MD-YOLOv8 strawberry diseased leaf detection method

Abstract

The invention discloses a strawberry diseased leaf detection method based on MD-YOLOv8, and belongs to the field of computer vision target detection. The method comprises the following steps: firstly, acquiring strawberry leaf disease images, marking five disease categories such as powdery mildew and gray mold, and dividing into a training set and a test set; and training an MD-YOLOv8 model on the training set, and testing by the test set to obtain the model for strawberry leaf disease detection. An MD-YOLOv8 network model is improved based on YOLOv8, and a multi-scale convolution attention MSCAN module is added into a trunk feature extraction network to improve the context information extraction capability; the neck network adopts a DySample lightweight dynamic up-sampling module, so that sampling points focus on a target area, and the background is prevented from being concerned. Experiments show that the model can quickly and accurately identify the types of diseased strawberry leaves, and can be effectively applied to strawberry disease detection scenes.

Description

Technical Field

[0001] The present invention relates to the field of computer vision object detection, and particularly to a method for detecting strawberry diseased leaves based on MD-YOLOv8. Background Art

[0002] As a crop with relatively high economic value, strawberries are vulnerable to diseases during their growth process, which seriously affects yield, quality and economic benefits. The detection of strawberry diseased leaves plays a crucial role in ensuring crop health, improving yield and quality, reducing disease spread, reducing the use of chemical pesticides, and enhancing the level of agricultural intelligence.

[0003] However, traditional plant disease monitoring methods rely on disease experts to conduct on-site inspections and confirm the disease situation. However, this method is time-consuming, laborious, costly, and it is difficult to accurately predict the spread of diseases in the field. On the other hand, biological detection is accurate but complex to operate and may have misjudgments; while traditional machine learning methods rely on feature engineering, resulting in relatively low recognition accuracy. By combining deep learning and computer vision technologies, accurate, efficient and automated disease monitoring can be achieved, providing strong technical support for smart agriculture.

[0004] With the rapid development of convolutional neural networks (CNNs), in recent years, more and more researchers have applied deep learning-based algorithms to crop disease detection. Compared with traditional object detection algorithms, CNN-based detection methods have stronger generalization ability and can automatically learn the deep semantic information of objects. Object detection methods based on deep convolutional networks are generally divided into two categories: one-stage methods, such as the YOLO series and SSD; two-stage methods, such as the region-based CNN series (R-CNN, Fast R-CNN, Faster R-CNN, etc.).

[0005] One-stage detection methods can significantly improve the detection speed at the expense of certain accuracy by predicting the recognition and localization tasks of object bounding boxes at one time; while two-stage methods first generate proposal boxes and then classify and localize them. Although the accuracy is higher, the speed is slower. Therefore, while ensuring that the detection speed meets the real-time requirements, improving the accuracy of one-stage detection methods is of great significance for improving strawberry yield and quality, and even promoting the development of agriculture towards modern intelligence.

[0006] Although existing diseased leaf detection algorithms have made significant progress in simple backgrounds, there are still deficiencies in detecting strawberry diseased leaves in complex field environments. First, the target areas of some strawberry diseased leaves containing pest parts are relatively small, which increases the recognition difficulty and thus affects the detection accuracy. There is an urgent need in this field for a method that can quickly and accurately identify strawberry diseased leaves. Summary of the Invention

[0007] The purpose of the present invention is to provide a strawberry diseased leaf detection method based on the MD-YOLOv8 model.

[0008] The technical solution of the present invention is to design and train the MD-YOLOv8 model, and use this model to identify the types of strawberry diseased leaves contained in the picture.

[0009] The MD-YOLOv8 model framework is as Figure 1 shown. Its overall architecture is inherited from the YOLOv8 model ( Figure 2 ). On the basis of inheriting this model framework, the main difference between MD-YOLOv8 and YOLOv8 lies in:

[0010] The backbone network of the MD-YOLOv8 network model uses CSPDarknet for feature extraction, and mainly completes the feature extraction operation through a series of Conv, C2f, and SPPF modules; we introduce a multi-scale convolutional attention MSCAN module behind the SPPF module to improve the ability of the backbone network to extract context information. Moreover, in the neck network, the DySample lightweight dynamic upsampling module is used to replace the upsampling module of the neck network, so that the sampling points are concentrated in the target area and do not focus on the background part, realizing the effective recognition of the detection target.

[0011] The total loss function of MD-YOLOv8 is composed of the classification loss function , the regression loss function and , and is defined by formula (1).

[0012] (1)

[0013] where , , are respectively defined by formulas (2), (3), and (4).

[0014] (2)

[0015] (3)

[0016] (4)

[0017] where, represents the true label of classification; represents the predicted value of the classification bounding box; and respectively represent the IoU loss function, the ground truth bounding box, and the predicted bounding box; the variable d is the Euclidean distance between the centers of these two bounding boxes, while c represents the smallest convex bounding box enclosing these two bounding boxes. In addition, v is a parameter for quantifying the aspect ratio consistency.

[0018] The MD-YOLOv8 model is iteratively trained using the training set. After the loss value stabilizes, it is saved as a trained model. The trained model is evaluated using the test set; if the evaluation result is not satisfactory, parameter adjustment is performed and retraining is carried out.

