Tea pest and disease detection method and computer device based on deep learning

Abstract

The present invention relates to the field of tea pest detection, and specifically to a tea pest detection method and computer device based on deep learning, which improves the accuracy of tea pest detection. The scheme includes: using RGB sensors and infrared sensors to collect images of tea; processing the collected images, processing the collected low-bit images through planned perceptual distortion, the low-bit images add offsets to the neighboring pixels of the pixels through a predefined planned perceptual distortion template, and the offset pixel values ​​enter the quantizer, and each offset signal is quantized into different quantization values ​​by the quantizer, and different quantization values ​​form multiple descriptions of the image signal; then feature extraction and reconstruction are performed, and the reconstructed features are input into a selective state space model, the selective state space model is trained and optimized, and finally the pest detection results are output. The present invention is suitable for tea pest detection.

Description

Technical Field

[0001] The present invention relates to the field of tea plant pest and disease detection, and in particular to a tea plant pest and disease detection method and a computer device based on deep learning. Background Art

[0002] In recent years, with the development of intelligent agriculture, the application of image recognition technology based on computer vision and deep learning in tea pest and disease monitoring has gradually attracted attention. Traditional image processing methods usually rely on manual feature extraction, and the effect is limited by the image clarity and the effectiveness of feature selection. For example, the use of image enhancement algorithms such as histogram equalization and contrast stretching can improve image quality, but often cannot fully capture the subtle differences of tea leaves under complex backgrounds.

[0003] In this context, convolutional neural networks have become mainstream due to their successful application in the field of image processing. However, existing convolutional neural network models still have certain limitations when processing multi-scale and complex agricultural images, such as insufficient recognition accuracy for specific pests and diseases. In addition, existing technologies often ignore the integration of prior knowledge, such as the growth environment of tea and the characteristics of pests and diseases, which limits the generalization ability and accuracy of the model.

[0004] Defects and shortcomings of existing technology

[0005] (1) Poor image processing effect: Traditional image enhancement technologies (such as histogram equalization and sharpening) cannot effectively solve problems such as image blur and uneven lighting when processing tea images, resulting in poor subsequent feature extraction and recognition effects.

[0006] (2) Low feature extraction efficiency: Existing manual feature extraction methods often rely on expert experience, and the processing process is cumbersome and inefficient, making it difficult to adapt to the needs of large-scale tea planting.

[0007] (3) Insufficient model generalization ability: Existing convolutional neural networks are prone to overfitting when trained on specific data sets, resulting in reduced recognition performance in new environments, especially when the types of tea pests and diseases are diverse and the samples are unbalanced.

[0008] (4) Lack of integration of prior knowledge: Many existing technologies fail to make full use of agricultural knowledge, such as the growing environment of tea and the characteristics of pests and diseases, resulting in poor performance of the model in specific scenarios.

[0009] (5) High cost and complexity: Some existing technologies require expensive equipment or complex software systems, which increases the burden on agricultural producers and is not conducive to promotion and application. Summary of the invention

[0010] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a tea pest and disease detection method and computer device based on deep learning, thereby improving the accuracy of tea pest and disease detection.

[0011] The present invention adopts the following technical solutions to achieve the above-mentioned purpose. In a first aspect, the present invention provides a tea pest and disease detection method based on deep learning, comprising:

[0012] S1, collect tea images;

[0013] S2, processing the collected images;

[0014] The collected low-bit image is processed by planned perceptual distortion. The low-bit image adds an offset to the neighboring pixels of the pixel through a predefined planned perceptual distortion template. The offset pixel value enters the quantizer. Each offset signal is quantized into different quantization values ​​by the quantizer. Different quantization values ​​form multiple descriptions of the image signal. The planned perceptual distortion template is defined as follows:

[0015] , PSD Indicates that the plan perception is distorted, represents the planned perceptual distortion template, k represents the template size, and Q represents the quantization step size;

[0016] Then, the tea surface image is smoothed using Gaussian filtering, and the adaptive threshold processing technology is applied to automatically adjust the binarization threshold according to the different lighting conditions of the tea leaves, so that the contours of the disease spots and insect pests on the entire leaf can be extracted more accurately.

[0017] S3, feature extraction and reconstruction;

[0018] The low-bit image after planned perceptual distortion processing is input into the pre-trained PSD-Mamba model, and the deep features of the tea surface details are extracted through the convolutional neural network in the PSD-Mamba model. The deep features are the features of the shape, color change and insect pest marks of the tea spots.

