Image recognition system for saline-alkali soil agricultural diseases and pests

Abstract

The invention relates to the technical field of image recognition, in particular to an image recognition system for saline-alkali soil agricultural diseases and pests, and the image recognition system comprises data collection, data annotation, feature extraction, model training, model verification and testing, deployment and application, result display and feedback, and continuous optimization. Through combination of multi-modal feature extraction and a deep learning technology, accuracy and environmental adaptability of disease and insect pest recognition are remarkably improved, in a feature extraction stage, the system simultaneously adopts a traditional method and deep features, local texture and global semantic information of disease and insect pests can be captured, and the method is suitable for the common problem of uneven illumination of saline-alkali soil. The system generates diversified samples through histogram equalization, color space conversion and data enhancement, enables a model to learn illumination invariant features, introduces a Part-based CNN and an attention mechanism for a shielding problem, focuses a target key region, and improves the utilization rate of local features in combination with multi-task learning.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and in particular to an image recognition system for agricultural pests and diseases in saline-alkali land. Background Art

[0002] Saline-alkali land refers to land with high levels of salt and alkalinity. This type of soil typically appears as a white or gray salt stain on the surface and has a high pH, ​​making it extremely unfavorable for plant growth. Saline-alkali land is primarily found in arid and semi-arid regions, as well as coastal areas and around inland salt lakes.

[0003] Saline-alkali agricultural pests and diseases refer to diseases and pests that occur in saline-alkali environments. Due to the high salinity and alkalinity of the soil, saline-alkali land has a significant impact on plant growth and also affects the occurrence and development of pests and diseases. Based on authorization announcement number CN113468984A, it discloses a crop pest and disease leaf recognition system, recognition method, and pest and disease prevention method, which belong to the field of image recognition. Problem. A crop pest and disease leaf recognition system, recognition method, and pest and disease prevention method use embedded image preprocessing technology to obtain information on the color and texture features of insect images to determine whether the leaves are likely to be infested by pests and diseases; collect images of crop pest and disease leafs, use the pytorch framework platform, combine the deep residual network and the Adam algorithm, and use the collected crop pest and disease leaf images to train a convolutional neural network; use the trained convolutional neural network to test a test set of images of leaves to be identified that may be infested by pests and diseases; classify the disease of the identified pest and disease leafs and take preventive measures. The present invention improves the intelligent recognition rate of crop pests and diseases.

[0004] Currently, in the process of agricultural pest and disease monitoring, image recognition problems include illumination changes, background changes, scale changes, perspective changes, occlusions, and noise interference. Since the main identification environment is saline-alkali land, the errors in the conclusions of conventional recognition due to background changes, scale changes, and perspective changes are relatively small. However, illumination changes, occlusions, and noise interference can still have a significant impact on the image recognition of pest and disease problems even with sufficient background environmental conditions. For example:

[0005] 1. Lighting changes: Images captured under different lighting conditions can have significant differences in color and brightness, affecting feature consistency.

[0006] 2. Occlusion: Part of the target object is blocked, resulting in incomplete features and affecting the recognition effect;

[0007] 3. Noise interference: Noise in the image will interfere with feature extraction, resulting in false detection or missed detection. Summary of the Invention

[0008] In view of the defects in the prior art, the present invention provides an image recognition system for agricultural pests and diseases in saline-alkali land, the image recognition system comprising:

[0009] S1. Data collection: Use high-resolution cameras, drones, or satellites to collect image data of saline-alkali land crops;

[0010] S2, data annotation: manual or semi-automatic annotation of the collected images;

[0011] S3. Feature extraction: Use traditional feature extraction methods or deep learning methods to extract the visual features of crop pests and diseases;

[0012] Based on traditional feature extraction methods:

[0013] f traditional =T(I)

[0014] in:

[0015] I is the input image;

[0016] T(·) represents the traditional feature extraction operator (such as the gradient direction histogram calculation of SIFT and HOG);

[0017] f traditional ∈R d Manually design feature vectors for the output d dimensions;

[0018] Based on deep feature extraction:

[0019]

[0020] in:

[0021] represents a deep neural network with parameters θ;

[0022] f deep ∈R k It is the k-dimensional deep feature output by a certain layer of the network;

[0023] S4. Model training: based on supervised learning or transfer learning;

[0024] S5. Model validation and testing: Evaluate model performance through cross-validation and independent test sets;

[0025] S6. Deployment application: Deploy the trained model to edge computing devices, cloud platforms or mobile applications;

[0026] S7. Result display and feedback: Output recognition results in a visual manner and provide user feedback mechanism for model optimization;

[0027] S8. Continuous optimization: Regularly update data sets and models, incorporating saline-alkali land environmental data.

