptability, high in accuracy and the like.

Description

Technical Field

[0001] The present invention relates to a method for grading the quality of agricultural products, and specifically to a method for grading the quality of Wogan oranges based on an improved EfficientNetV2. Background Art

[0002] The quality grading of Wogan oranges is a process of classifying Wogan oranges according to their attributes such as size, shape, color, and maturity. This process is very important for all links in the Wogan orange supply chain (such as farms, sorting centers, supermarkets, and consumers). However, the automation of Wogan orange quality grading still faces many challenges, such as increasing the detection speed to adapt to the industrial production rhythm, reducing costs so that small and medium-sized producers can also adopt it, improving the accuracy and reliability of internal quality detection, and reasonably utilizing the large amount of data collected.

[0003] At present, the Wogan orange sorting system, such as the method for non-destructive detection of fruit defects and fruit grading method based on neural network with the patent publication number CN112697984A, has deficiencies such as high cost, complex operation, and poor applicability, and has high requirements for production equipment, and is difficult to use when computing resources are limited. Summary of the Invention

[0004] In order to overcome the defects and deficiencies existing in the prior art, the present invention provides a method for grading the quality of Wogan oranges based on an improved EfficientNetV2. In the process of classifying Wogan oranges of the present invention, Wogan oranges can be intelligently classified according to different quality levels, and it has the advantages of low cost, good adaptability, and high accuracy.

[0005] In order to achieve the above purpose, the present invention adopts the following technical solutions:

[0006] The present invention provides a method for grading the quality of Wogan oranges based on an improved EfficientNetV2, including:

[0007] Obtain Wogan orange images and perform data augmentation;

[0008] Preprocess the Wogan orange images;

[0009] Divide the preprocessed Wogan orange images into a training set, a validation set, and a test set, and mark different quality levels;

[0010] Construct a convolutional neural network model EfficientNetV2, which sequentially includes: a convolutional layer, an SPPF module, multiple ECA-Fused-MBConv modules, and a fully connected layer, and replace the SENet module of the Fused-MBConv module with an ECANet module to obtain an ECA-Fused-MBConv module;

[0011] Train the convolutional neural network model EfficientNetV2 based on the training set to obtain the trained convolutional neural network model;

[0012] Test the convolutional neural network model EfficientNetV2 based on the test set and output the accuracy rate of ponkan grading;

[0013] Obtain the predicted ponkan quality grading information based on the trained convolutional neural network model.

[0014] As a preferred technical solution, obtain ponkan images and perform data augmentation, specifically including:

[0015] Perform data augmentation operations on ponkan images of different categories, including translation, flipping, rotation, and adding noise.

[0016] As a preferred technical solution, preprocess the ponkan images, specifically including:

[0017] Use the Gaussian filtering method to perform noise reduction processing on the ponkan images. Scan each pixel in the ponkan images with a filtering template. The gray value of each pixel point is obtained by weighted averaging of itself and the pixel values in the neighborhood. The size of the neighborhood and the weights of the pixels in the neighborhood are determined by the filtering template;

[0018] Perform image enhancement on the filtered ponkan images. Based on gamma correction, perform power function transformation on the pixel values of the ponkan images to adjust the brightness and contrast of the ponkan images.

[0019] As a preferred technical solution, the ECANet module replaces two fully connected layers with one-dimensional convolution, and its kernel size is adaptively determined through non-linear mapping of the channel dimension;

[0020] The ECANet module implements channel global average pooling on the input features, and then through one-dimensional convolution, uses the Sigmoi function to generate the weights of each channel, and combines the input features with the channel weights to obtain the final output.

[0021] As a preferred technical solution, the SPPF module includes an input module, a ConvBNSiLU module, multiple max pooling layers, a splicing layer, and an output module;

[0022] The splicing layer splices the results before pooling and after each pooling.