[0019] Then, the features extracted by convolutional layers of different depths of the convolutional neural network are fused and reconstructed through a densely connected network;

[0020] Finally, the attention mechanism is used to strengthen the fused and reconstructed features;

[0021] S4, Optimization of pest and disease detection based on selective state space model;

[0022] The reconstructed features are input into the selective state space model, and the selective state space model is trained. During the training process, for disease spots and insect pest marks, the selective state space model will selectively enhance the color contrast and edge texture of the damaged area. When processing the internal structure of the leaves, the selective state space model will strengthen the water loss and temperature changes caused by the disease. For the gnawing marks on the edge of the tea leaves, the selective state space model will dynamically adjust the feature transmission of the gnawing mark area on the edge of the tea leaves, and optimize the disease and insect pest detection process by dynamically adjusting the transmission path of the input data features.

[0023] S5. Output and quality assessment of pest and disease detection results;

[0024] Tea pests and diseases are detected through trained and optimized selective state space models, and the detection results are generated through image analysis models, covering tea spots, pest traces and leaf tissue damage data. The detection results of each tea leaf are evaluated for accuracy through quantitative analysis algorithms.

[0025] Furthermore, step S1 specifically includes:

[0026] An RGB sensor and an infrared sensor are used to collect images of tea leaves. The RGB sensor is used to capture the color changes, disease spots and visual features of insect bites on the surface of the tea leaves, and the infrared sensor is used to detect temperature changes and internal tissue damage caused by diseases and insect pests. The analog signals collected by the RGB sensor and the infrared sensor are then converted into digital signals through an analog-to-digital converter, and the damage features of the disease and insect pest areas are preferentially collected by dynamically adjusting the quantization parameters.

[0027] Furthermore, in step S5, the evaluation process specifically includes:

[0028] Through quantitative analysis algorithms, the color changes of detected lesions, the degree of damage to leaf edges, and the temperature data in infrared images are quantitatively analyzed. The lesion area, shape, and depth data are compared with the corresponding data in the standard feature library to verify the reliability and consistency of the test results.

[0029] In a second aspect, the present invention provides a computer device, including a memory, wherein the memory stores program instructions, and when the program instructions are run, the above-mentioned tea disease and pest detection method based on deep learning is executed.

[0030] The beneficial effects of the present invention are:

[0031] The present invention introduces planned perceptual distortion, which can accurately capture the subtle features of tea pests and diseases, such as small spots or pests on tea buds, reduce the estimation error in feature recognition, and suppress artifacts caused by factors such as low light and complex background.

[0032] The present invention adds a PSD-Mamba model, inputs a low-bit image after planned perceptual distortion processing into a pre-trained PSD-Mamba model, extracts features of disease spots, insect damage and leaf tissue changes on tea leaves, and then uses a convolutional neural network to extract features of the shape, color changes and insect traces of tea leaf disease spots on the surface details of the tea leaves. Finally, the extracted features are integrated through a densely connected network. This process ensures that the characteristic information of diseases and pests can be efficiently extracted in different parts of the leaves and under different lighting conditions, thereby improving the accuracy and comprehensiveness of tea disease and pest identification.

[0033] The present invention introduces a selective state space model, realizes an efficient linear expansion training process in the task of pest and disease detection, can selectively process the input pest and disease data, adjust the signal transmission path based on the input dynamic feature information, reduce the propagation of redundant information, and reduce the overall computational complexity. At the same time, when processing long-sequence tea pest and disease image data, it can significantly improve the response speed and data processing efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flow chart of a tea disease and pest detection method based on deep learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0035] To make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0036] The present invention provides a tea pest detection method based on deep learning, such as Figure 1 As shown, specifically including:

[0037] S1, collect tea images;

[0038] The present invention uses high-definition optical sensors (such as RGB sensors) and multi-spectral sensors (such as infrared sensors) to collect images of tea leaves. The RGB sensors mainly capture visual features such as color changes, disease spots, and insect bites on the surface of the tea leaves, while the infrared sensors detect temperature changes and internal tissue damage caused by diseases and insect pests.

[0039] The analog signals collected by these sensors are converted into digital signals through analog-to-digital converters, and the key obvious damage features of the pest and disease areas are collected first by dynamically adjusting the quantitative parameters to ensure that no important details are missed during the collection process. Then, through PSD (Planned Sensor Distortion), the contrast and clarity of subtle features such as disease spots or leaf damage are enhanced in low light or complex environments, making them easier to identify and reducing the neglect of damaged leaves. Multi-description technology generates multiple versions of image descriptions by fusing RGB and infrared images, and processes different spectral information in high-brightness and low-brightness areas to ensure that even if some areas cannot be accurately captured due to light or noise interference, other descriptions can still retain the details of pests and diseases. These technologies work together in multiple aspects such as tea disease spots, leaf edge pest damage, and changes in tea moisture content. Through precise data processing and fusion, the accuracy and reliability of tea disease and pest detection are comprehensively improved.