[0028] Furthermore, the feature extraction includes conventional extraction and extraction and recognition problem processing. The extraction and recognition problem processing includes the influence of illumination changes, occlusion and obstruction, and noise interference. The feature extraction uses a convolutional neural network for automatic extraction.

[0029] Furthermore, the processing of the illumination change impact recognition includes image preprocessing, color space conversion, extraction of illumination invariant features, processing illumination changes based on convolutional neural networks, and feature fusion and optimization.

[0030] Furthermore, the convolutional neural network handles the problem of illumination changes:

[0031] Convolutional neural networks: Use images under different lighting conditions during training to enable the model to automatically learn lighting-invariant features;

[0032] Data augmentation: Randomly adjust brightness, contrast, and color temperature during training to improve model generalization capabilities.

[0033] Furthermore, the specific content of the convolutional neural network and data enhancement processing of illumination changes is as follows:

[0034] min θ E (x,y)-D |L(f θ (T(x)),y)|+λ·Ω(θ)

[0035] in:

[0036] Input transformation: T(x) = t aug (t light (x))

[0037] t aug (x) is the natural illumination variation, inherent in the dataset;

[0038] t light (x) is data augmentation, random bright color, contrast and color temperature edge ring;

[0039] Feature extraction: f θ (x) Automatically learn illumination-invariant features through CNN hierarchical structure;

[0040] The shallow convolution kernel response is h (1) =σ(W (1) *x+b (1) )

[0041] The deep feature is expressed as h (L) =φ(x;θ)

[0042] Optimization goal:

[0043] L(·) is the loss function of the classification or detection task;

[0044] Ω(θ) is the weight regularization term;

[0045] λ is the regularization function.

[0046] Furthermore, the processing of occlusion obstruction impact recognition includes data enhancement, partial feature matching, image segmentation, processing occlusion obstruction based on convolutional neural network, and prior knowledge fusion. Among them, for partial feature matching, the RANSAC algorithm is used for random sampling and consistency verification to restore the complete matching relationship from partial features, and local features are used to extract features that are robust to occlusion.

[0047] Furthermore, the convolutional neural network handles occlusion problems:

[0048] A partial convolutional neural network (CNN) is designed to capture the local features of the target object. An attention mechanism is introduced to enable the model to focus on the key areas of the target. Multi-task learning is combined to improve occlusion robustness.

[0049] Furthermore, the following content is used to handle occlusion based on Part-based CNN, attention mechanism, and multi-task learning:

[0050]

[0051] in:

[0052] For local feature extraction;

[0053] σ(Attn(f vis )) is the attention weight;

[0054] λ·∑ t∈{det，seg，rec} ω t L t For multi-task constraints;

[0055] Part-based CNN part:

[0056] K represents the preset number of components;

[0057] α k is the learnable weight of the kth component;

[0058] PartConv(·) uses partial convolution operation: I k *W k +b k , where I k is the component area obtained by ROI alignment;

[0059] Attention mechanism part:

[0060] σ(·) represents the Sigmoid activation function;

[0061] Attn(·) implements channel and spatial dual attention: Attn(f) = SpatialAttn(ChannelAttn(f));

[0062] f vis is the visibility score vector, which is calculated by 1*1 convolution;

[0063] Multi-task learning part:

[0064] L det : Detection task loss (such as Focal Loss);

[0065] L seg : Segmentation loss (Dice Loss);

[0066] L rec : Recognition loss (cross entropy);

[0067] wt is the dynamically adjusted task weight, which is automatically learned through uncertainty;

[0068] Hyperparameters:

[0069] λ controls the multi-task constraint strength;

[0070] ⊙ represents Hadamard product (element-wise multiplication);

[0071] B5. Prior knowledge integration:

[0072] Combine shape priors, such as shape recovery algorithms, to complete occluded areas;

[0073] Utilize structural priors, such as structure recovery algorithms, to reconstruct the complete structure of the target object.