[0023] The present invention also provides a ponkan quality grading system based on the improved EfficientNetV2, including: an image acquisition module, a data augmentation module, an image preprocessing module, a data division module, a model construction module, a model training module, a model testing module, and a ponkan quality grading prediction module;

[0024] The image acquisition module is used to acquire citrus reticulata blanco images;

[0025] The data augmentation module is used to perform data augmentation on the citrus reticulata blanco images;

[0026] The image preprocessing module is used to preprocess the citrus reticulata blanco images;

[0027] The data division module is used to divide the preprocessed citrus reticulata blanco images into a training set, a validation set and a test set, and mark different quality levels;

[0028] The model construction module is used to construct a convolutional neural network model EfficientNetV2, which successively includes: a convolutional layer, an SPPF module, multiple ECA-Fused-MBConv modules and a fully connected layer, and the SENet module of the Fused-MBConv module is replaced with an ECANet module to obtain the ECA-Fused-MBConv module;

[0029] The model training module is used to train the convolutional neural network model EfficientNetV2 based on the training set to obtain a trained convolutional neural network model;

[0030] The model testing module is used to test the convolutional neural network model EfficientNetV2 based on the test set and output the accuracy rate of citrus reticulata blanco grading;

[0031] The citrus reticulata blanco quality grading prediction module is used to obtain the predicted citrus reticulata blanco quality grading information based on the trained convolutional neural network model.

[0032] As a preferred technical solution, acquiring citrus reticulata blanco images and performing data augmentation specifically includes:

[0033] Performing data augmentation operations on citrus reticulata blanco images of different categories, including translation, flipping, rotation, and adding noise.

[0034] As a preferred technical solution, preprocessing the citrus reticulata blanco images specifically includes:

[0035] Using the method of Gaussian filtering to perform noise reduction processing on the citrus reticulata blanco images, scanning each pixel in the citrus reticulata blanco images with a filtering template, and the gray value of each pixel point is obtained by weighted average of itself and the pixel values in the neighborhood, and the size of the neighborhood and the weights of the pixels in the neighborhood are determined by the filtering template;

[0036] Performing image enhancement on the filtered citrus reticulata blanco images, performing power function transformation on the pixel values of the citrus reticulata blanco images based on gamma correction, and adjusting the brightness and contrast of the citrus reticulata blanco images.

[0037] As a preferred technical solution, the ECANet module replaces two fully connected layers with one-dimensional convolutions, and the kernel size is adaptively determined through the non-linear mapping of the channel dimension;

[0038] The ECANet module performs channel global average pooling on the input features, and then through one-dimensional convolution, uses the Sigmoid function to generate the weights of each channel, and combines the input features with the channel weights to obtain the final output.

[0039] As a preferred technical solution, the SPPF module includes an input module, a ConvBNSiLU module, multiple max-pooling layers, a splicing layer, and an output module;

[0040] The splicing layer splices the results before pooling and after each pooling.

[0041] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0042] (1) By introducing the SPPF module into the EfficientNetV2 network model, the present invention improves the receptive field and feature expression ability of the network, and significantly improves the classification performance of the model by improving the Fused-MBConv module in the EfficientNetV2 network model.

[0043] (2) The classification accuracy of the convolutional neural network model proposed by the present invention is relatively high, reaching more than 95%, and it has good stability and strong applicability. Description of the Drawings

[0044] Figure 1 It is a schematic flow chart of the method for grading the quality of ponkan oranges based on the improved EfficientNetV2 of the present invention;

[0045] Figure 2 It is a schematic diagram of the comparison of the Fused-MBConv structures before and after the improvement of the present invention;

[0046] Figure 3 It is a schematic diagram of the ECANet structure of the present invention;

[0047] Figure 4 It is a schematic diagram of the SPPF module structure of the present invention;

[0048] Figure 5 It is a schematic diagram of the structure of the improved EfficientNetV2 of the present invention. Detailed Embodiments

[0049] In order to make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0050] Embodiment 1

[0051] As Figure 1 shown, this embodiment provides a method for grading the quality of ponkan oranges based on the improved EfficientNetV2, including the following steps:

[0052] S1: Obtain ponkan orange images through the Internet and an image acquisition device, perform data augmentation, and specifically amplify images of different categories;

[0053] In this embodiment, ponkan orange images can be obtained from the Internet. Through various search engines and databases, relevant ponkan orange image samples can be collected. Some image sharing websites such as Flickr, Pinterest, Instagram, etc., or image databases opened by some academic institutions, as well as some agricultural information sites and other channels can be used to collect relevant ponkan orange image samples. And attention should be paid to the balance of ponkan orange image samples to avoid a situation where the number of samples in different categories varies greatly. In addition, ponkan orange images can also be obtained through an image acquisition device. The image acquisition device includes parts such as a light box, a camera, and an LED light source. The ponkan oranges are placed in the light box for image acquisition. To simulate the actual production line environment, the ponkan oranges can also be placed on a conveyor belt, and multiple high-precision cameras can collect images of the ponkan oranges from multiple angles;