[0040] S2, processing the collected images;

[0041] The collected low-bit images will first be processed by planned perceptual distortion to ensure that details such as color changes, lesions and insect damage of tea leaves still maintain the integrity of the dynamic range under low light or complex conditions. In the PSD algorithm, the low-bit image adds an offset to the neighboring pixels of the pixel through a predefined PSD template. The offset pixel value enters the quantizer, and each offset signal is quantized by the quantizer into different quantization values. These different quantization values ​​form multiple descriptions of the image signal. The design of the PSD template has a key influence on the effect of the offset. The commonly used template size is k = 3, that is, a 3×3 neighborhood block. The template is selected to ensure that the neighboring pixels satisfy the Markov Property, that is, the value of a pixel depends on its neighboring pixels, so that an effective multiple description can be constructed. The template is defined as follows:

[0042] , It represents the planned perceptual distortion template, Q represents the quantization step size, and is a parameter used to determine the quantization level interval during the quantization process.

[0043] Next, the surface image of the tea leaves is smoothed using Gaussian filtering, especially at the edges of the tea leaves and in the pest and disease areas, to enhance the clarity of the feature areas by reducing noise and small interference. Subsequently, the adaptive threshold processing technology is applied to automatically adjust the binarization threshold according to the different lighting conditions of the tea leaves, so that the contours of the disease spots and pest marks on the entire leaf can be extracted more accurately, avoiding the pseudo-contour effect caused by excessive binarization. This process ensures that the characteristics of tea pests and diseases can be captured clearly and accurately in all kinds of collection environments. This technology is particularly suitable for identifying subtle features such as mold spots, yellow spots, and insect bites on tea leaves. Through the optimized distortion insertion process, the algorithm can significantly improve the display effect of pest and disease characteristics in complex tea garden environments, generating clearer and more natural images.

[0044] S3, feature extraction and reconstruction;

[0045] The low-bit images that have been processed with planned perceptual distortion are input into the pre-trained PSD-Mamba model, which is specifically designed to detect disease spots, insect damage, and leaf tissue changes on tea leaves. The convolutional neural network in the PSD-Mamba model extracts deep features of the surface details of the tea leaves to identify the shape of the disease spots, color changes, and traces of insect pests. At the same time, the densely connected network DenseNet is used to further fuse and reconstruct features at different levels to capture complex information such as subtle damage to the internal tissue of the tea leaves and temperature changes. The different levels refer to convolutional layers of different depths in the convolutional neural network, and the features extracted in each layer have different levels of abstraction and spatial resolution:

[0046] Shallow features: The layers close to the input layer extract relatively basic features, such as edges, textures, and other low-level visual information. These features are very useful for capturing detailed information.

[0047] Mid-level features: The layers in the middle of the network can extract more complex patterns, such as shapes, partial structures, and other mid-level visual information. These features are very important for local descriptions of objects.

[0048] Deep features: The layers close to the output layer extract high-level semantic features, such as object categories, scenes, and other high-level visual information. These features are very important for global understanding.

[0049] Each layer in DenseNet receives not only the output of the previous layer, but also the output of all predecessor layers as input, which are concatenated in the channel dimension. In this way, each layer can directly access the feature maps of all previous layers, achieving effective fusion of features at different levels.

[0050] This process ensures that the characteristic information of pests and diseases can be efficiently extracted from different parts of the leaves and under different lighting conditions, improving the accuracy and comprehensiveness of tea pest and disease identification. Combining the multi-description signals of tea pests and diseases, it can effectively identify the tiny features of pests and diseases such as aphids, mites, and leaf rollers, and enhance feature expression through the Attention Mechanism to improve the accuracy of pest and disease detection.

[0051] S4, Optimization of pest and disease detection based on selective state space model;

[0052] SSM (Selective State Space Models) is used to finely process local areas to improve the detection effect. First, for disease spots and pest traces, SSM selectively enhances the color contrast and edge texture of the damaged area to ensure that these key features are not ignored in complex environments. Second, when processing the internal structure of the leaf, SSM focuses on strengthening the water loss and temperature changes caused by the disease to ensure that the local temperature difference in the infrared image is accurately captured. In addition, SSM pays special attention to the subtle nibbling marks on the edges of tea leaves, dynamically adjusts the feature transfer of these areas, and avoids the influence of uneven lighting or shadow interference on the detection results. Through these local processing, SSM significantly improves the accuracy and comprehensiveness of tea pest and disease detection. The model can also process a large amount of tea garden image data in real time, especially for the rapid and efficient detection and identification of the distribution of pests and diseases in large areas of tea gardens. At the same time, SSM can adjust the model parameters according to the tea growth cycle to ensure the efficient detection of pest and disease characteristics at different growth stages of tea.