[0074] Furthermore, the processing of image noise impact identification includes image denoising, processing noise interference based on convolutional neural network, extracting robust features, image preprocessing to enhance contrast, and multi-scale analysis.

[0075] Furthermore, the convolutional neural network processes noise interference:

[0076] Denoising autoencoder: trains the model to recover clean images from noisy images;

[0077] Convolutional Neural Networks: Add noise to training data to improve the robustness of feature extraction.

[0078] Innovations of this solution:

[0079] 1. Multimodal data fusion and adaptability to saline-alkali land

[0080] For the first time, saline-alkali land environmental data (soil moisture and salinity) are combined with visual features to improve the accuracy of pest and disease detection through multimodal data fusion, and to specifically address the particularities of saline-alkali land agricultural scenarios.

[0081] 2. Composite feature extraction framework

[0082] A collaborative extraction strategy of traditional features (SIFT, HOG) and deep features (CNN) is proposed, combined with PCA dimensionality reduction and feature fusion (such as SIFT+LBP+CNN) to enhance the robustness of the model to complex scenes.

[0083] 3. Innovation in illumination invariance technology

[0084] A cascade processing flow of input transformation + data enhancement + deep feature learning (such as HSV, Lab space conversion, SIFT, SURF illumination invariant features, and CNN automatic learning) is designed to systematically solve the problem of strong light changes in saline-alkali land.

[0085] 4. Multi-task dynamic learning of occlusion handling

[0086] The innovative integration of Part-based CNN, dual attention mechanism (channel + space) and multi-task learning (detection, segmentation, recognition) significantly improves the recognition rate in occluded scenes through dynamic weight adjustment and visibility scoring.

[0087] 5. Noise robustness layered processing

[0088] A three-layer anti-noise system is constructed, consisting of physical denoising (median / bilateral filtering) + deep learning (denoising autoencoder) + multi-scale analysis (wavelet, multi-scale CNN), among which noise adversarial training has methodological innovation.

[0089] The beneficial effects of the present invention are as follows: 1. By combining multimodal feature extraction with deep learning technology, the accuracy and environmental adaptability of pest and disease identification are significantly improved. In the feature extraction stage, the system uses both traditional methods and deep features to capture the local texture and global semantic information of pests and diseases. To address the common problem of uneven illumination in saline-alkali land, the system generates diversified samples through histogram equalization, color space conversion, and data enhancement, enabling the model to learn illumination-invariant features. To address occlusion issues, the system introduces Part-based CNN and attention mechanisms to focus on key target areas, and combines multi-task learning (detection, segmentation, and recognition) to improve the utilization of local features.

[0090] 2. As described in 1, by optimizing computing processes and edge deployment, the problem of limited equipment resources in saline-alkali land agricultural scenarios is solved. During the model training phase, transfer learning is used to reduce dependence on large-scale labeled data. To address noise interference, the system integrates a denoising autoencoder and a robust CNN. During the preprocessing phase, median filtering and non-local mean filtering are used to remove salt and pepper noise and Gaussian noise, and a multi-scale CNN is used to extract noise-resistant features.

[0091] 3. As described in 2, a multimodal analysis framework is constructed by combining environmental data and visual features, which significantly improves the comprehensiveness of pest and disease prediction. During the continuous optimization stage, the system regularly collects saline-alkali land environmental parameters, associates soil data with image features through feature fusion, and establishes a probability model for pest and disease occurrence. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly describes the drawings required for the specific embodiments or the description of the prior art. Similar elements or parts are generally identified by similar reference numerals throughout the drawings. Elements or parts in the drawings are not necessarily drawn to scale.