[0054] In this embodiment, after collecting the ponkan orange images, perform data augmentation operations on the obtained ponkan orange images. Operations such as translation, flipping, rotation, and adding noise can be specifically performed on ponkan orange images of different categories to increase the number of the sample data set, enable the network model to be fully trained, and make the distribution of the sample data set more balanced, avoiding the situation that the model tends to focus on the multi-sample category during the training process due to the excessive number of samples in certain categories, resulting in a decrease in the classification accuracy of the few-sample categories;

[0055] S2: Preprocess the obtained ponkan orange images, including methods such as Gaussian filtering and image enhancement;

[0056] In this embodiment, first, the Gaussian filtering method is used to denoise the ponkan orange image to avoid the influence of noise on image recognition. Specifically, each pixel in the image is scanned with a filtering template, and the gray value of each pixel point is obtained by weighted averaging of its own and the pixel values in the neighborhood. The size of the neighborhood and the weights of the pixels in the neighborhood are determined by the filtering template, and the size of the filtering template is 3*3. Then, the filtered image is enhanced by image enhancement, and the power function transformation is performed on the pixel values of the image by using the gamma correction method to adjust the brightness and contrast of the image, improve the contrast between the defective area and the normal area of the ponkan orange, make the details of the image clearer, and improve the recognition ability of the model for non-obvious defects.

[0057] S3: Divide the preprocessed ponkan orange image dataset into a training set, a validation set, and a test set according to a set ratio.

[0058] In this embodiment, the training set accounts for 70%, the validation set accounts for 15%, and the test set accounts for 15%. According to the shape characteristics, fruit surface defects, fruit surface color, etc. of the ponkan orange, the ponkan oranges are divided into four categories: super-grade fruits, first-grade fruits, second-grade fruits, and unqualified fruits, and the ponkan orange images are saved in different folders according to the categories.

[0059] In this embodiment, taking citrus as the grading object, the grading standard for super-grade fruits is: short oval shape, regular and round fruit appearance, clean fruit surface, smooth peel, allowing extremely slight defects, no fruit spots, no edema, withered water, and puffy peel fruits, no mechanical damage, and the fruit surface is orange-red or orange-red in color with uniform coloring; the grading standard for first-grade fruits is: short oval or oblate shape, relatively regular and round fruit appearance, clean fruit surface, relatively smooth peel, allowing slight defects, the number of fruit spots does not exceed 1, the diameter of the fruit spot < 2.5mm, no edema, withered water, and puffy peel fruits, no mechanical damage, and the fruit surface is light orange-red or light red in color with uniform coloring; the grading standard for second-grade fruits is: short oval or oblate shape, relatively regular and round fruit appearance, relatively clean fruit surface, allowing slight defects, the number of fruit spots exceeds 1, the diameter of the fruit spot < 3mm, allowing slight edema, withered water, and puffy peel fruits, allowing slight mechanical damage, and the fruit surface is light orange-yellow in color with relatively uniform coloring; those that do not meet the above standards are unqualified fruits.

[0060] S4: Construct a convolutional neural network model EfficientNetV2, adjust and optimize its structure, add an SPPF module and improve the Fused-MBConv module, which can improve the applicability and classification accuracy of the model without greatly affecting the computational amount and training time of the network.

[0061] In this embodiment, by changing the SENet part inside the Fused-MBConv module to the ECANet, and by introducing a one-dimensional convolutional kernel, the information of each channel and its adjacent channels can be better retained and interacted, so as to capture local features more effectively;

[0062] In this embodiment, the SPPF module includes an input-output module, a synthesis module of convolution plus batch normalization plus SiLU activation function, a max pooling layer, and a concatenation layer. By adding the SPPF module to the EfficientNetV2 network, problems such as image distortion caused by cropping and scaling operations on image regions can be avoided.

[0063] In this embodiment, a citrus reticulata blanco image classification network model is constructed, which is improved based on the EfficientNetV2 model. As Figure 5 shown, in the improved network structure, Input represents the input feature map, Conv represents the convolutional layer, n*n represents the size of the convolutional kernel, Layers represents the number of layers, MBConv represents an inverted linear bottleneck layer with depthwise separable convolution, which was proposed in MobileNetV2, Concat represents the concatenation operation, and FC represents the fully connected layer.