[0019] The present invention proposes a detection method for identifying the types of strawberry diseased leaves contained in RGB images based on the MD-YOLOv8 model. This method breaks through the traditional manual-dependent plant disease monitoring method and can quickly and accurately identify the strawberry diseased leaf information in the image, promoting the development of agriculture towards the direction of modern intelligence. Brief Description of the Drawings

[0020] Figure 1 MD-YOLOv8 Model Architecture

[0021] Figure 2 YOLOv8 Model Architecture

[0022] Figure 3 5 Annotated Examples of Strawberry Diseased Leaves

[0023] Figure 4 Examples of Strawberry Diseased Leaf Recognition Results

[0024] Figure 5 Flowchart of Strawberry Diseased Leaf Detection Based on MD-YOLOv8 Detailed Embodiments

[0025] The present invention is a method for detecting strawberry diseased leaves based on the MD-YOLOv8 model, and the specific steps are as follows:

[0026] (1) Approximately 2000 images containing strawberry diseased leaves were collected from the Internet, and the strawberry diseased leaf objects contained in the images were annotated according to 5 states: powdery mildew, gray mold, black spot, withered flower, and fruit rot. Figure 3 Two sets of annotation examples were given for each state.

[0027] (2) The correctly annotated images were randomly divided into a training set and a test set in a ratio of 7:3.

[0028] (3) Set the parameters related to the training of the MD-YOLOv8 model, including: the learning rate is set to 0.001, the weight decay is set to 0.01, and the momentum is set to 0.937.

[0029] (4) Iteratively train the MD-YOLOv8 model with the images in the training set until the total loss value of the model tends to be stable. Save the trained MD-YOLOv8 model.

[0030] (5) Evaluate the performance of the saved MD-YOLOv8 model using the test set. If the expected effect is not achieved, adjust the parameters related to the training of the MD-YOLOv8 model and restart step (4).

[0031] (6) Use the MD-YOLOv8 model that has passed the evaluation to identify the strawberry diseased leaves contained in the image and the types of strawberry diseased leaves they belong to. Figure 4 Ten examples of the recognition results are given. Among them, the strawberry diseased leaves detected by the MD-YOLOv8 model are marked with purple frames. The letter before the vertical line in the frame represents the abbreviation of the type of strawberry diseased leaf (pm represents powdery mildew, gm represents gray mold, bs represents black spot, wf represents blossom blight, fr represents fruit rot), and the number after the vertical line represents the confidence level of the prediction result. The larger the value, the more credible it is.

Claims

1. A strawberry diseased leaf detection method based on MD-YOLOv8, characterized in that, Specifically, it includes the following steps: S1: Collect strawberry diseased leaf images, label the pictures according to five categories of powdery mildew, gray mold, black spot, blossom blight, and fruit rot, and divide the dataset into a training set and a test set; S2: Construct an MD-YOLOv8 network model; S3: Use the training set described in S1 to train the MD-YOLOv8 network model described in S2 to obtain an MD-YOLOv8 strawberry diseased leaf detection model; S4: Use the trained MD-YOLOv8 network model in S3 to perform strawberry diseased leaf detection tests in the test set and predict the specific categories of the diseased leaves.

2. The method for detecting strawberry diseased leaves based on MD-YOLOv8 according to claim 1, characterized in that, The main features of the MD-YOLOv8 network model in step S2 are as follows: S2.1: The MD-YOLOv8 network model is improved based on YOLOv8, and its overall architecture is the same as that of YOLOv8, which successively includes a backbone network, a neck network, and a detection head; S2.2: The backbone network of the MD-YOLOv8 network model uses CSPDarknet for feature extraction, and mainly completes the feature extraction operation through a series of Conv, C2f, and SPPF modules; we introduce a multi-scale convolutional attention MSCAN module behind the SPPF module to improve the ability of the backbone network to extract context information; S2.3: The neck network of the MD-YOLOv8 network model is PA-FPN. In the neck network, the DySample lightweight dynamic upsampling module is used to replace the upsampling module of the neck network, so that the sampling points are concentrated in the target area and do not focus on the background part, realizing the effective recognition of the detection target; S2.4: The detection head of the MD-YOLOv8 network model is the same as that of YOLOv8, including two types of heads: one for regression and localization, and the other for classification.

3. A strawberry diseased leaf detection method based on MD-YOLOv8 according to claim 1, characterized in that, The main features of the loss function used in the training of the MD-YOLOv8 network model in step S3 are as follows: S3.1 Total loss function of MD-YOLOv8 Consisting of the classification loss function , the regression loss function and is defined by Equation (1): (1) S3.2 and and are defined by formulas (2), (3), and (4) respectively; (2) (3) (4) Among them, represents the true label of the classification; represents the predicted value of the classification bounding box; and respectively represent the IoU loss function, the true box and the predicted box; the variable d represents the Euclidean distance between the centers of the two bounding boxes, and c represents the smallest convex hull box enclosing the two bounding boxes; in addition, v is a parameter for quantifying the aspect ratio consistency.

Images