[0093] Figure 1 This is a system principle diagram of an image recognition system for agricultural pests and diseases in saline-alkali land according to the present invention;

[0094] Figure 2 A schematic diagram of illumination change processing in an image recognition system for agricultural pests and diseases in saline-alkali land according to the present invention;

[0095] Figure 3 A schematic diagram of occlusion and obstruction processing of an image recognition system for agricultural pests and diseases in saline-alkali land according to the present invention;

[0096] Figure 4 The figure is a schematic diagram of image denoising processing for an image recognition system for agricultural pests and diseases in saline-alkali land according to the present invention. DETAILED DESCRIPTION

[0097] The following embodiments of the technical solution of the present invention will be described in detail with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore only examples and are not intended to limit the scope of protection of the present invention.

[0098] It should be noted that, unless otherwise specified, the technical or scientific terms used in this application should have the common meanings understood by those skilled in the art to which the present invention belongs.

[0099] like Figure 1-Figure 4 As shown, an image recognition system for agricultural pests and diseases in saline-alkali land includes:

[0100] S1. Data collection: Use high-resolution cameras, drones, or satellites to collect image data of saline-alkali crops, and perform image preprocessing such as denoising, enhancement, and standardization.

[0101] S2. Data annotation: Manually or semi-automatically annotate the collected images, mark the types and locations of pests and diseases, and the characteristics of healthy crops, and construct a training dataset;

[0102] S3. Feature extraction: Use traditional feature extraction methods, such as SIFT and HOG, or deep learning methods, such as CNN, to extract visual features of crop pests and diseases;

[0103] Based on traditional feature extraction methods:

[0104] f traditional =T(I)

[0105] in:

[0106] I is the input image;

[0107] T(·) represents the traditional feature extraction operator (such as the gradient direction histogram calculation of SIFT and HOG);

[0108] f traditional ∈R d Manually design feature vectors for the output d dimensions;

[0109] Based on deep feature extraction:

[0110]

[0111] in:

[0112] represents a deep neural network with parameters θ;

[0113] f deep ∈R k It is the k-dimensional deep feature output by a certain layer of the network;

[0114] S4. Model training: Based on supervised learning or transfer learning, such as ResNet and VGG fine-tuning, train pest and disease classification or detection models and optimize model parameters;

[0115] In the training of pest and disease classification or detection models based on supervised learning or transfer learning, the common optimization goal is to minimize the loss function, which is analyzed using cross entropy loss. Specifically,

[0116]

[0117] in:

[0118] N is the sample size;

[0119] C is the total number of categories, such as the number of categories of pests and diseases;

[0120] y i，c is the true label of the i-th sample (1 if it belongs to category C, otherwise 0);

[0121] P i，c is the model's predicted probability that the i-th sample belongs to category C (Softmax output).

[0122] Optimization goal:

[0123] Adjust the model parameters θ through back propagation and gradient descent (such as Adam, SGD) to minimize the loss function, specifically:

[0124]

[0125] S5. Model validation and testing: Evaluate model performance through cross-validation and independent test sets, and calculate accuracy, recall, and F1 score indicators;

[0126] S6. Deployment and application: Deploy the trained model to edge computing devices, cloud platforms, or mobile applications to achieve real-time or remote pest and disease identification;

[0127] S7. Result display and feedback: Output recognition results in a visual manner and provide user feedback mechanism for model optimization;

[0128] S8. Continuous Optimization: Regularly update datasets and models, combine saline-alkali land environmental data, such as soil moisture and salinity, perform multimodal data fusion, and improve detection accuracy;

[0129] Feature extraction includes conventional extraction and extraction and recognition problem processing. The extraction and recognition problem processing includes the influence of lighting changes, occlusion and noise interference. Feature extraction uses convolutional neural networks for automatic extraction.

[0130] The processing for identifying the effects of illumination changes includes image preprocessing, color space conversion, extraction of illumination-invariant features, processing illumination changes based on convolutional neural networks, and feature fusion and optimization. Specifically:

[0131] A1. Image preprocessing:

[0132] Histogram equalization: enhances image contrast and reduces the impact of uneven lighting;

[0133] Gamma correction: adjusts image brightness to improve visual effects under different lighting conditions;

[0134] White balance adjustment: Correct color temperature and hue to keep colors consistent under different lighting conditions;

[0135] A2. Color space conversion:

[0136] Convert to HSV or Lab color space to separate brightness (V / L) and color information (H / S or a / b) to reduce the impact of lighting;

[0137] A3. Extracting illumination-invariant features:

[0138] Traditional feature extraction methods:

[0139] SIFT or SURF: Based on local gradients, robust to lighting changes;

[0140] ORB: combines FAST corner points and BRIEF descriptors, suitable for scenes with changing lighting;

[0141] Local binary pattern: extracts texture features and is insensitive to lighting;

[0142] Gray-level co-occurrence matrix: analyzes spatial gray-level relationships and reduces the impact of lighting;

[0143] A4. Processing illumination changes based on convolutional neural networks:

[0144] Convolutional neural networks: Use images under different lighting conditions during training to enable the model to automatically learn lighting-invariant features;

[0145] Data augmentation: Randomly adjust brightness, contrast, and color temperature during training to improve model generalization capabilities;

[0146] The specific content of processing illumination changes based on convolutional neural networks and data enhancement is as follows:

[0147] min θ E (x,y)-D |L(f θ (T(x)),y)|+λ·Ω(θ)

[0148] in:

[0149] Input transformation: T(x) = t aug (t light (x))

[0150] t aug (x) is the natural illumination variation, inherent in the dataset;

[0151] t light (x) is data augmentation, random bright color, contrast and color temperature edge ring;

[0152] Feature extraction: f θ (x) Automatically learn illumination-invariant features through CNN hierarchical structure;

[0153] The shallow convolution kernel response is h (1) =σ(W (1) *x+b (1) )

[0154] The deep feature is expressed as h (L) =φ(x;θ)

[0155] Optimization goal:

[0156] L(·) is the loss function of the classification or detection task;

[0157] Ω(θ) is the weight regularization term;

[0158] λ is the regularization function.

[0159] A5. Feature Fusion and Optimization:

[0160] Combining multiple features, such as SIFT+LBP+CNN features, to improve robustness;

[0161] Use PCA dimensionality reduction method to optimize features and reduce redundant information.

[0162] The processing of occlusion and obstruction impact recognition includes data enhancement, partial feature matching, image segmentation, occlusion and obstruction processing based on convolutional neural networks, and prior knowledge fusion. Specifically:

[0163] B1. Data enhancement:

[0164] Simulate random occlusions during training to generate diverse samples;

[0165] Use multi-view image training to enhance the model's generalization ability to occlusion;

[0166] B2. Partial feature matching:

[0167] The RANSAC algorithm is used for random sampling and consistency verification to restore the complete matching relationship from partial features;

[0168] Use local features, such as SIFT, SURF, and ORB, to extract features that are robust to occlusion;

[0169] B3. Image segmentation technology:

[0170] Apply semantic segmentation, such as FCN and U-Net, to separate the target from the background;

[0171] Use instance segmentation, such as Mask R-CNN, to distinguish different instances and reduce occlusion interference;

[0172] B4. Handling occlusion problems based on convolutional neural networks:

[0173] Design a Part-based CNN to capture the local features of the target object;

[0174] Introducing the attention mechanism to make the model focus on the key areas of the target;

[0175] Combine multi-task learning, such as detection, segmentation, and recognition, to improve occlusion robustness;

[0176] The following are the details of occlusion processing based on Part-based CNN, attention mechanism, and multi-task learning:

[0177]

[0178] in:

[0179] For local feature extraction;

[0180] σ(Attn(f vis )) is the attention weight;

[0181] λ·∑ t∈{det，seg，rec} ω t L t For multi-task constraints;

[0182] Part-based CNN part:

[0183] K represents the preset number of components;

[0184] α k is the learnable weight of the kth component;

[0185] PartConv(·) uses partial convolution operation: I k *W k +b k , where I k is the component area obtained by ROI alignment;

[0186] Attention mechanism part:

[0187] σ(·) represents the Sigmoid activation function;

[0188] Attn(·) implements channel and spatial dual attention: Attn(f) = SpatialAttn(ChannelAttn(f));

[0189] f vis is the visibility score vector, which is calculated by 1*1 convolution;

[0190] Multi-task learning part:

[0191] L det: Detection task loss (such as Focal Loss);

[0192] L seg : Segmentation loss (Dice Loss);

[0193] L rec : Recognition loss (cross entropy);

[0194] wt is the dynamically adjusted task weight, which is automatically learned through uncertainty;

[0195] Hyperparameters:

[0196] λ controls the multi-task constraint strength;

[0197] ⊙ represents Hadamard product (element-wise multiplication);

[0198] B5. Prior knowledge integration:

[0199] Combine shape priors, such as shape recovery algorithms, to complete occluded areas;

[0200] Utilize structural priors, such as structure recovery algorithms, to reconstruct the complete structure of the target object.