[0064] In this embodiment, ECA-Fused-MBConv represents a module improved based on Fused-MBConv proposed in EfficientNetV2. The comparison before and after the improvement is as Figure 2 shown, the improved ECANet structure is as Figure 3 shown. The ECANet module replaces two fully connected layers with one-dimensional convolutions, and its kernel size can be adaptively determined through non-linear mapping of the channel dimension; first, channel-wise global average pooling (GAP) is implemented on the input features without dimensionality reduction, and then through a fast 1D convolution with a size of k, the weights of each channel can comprehensively consider the information of its adjacent channels. Then, the Sigmoid function is used to generate the weights of each channel, and the input features are combined with the channel weights (the same as the final processing process in SENet) to obtain the final output. SENet uses fully connected layers for global information aggregation, which will ignore local interaction information. In contrast, ECANet makes the information of each channel and its adjacent channels better retained and interacted by introducing a one-dimensional convolutional kernel, so as to capture local features more effectively. The introduction of fully connected layers in the SENet module increases the computational complexity and the number of parameters, while ECANet avoids dimensionality reduction and upsampling operations by using one-dimensional convolutions, can reduce the complexity of the model, and makes the model more lightweight.

[0065] In this embodiment, the core component of the EfficientNetV2 network is the Fused-MBConv module. The improvement of this module aims to reduce the model complexity while increasing the recognition rate of ponkan samples. In the improved ECA-Fused-MBConv module, first, a dimensionality increase operation of Conv 3×3 is performed, then the feature matrix is adjusted using the attention mechanism SENet module, and the importance of each feature channel is automatically determined through learning to better retain key feature information. Finally, dimensionality reduction output is performed through Conv 1×1. When adjusting the attention mechanism, the ECANet module is used to replace the SENet module in the original model to ensure automatic learning of the importance of each channel, while effectively reducing the model complexity, thereby improving the performance.

[0066] In addition, an SPPF module is added to the network, and its structural diagram is as Figure 4 shown. In the figure, Input and Output represent the input feature map and the output feature map, ConvBNSiLU represents a combined module of convolution, batch normalization, and SiLU activation function, MaxPool2d represents the max pooling layer, n×n represents the size of the pooling kernel, and Concat represents the concatenation operation. By adding the Spatial Pyramid Pooling-Fast (SPPF) module to the EfficientNetV2 network, problems such as image distortion caused by cropping and scaling operations on image regions can be effectively avoided. At the same time, the problem of repetitive feature extraction of convolutional neural networks for images is solved, and the computational cost is saved. The SPPF module is improved based on the SPP module, and its speed has been significantly improved compared to SPP. While ensuring multi-scale fusion, the computational amount is reduced, and the receptive field and feature expression ability of the network are improved. In the SPPF structure, first, a residual structure is passed through, then max pooling is performed three times continuously, the size of the convolutional kernel is uniformly 5*5, and finally, the results before pooling and after each pooling are concatenated.

[0067] S5: Set the training parameters. After a large number of trainings and debuggings, the convolutional neural network model with the best defect recognition effect is obtained;

[0068] In this embodiment, the epoch is set to 200, the batch size is set to 16, and the learning rate is set to 0.0001. The exponential decay method is used to dynamically adjust the learning rate. At the beginning of the training, an initial learning rate and a decay rate are set. During the training process, the current learning rate is calculated based on the comparison between the global step and the decay step, and the iteration continues until the training ends. The learning rate will continuously decay according to the exponential function. Through the learning rate exponential decay method, the learning rate can be better adjusted, the training effect of the model can be improved, and the situation of unstable training or slow convergence caused by unreasonable learning rate settings can be avoided. The Adam optimizer is used, and through a large amount of training and debugging, the network model with the best defect recognition effect is obtained.

[0069] S6: Call the convolutional neural network model to perform classification tests on the test set;

[0070] In this embodiment, call the convolutional neural network model to perform classification tests on the test set, use the classification accuracy as the model evaluation criterion to verify the model performance. By comparing the classification results with the true categories of ponkan oranges, it can be detected whether the grading method has the ability to grade the quality of ponkan oranges, and the accuracy of ponkan orange grading is output.

[0071] S7: Deploy the trained convolutional neural network model to the PC side and combine it with the motion control module to achieve ponkan orange sorting;

[0072] In this embodiment, deploy the trained convolutional neural network model to the PC side. When performing ponkan orange sorting operations, first collect ponkan orange images through a camera, then preprocess the images, and then input them into the convolutional neural network model to obtain the predicted ponkan orange grade information. Finally, send the prediction results to the control center and cooperate with the motion control module to achieve the sorting of ponkan oranges, thus completing the ponkan orange quality grading of the improved EfficientNetV2.