[0053] S5. Output and quality assessment of pest and disease detection results;

[0054] The final test results are generated by the image analysis model, mainly covering data such as tea lesions, insect pest traces and leaf tissue damage. Through the quantitative analysis algorithm, the accuracy of the test results of each tea leaf is evaluated to ensure that the identified diseased areas (such as the area of ​​lesions, edge details of insect pests, temperature differences, etc.) have high accuracy. Specifically, the algorithm conducts a refined quantitative analysis of the color changes of the detected lesions, the degree of damage to the leaf edges, and the temperature data in the infrared image, and compares the area, shape and depth data of the lesions with the corresponding data of the standard feature library to verify the reliability and consistency of the test results. This process not only accurately evaluates the test results of each disease, but also ensures a comprehensive analysis of the pest and disease characteristics of tea leaves under different conditions, thereby improving the accuracy of the overall detection. In tea garden scenes with a large dynamic range, especially in the case of complex coexistence of multiple pests and diseases, this method can ensure a high degree of restoration of the test results and provide accurate information on pests and diseases to tea garden managers in a timely manner to assist in decision-making.

[0055] In summary, the present invention achieves efficient tea pest and disease detection by introducing a planned perception distortion strategy and a selective state space model. While coping with the complex environment of the tea garden, the accuracy and computational efficiency of pest and disease identification are significantly improved. Through the optimization of deep learning algorithms and multimodal data fusion, an intelligent pest and disease detection solution is provided for modern tea planting, which is suitable for large-scale tea gardens and complex scenarios where multiple pests and diseases coexist.

[0056] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used in various other combinations, modifications and environments, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art shall not deviate from the spirit and scope of the present invention, and shall be within the scope of protection of the claims attached to the present invention.

Claims

1. A tea pest and disease detection method based on deep learning, characterized in that: include: S1, collect tea images; S2, processing the collected images; The collected low-bit image is processed by planned perceptual distortion. The low-bit image adds an offset to the neighboring pixels of the pixel through a predefined planned perceptual distortion template. The offset pixel value enters the quantizer. Each offset signal is quantized into different quantization values ​​by the quantizer. Different quantization values ​​form multiple descriptions of the image signal. The planned perceptual distortion template is defined as follows: , PSD Indicates that the plan perception is distorted, represents the planned perceptual distortion template, k represents the template size, and Q represents the quantization step size; Then, the tea surface image is smoothed using Gaussian filtering, and the adaptive threshold processing technology is applied to automatically adjust the binarization threshold according to the different lighting conditions of the tea leaves; S3, feature extraction and reconstruction; The low-bit image after planned perceptual distortion processing is input into the pre-trained PSD-Mamba model, and the deep features of the tea surface details are extracted through the convolutional neural network in the PSD-Mamba model. The deep features are the features of the shape, color change and insect pest marks of the tea spots. Then, the features extracted by convolutional layers of different depths of the convolutional neural network are fused and reconstructed through a densely connected network; Finally, the attention mechanism is used to strengthen the fused and reconstructed features; S4, Pest and disease detection optimization based on selective state space model; The reconstructed features are input into the selective state space model, and the selective state space model is trained. During the training process, for disease spots and insect pest marks, the selective state space model will selectively enhance the color contrast and edge texture of the damaged area. When processing the internal structure of the leaves, the selective state space model will strengthen the water loss and temperature changes caused by the disease. For the gnawing marks on the edge of the tea leaves, the selective state space model will dynamically adjust the feature transmission of the gnawing mark area on the edge of the tea leaves, and optimize the disease and insect pest detection process by dynamically adjusting the transmission path of the input data features. S5. Output and quality assessment of pest and disease detection results; Tea pests and diseases are detected through trained and optimized selective state space models. The detection results are generated through image analysis models, covering data on tea spots, pest traces and leaf tissue damage. The detection results of each tea leaf are then evaluated for accuracy through quantitative analysis algorithms.

2. The tea plant pest and disease detection method based on deep learning according to claim 1, characterized in that: Step S1 specifically includes: An RGB sensor and an infrared sensor are used to collect images of tea leaves. The RGB sensor is used to capture the color changes, disease spots and visual features of insect bites on the surface of the tea leaves, and the infrared sensor is used to detect temperature changes and internal tissue damage caused by diseases and insect pests. The analog signals collected by the RGB sensor and the infrared sensor are then converted into digital signals through an analog-to-digital converter, and the damage features of the disease and insect pest areas are preferentially collected by dynamically adjusting the quantization parameters.

3. The tea plant pest and disease detection method based on deep learning according to claim 1, characterized in that: In step S5, the evaluation process specifically includes: Through quantitative analysis algorithms, the color changes of detected lesions, the degree of damage to leaf edges, and the temperature data in infrared images are quantitatively analyzed. The lesion area, shape, and depth data are compared with the corresponding data in the standard feature library to verify the reliability and consistency of the test results.

4. A computer device comprising a memory storing program instructions, characterized in that: When the program instructions are executed, the tea pest and disease detection method based on deep learning as described in any one of claims 1 to 3 is executed.

Images