[0201] The processing for identifying the impact of image noise includes image denoising, noise interference processing based on convolutional neural networks, robust feature extraction, image preprocessing to enhance contrast, and multi-scale analysis. Specifically:

[0202] C1. Image denoising:

[0203] Median filter: suitable for removing salt and pepper noise;

[0204] Gaussian filtering: suitable for removing Gaussian noise;

[0205] Bilateral filtering: retains edge information while removing noise;

[0206] Non-local mean filtering: uses image self-similarity to remove noise and retain details;

[0207] C2. Processing noise interference based on convolutional neural network:

[0208] Denoising autoencoder: trains the model to recover clean images from noisy images;

[0209] Convolutional neural networks: add noise to training data to improve the robustness of feature extraction;

[0210] Noise processing based on denoising autoencoders and robust convolutional neural networks, the specific contents are as follows:

[0211]

[0212] in:

[0213] is the denoising autoencoder loss;

[0214] is the robust classification loss;

[0215] Key variables:

[0216] x is a clean image, (noised image,∈~N(0,σ 2 ));

[0217] f θ For the encoder, realize denoising and feature extraction;

[0218] is a classifier;

[0219] L cls is the classification loss (such as cross entropy);

[0220] λ is a hyperparameter that balances the two losses;

[0221] Specific process:

[0222] Noise layer:

[0223] Feature extraction:

[0224] Denoising and reconstruction:

[0225] Classification output:

[0226] C3. Extract robust features:

[0227] Scale-invariant feature transformation: based on gradient information, insensitive to noise;

[0228] Accelerated robust features: similar to SIFT, but with higher computational efficiency;

[0229] ORB: combines FAST corner points and BRIEF descriptors, suitable for noisy environments;

[0230] C4. Image preprocessing to enhance contrast:

[0231] Histogram equalization: adjust image contrast and reduce noise impact;

[0232] Gamma correction: optimizes brightness distribution and improves noise robustness;

[0233] C5. Multi-scale analysis:

[0234] Wavelet transform: decomposes the image into different frequency bands and performs targeted denoising;

[0235] Multi-scale CNN: extracts features at different scales to improve noise resistance.

[0236] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and description of the present invention.

Claims

1. An image recognition system for agricultural pests and diseases in saline-alkali land, characterized by: The image recognition system includes: S1. Data collection: Use high-resolution cameras, drones, or satellites to collect image data of saline-alkali land crops; S2, data annotation: manual or semi-automatic annotation of the collected images; S3. Feature extraction: Use traditional feature extraction methods or deep learning methods to extract the visual features of crop pests and diseases; Based on traditional feature extraction methods: f traditional =T(I) in: I is the input image; T(·) represents the traditional feature extraction operator (such as the gradient direction histogram calculation of SIFT and HOG); f traditional ∈R d Manually design feature vectors for the output d dimensions; Based on deep feature extraction: in: represents a deep neural network with parameters θ; f deep ∈R k It is the k-dimensional deep feature output by a certain layer of the network; S4. Model training: based on supervised learning or transfer learning; S5. Model validation and testing: Evaluate model performance through cross-validation and independent test sets; S6. Deployment application: Deploy the trained model to edge computing devices, cloud platforms or mobile applications; S7. Result display and feedback: Output recognition results in a visual manner and provide user feedback mechanism for model optimization; S8. Continuous optimization: Regularly update data sets and models, incorporating saline-alkali land environmental data.

2. The image recognition system for agricultural pests and diseases in saline-alkali land according to claim 1, characterized in that: The feature extraction includes conventional extraction and extraction and recognition problem processing. The extraction and recognition problem processing includes the influence of illumination changes, occlusion and obstruction, and noise interference. The feature extraction uses a convolutional neural network for automatic extraction.