[0073] In this embodiment, when performing ponkan orange grading operations, first run the vision code on the PC side to control the camera to collect ponkan orange images, then perform the above preprocessing operations on the obtained images, and then use the preprocessed images as the input of the network model. After the network model outputs the grade information of ponkan oranges, send the prediction results to the STM32 single-chip microcomputer, and control the steering angle of the servo by outputting different PWM signals, thereby controlling the sorting disk to perform sorting work and transporting the ponkan oranges to the corresponding grade positions.

[0074] The present invention can effectively improve efficiency and reduce the consumption of unnecessary human and material resources, effectively alleviate the difficulty of screening and classification in the case of a large number of ponkan oranges, and can effectively classify various ponkan oranges by using convolutional neural network detection.

[0075] Embodiment 2

[0076] This embodiment provides a ponkan quality grading system based on the improved EfficientNetV2, which is used to implement the above-mentioned ponkan quality grading method based on the improved EfficientNetV2. The system includes: an image acquisition module, a data augmentation module, an image preprocessing module, a data partitioning module, a model construction module, a model training module, a model testing module, and a ponkan quality grading prediction module;

[0077] In this embodiment, the image acquisition module is used to acquire ponkan images;

[0078] In this embodiment, the data augmentation module is used to perform data augmentation on the ponkan images;

[0079] In this embodiment, the image preprocessing module is used to preprocess the ponkan images;

[0080] In this embodiment, the data partitioning module is used to partition the preprocessed ponkan images into a training set, a validation set, and a test set, and mark different quality levels;

[0081] In this embodiment, the model construction module is used to construct a convolutional neural network model EfficientNetV2, which sequentially includes: a convolutional layer, an SPPF module, multiple ECA-Fused-MBConv modules, and a fully connected layer. The SENet module of the Fused-MBConv module is replaced with an ECANet module to obtain the ECA-Fused-MBConv module;

[0082] In this embodiment, the model training module is used to train the convolutional neural network model EfficientNetV2 based on the training set to obtain a trained convolutional neural network model;

[0083] In this embodiment, the model testing module is used to test the convolutional neural network model EfficientNetV2 based on the test set and output the accuracy rate of ponkan grading;

[0084] In this embodiment, the ponkan quality grading prediction module is used to obtain the predicted ponkan quality grading information based on the trained convolutional neural network model.

[0085] In this embodiment, acquiring ponkan images and performing data augmentation specifically includes:

[0086] Performing data augmentation operations on ponkan images of different categories, including translation, flipping, rotation, and adding noise.

[0087] In this embodiment, preprocessing the ponkan images specifically includes:

[0088] The wogan image is denoised by using the Gaussian filtering method. Each pixel in the wogan image is scanned with a filtering template, and the gray value of each pixel point is obtained by weighted averaging of its own and the pixel values in the neighborhood. The size of the neighborhood and the weights of the pixels in the neighborhood are determined by the filtering template;

[0089] The filtered wogan image is enhanced. Based on gamma correction, the pixel values of the wogan image are subjected to a power function transformation to adjust the brightness and contrast of the wogan image.

[0090] In this embodiment, the ECANet module replaces two fully connected layers with one-dimensional convolutions, and its kernel size is adaptively determined by the non-linear mapping of the channel dimension;

[0091] The ECANet module performs channel global average pooling on the input features, and then through one-dimensional convolution, uses the Sigmoid function to generate the weights of each channel, and combines the input features with the channel weights to obtain the final output.

[0092] In this embodiment, the SPPF module includes an input module, a ConvBNSiLU module, multiple max-pooling layers, a splicing layer, and an output module;

[0093] The splicing layer splices the results before pooling and after each pooling.

[0094] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited by the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A quality grading method for mandarin oranges based on improved EfficientNetV2, characterized in that: include: Obtain Wogan images and perform data augmentation; Preprocess the Wogan images; The preprocessed mandarin orange images were divided into training set, validation set and test set, and marked with different quality levels; Construct a convolutional neural network model EfficientNetV2, which includes: convolutional layer, SPPF module, multiple ECA-Fused-MBConv modules and fully connected layer. Replace the SENet module of the Fused-MBConv module with the ECANet module to obtain the ECA-Fused-MBConv module. The convolutional neural network model EfficientNetV2 is trained based on the training set to obtain the trained convolutional neural network model; The convolutional neural network model EfficientNetV2 was tested based on the test set, and the accuracy of grading of mandarin oranges was output; The predicted quality grading information of Wogan oranges was obtained based on the trained convolutional neural network model.