3. The image recognition system for agricultural pests and diseases in saline-alkali land according to claim 2, characterized in that: The processing of illumination change impact recognition includes image preprocessing, color space conversion, extraction of illumination invariant features, processing of illumination changes based on convolutional neural networks, and feature fusion and optimization.

4. The image recognition system for agricultural pests and diseases in saline-alkali land according to claim 3, characterized in that: The convolutional neural network handles the problem of illumination variation: Convolutional neural networks: Use images under different lighting conditions during training to enable the model to automatically learn lighting-invariant features; Data augmentation: Randomly adjust brightness, contrast, and color temperature during training to improve model generalization capabilities.

5. The image recognition system for agricultural pests and diseases in saline-alkali land according to claim 4, characterized in that: The specific content of the convolutional neural network and data enhancement processing of illumination changes is as follows: minutes θ E (x,y)-D |L(f θ (T(x)),y)|+λ·Ω(θ) in: Input transformation: T(x) = t aug (t light (x)) t aug (x) is the natural illumination variation, inherent in the dataset; t light (x) is data augmentation, random bright color, contrast and color temperature edge ring; Feature extraction: f θ (x) Automatically learn illumination-invariant features through CNN hierarchical structure; The shallow convolution kernel response is h (1) =σ(W (1) *x+b (1) ) The deep feature is expressed as h (L) =φ(x;θ) Optimization goal: L(·) is the loss function of the classification or detection task; Ω(θ) is the weight regularization term; λ is the regularization function.

6. The image recognition system for agricultural pests and diseases in saline-alkali land according to claim 5, characterized in that: The processing of occlusion obstruction impact recognition includes data enhancement, partial feature matching, image segmentation, processing occlusion obstruction based on convolutional neural network, and prior knowledge fusion. For partial feature matching, the RANSAC algorithm is used for random sampling and consistency verification to restore the complete matching relationship from partial features, and local features are used to extract features that are robust to occlusion.

7. The image recognition system for agricultural pests and diseases in saline-alkali land according to claim 6, characterized in that: The convolutional neural network handles occlusion problems: A partial convolutional neural network (CNN) is designed to capture the local features of the target object. An attention mechanism is introduced to enable the model to focus on the key areas of the target. Multi-task learning is combined to improve occlusion robustness.

8. The image recognition system for agricultural pests and diseases in saline-alkali land according to claim 7, characterized in that: The content of occlusion processing based on Part-based CNN, attention mechanism and multi-task learning is as follows: in: For local feature extraction; σ(Attn(f vis )) is the attention weight; λ·∑ t∈{det，seg，rec} ω t L t For multi-task constraints; Part-basedCNN part: K represents the preset number of components; α k is the learnable weight of the kth component; PartConv(·) uses partial convolution operation: I k *W k +b k , where I k is the component area obtained by ROI alignment; Attention mechanism part: σ(·) represents the Sigmoid activation function; Attn(·) implements channel and spatial dual attention: Attn(f) = SpatialAttn(ChannelAttn(f)); f vis is the visibility score vector, which is calculated by 1*1 convolution; Multi-task learning part: L det : Detection task loss (such as FocalLoss); L seg : Segmentation loss (DiceLoss); L rec : Recognition loss (cross entropy); wt is the dynamically adjusted task weight, which is automatically learned through uncertainty; Hyperparameters: λ controls the multi-task constraint strength; ⊙ represents Hadamard product (element-wise multiplication); B5. Prior knowledge integration: Combine shape priors, such as shape recovery algorithms, to complete occluded areas; Utilize structural priors, such as structure recovery algorithms, to reconstruct the complete structure of the target object.

9. The image recognition system for agricultural pests and diseases in saline-alkali land according to claim 8, characterized in that: The processing of image noise impact identification includes image denoising, processing noise interference based on convolutional neural network, extracting robust features, image preprocessing to enhance contrast, and multi-scale analysis.

10. The image recognition system for agricultural pests and diseases in saline-alkali land according to claim 9, characterized in that: The convolutional neural network handles noise interference: Denoising autoencoder: trains the model to recover clean images from noisy images; Convolutional Neural Networks: Add noise to training data to improve the robustness of feature extraction.

Images