2. The quality grading method of mandarin oranges based on improved EfficientNetV2 according to claim 1, characterized in that: Obtain the mandarin orange image and perform data enhancement, including: Data augmentation operations are performed on different categories of mandarin orange images, including translation, flipping, rotation, and adding noise.

3. The quality grading method of mandarin oranges based on improved EfficientNetV2 according to claim 1, characterized in that: Preprocessing of the mandarin orange image includes: The Gaussian filtering method is used to reduce the noise of the mandarin orange image. A filtering template is used to scan each pixel in the mandarin orange image. The gray value of each pixel is obtained by weighted average of its own pixel value and the pixel value in the neighborhood. The size of the neighborhood and the weight of the pixels in the neighborhood are determined by the filtering template. The filtered mandarin orange image is enhanced by image enhancement, and the pixel value of the mandarin orange image is transformed by power function based on gamma correction to adjust the brightness and contrast of the mandarin orange image.

4. The quality grading method of mandarin oranges based on improved EfficientNetV2 according to claim 1, characterized in that: The ECANet module replaces two fully connected layers with one-dimensional convolutions, whose kernel size is adaptively determined by nonlinear mapping of the channel dimension; The ECANet module implements channel global average pooling on the input features, and then generates the weight of each channel through one-dimensional convolution using the Sigmoi function, combining the input features with the channel weights to obtain the final output.

5. The quality grading method of mandarin oranges based on improved EfficientNetV2 according to claim 1, characterized in that: The SPPF module includes an input module, a ConvBNSiLU module, multiple maximum pooling layers, a splicing layer, and an output module; The concatenation layer concatenates the results before and after pooling.

6. A quality grading system for mandarin oranges based on improved EfficientNetV2, characterized in that: include: Image acquisition module, data enhancement module, image preprocessing module, data partitioning module, model building module, model training module, model testing module, and Wogan quality grading prediction module; The image acquisition module is used to acquire the image of the mandarin orange; The data enhancement module is used to perform data enhancement on the mandarin orange image; The image preprocessing module is used to preprocess the mandarin orange image; The data division module is used to divide the pre-processed mandarin orange images into a training set, a validation set and a test set, and mark different quality levels; The model construction module is used to construct a convolutional neural network model EfficientNetV2, which includes: a convolutional layer, an SPPF module, multiple ECA-Fused-MBConv modules and a fully connected layer, and the SENet module of the Fused-MBConv module is replaced with the ECANet module to obtain the ECA-Fused-MBConv module; The model training module is used to train the convolutional neural network model EfficientNetV2 based on the training set to obtain a trained convolutional neural network model; The model testing module is used to test the convolutional neural network model EfficientNetV2 based on the test set and output the accuracy of the grading of mandarin oranges; The Wogan quality grading prediction module is used to obtain predicted Wogan quality grading information based on the trained convolutional neural network model.

7. The quality grading system of mandarin oranges based on improved EfficientNetV2 according to claim 6 is characterized in that: Obtain the mandarin orange image and perform data enhancement, including: Data augmentation operations are performed on different categories of mandarin orange images, including translation, flipping, rotation, and adding noise.

8. The quality grading system for mandarin oranges based on improved EfficientNetV2 according to claim 6, characterized in that: Preprocessing of the mandarin orange image includes: The Gaussian filtering method is used to reduce the noise of the mandarin orange image. A filtering template is used to scan each pixel in the mandarin orange image. The gray value of each pixel is obtained by weighted average of its own pixel value and the pixel value in the neighborhood. The size of the neighborhood and the weight of the pixels in the neighborhood are determined by the filtering template. The filtered mandarin orange image is enhanced by image enhancement, and the pixel value of the mandarin orange image is transformed by power function based on gamma correction to adjust the brightness and contrast of the mandarin orange image.

9. The quality grading system of mandarin oranges based on improved EfficientNetV2 according to claim 6, characterized in that: The ECANet module replaces two fully connected layers with one-dimensional convolutions, whose kernel size is adaptively determined by nonlinear mapping of the channel dimension; The ECANet module implements channel global average pooling on the input features, and then generates the weight of each channel through one-dimensional convolution using the Sigmoi function, combining the input features with the channel weights to obtain the final output.

10. The quality grading system of mandarin oranges based on improved EfficientNetV2 according to claim 6, characterized in that: The SPPF module includes an input module, a ConvBNSiLU module, multiple maximum pooling layers, a splicing layer, and an output module; The concatenation layer concatenates the results before and after pooling.

Images