Animal and plant image recognition method based on deep learning and electronic equipment

Abstract

The invention relates to the technical field of image recognition, in particular to an animal and plant image recognition method based on deep learning and electronic equipment, and the method comprises the steps: obtaining and preprocessing sample image data, and constructing a sample image database; and constructing a data model, setting training parameters, and inputting image data and labels for training. And evaluating the model by using the verification set, and outputting after optimization. And evaluating the performance by using the test set, and deploying the model for intelligent identification. According to the invention, data diversity is increased by using a data enhancement technology, a model structure and a training strategy are optimized, and an under-fitting problem is solved. And by increasing the complexity of the neural network, selecting a proper weight initialization method and a proper learning rate strategy, and increasing training rounds, the generalization ability of the model is improved. Meanwhile, different attitude changes are simulated by using a data enhancement technology, multi-attitude image data are collected, an attention mechanism and an attitude estimation model are introduced, and differential recognition of animals and plants under different attitudes is realized by combining a multi-scale feature extraction and fusion strategy.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and particularly relates to a method for identifying plant and animal images based on deep learning and an electronic device. Background Art

[0002] In the field of plant and animal image recognition, a patented technical solution of an intelligent deep learning algorithm usually involves multiple key steps, including data preprocessing, model construction, feature extraction and classification, and post-processing and optimization. The following is the detailed content of this technical solution:

[0003] (1) Data Preprocessing

[0004] Image Acquisition:

[0005] Advanced devices such as high-definition cameras and infrared cameras are used to collect plant and animal images to ensure the clarity and integrity of the images. The collected images are accurately labeled to construct a label data set containing plant and animal category information.

[0006] Image Enhancement:

[0007] Preprocessing steps such as image denoising, enhancement, and cropping are performed to improve the image quality and reduce the difficulty of subsequent recognition tasks.

[0008] Image enhancement techniques such as contrast adjustment and brightness adjustment are used to highlight the features of plants and animals in the image.

[0009] Image Processing Fourier Transform:

[0010] The image processing problem is simplified through image transformation, facilitating the extraction of image features and thus conceptually deepening the understanding of image information.

[0011] Fourier transform formula in one-dimensional discrete form (DFT):

[0012]

[0013] Data Augmentation:

[0014] Through image transformation techniques such as rotation, scaling, flipping, and cropping, the diversity of training data is increased, and the generalization ability of the model is improved.

[0015] Data augmentation techniques are used to generate more training samples to alleviate the dependence of deep learning models on a large amount of data.

[0016] (2) Model Construction

[0017] Deep Learning Model Selection:

[0018] Select deep learning models suitable for plant and animal image recognition tasks, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), etc.

[0019] Customize and optimize the model according to specific application scenarios and requirements to improve recognition accuracy and efficiency.

[0020] Network architecture design:

[0021] Design a network architecture suitable for plant and animal image recognition, covering the input layer, convolutional layer, pooling layer, fully connected layer, etc.

[0022] Adopt advanced network architectures such as Residual Network (ResNet) and Dense Connected Network (DenseNet) to enhance the model's expressive ability and generalization ability.

[0023] Loss function and optimizer:

[0024] Select an appropriate loss function, such as cross-entropy loss, Focal Loss, etc., to evaluate the difference between the model's prediction results and the true labels.

[0025] As Figure 1 shown, the loss function (los funcion) is used to measure the inconsistency between the model's predicted value f(x) and the true value y. The smaller the loss function value, usually the stronger the robustness of the model. The loss function guides the learning process of the model.

[0026] Use an optimizer, such as Adam, SGD, etc., to update and optimize the model parameters to improve the training efficiency and performance of the model.

[0027] (3) Feature extraction and classification

[0028] Feature extraction:

[0029] Use a deep learning model to automatically extract features from images. These features may include visual information such as shape, texture, color, etc.

[0030] Adopt advanced techniques such as attention mechanisms and feature pyramids to improve the accuracy and robustness of feature extraction.

[0031] Classifier design:

[0032] Input the extracted features into a classifier, such as a Softmax classifier, support vector machine (SVM), etc., to achieve image classification and recognition.

[0033] Adopt ensemble learning methods, such as random forests, gradient boosting trees, etc., and combine the prediction results of multiple classifiers to improve the accuracy and stability of recognition.

[0034] (4) Post-processing and optimization

[0035] Result screening and verification:

[0036] Screen and verify the prediction results of the classifier, and eliminate the incorrect recognition results to improve the recognition accuracy.

[0037] Use algorithms such as non-maximum suppression (NMS) to merge and optimize the overlapping recognition results.

[0038] Model optimization and update:

[0039] According to new data and requirements, fine-tune or retrain the model to adapt to new application scenarios and plant and animal species.

[0040] Utilize technologies such as online learning and incremental learning to achieve continuous update and optimization of the model, so as to improve the real-time performance and accuracy of recognition.

[0041] In summary, the intelligent deep learning algorithm patent technical solutions for plant and animal image recognition cover multiple aspects such as data preprocessing, model construction, feature extraction and classification, and post-processing and optimization. These technical solutions cooperate with and promote each other, jointly improving the accuracy and efficiency of plant and animal image recognition.

[0042] However, there are still several limitations in the current technology:

[0043] Limitation 1:

[0044] Machine learning technology highly depends on a large number of precisely labeled training data sets. For the field of plant and animal image recognition, it is quite challenging to obtain high-quality labeled data sets. Given the wide variety and diverse forms of plants and animals, the labeling process not only requires relevant professional knowledge but also a large amount of time and labor.

[0045] Limitation 2:

[0046] The artificial intelligence learning model has deficiencies in learning and fitting the true distribution of training data, resulting in underfitting of the model, which in turn affects its performance on the training data set. In the task of plant and animal image recognition, if the model structure is too simplified or the training is insufficient, it may not be able to effectively capture the key features in the image, thus affecting the recognition accuracy.

[0047] Limitation 3:

[0048] The images of plants and animals in different poses show significant differences. For example, animals may exhibit different poses such as standing, walking, running, etc., while plants may be in different stages such as growth, flowering, and fruiting. These pose changes may pose difficulties for the accurate recognition of the model. Summary of the Invention

[0049] In view of the deficiencies of the prior art, the present invention discloses a method and an electronic device for identifying animal and plant images based on deep learning to solve the above problems.

[0050] The present invention is realized through the following technical solutions:

[0051] In a first aspect, the present invention provides a method for identifying animal and plant images based on deep learning, including the following steps:

[0052] Obtain sample image data, and perform preprocessing and data augmentation on the obtained sample image data to form a sample image database;

[0053] Select and construct a data model, set model training parameters, and input the image data and corresponding labels in the sample image database into the data model for training;

[0054] Use the validation set to evaluate the model, and optimize the data model according to the evaluation results to output a trained data model;

[0055] Use the test set to evaluate the performance of the data model, and deploy the data model to the actual application scenario to achieve intelligent recognition of the image data in the scenario.

[0056] Furthermore, the data preprocessing steps include:

[0057] Format conversion: Convert the images to a unified format;

[0058] Data annotation: Annotate the target objects or regions in the images to generate annotation data;

[0059] Data deduplication: Delete duplicate images;

[0060] Data filtering: Delete low-quality, non-standard, and unreasonable images;

[0061] Data repair: Repair the problems of missing, noise, and artifacts in the images.

[0062] Furthermore, the image data augmentation includes:

[0063] Rotation: Rotate the images by a set angle;

[0064] Translation: Perform a translation operation on the images in the image plane;

[0065] Scaling: Change the size of the images, including magnification and reduction;

[0066] Flipping: Flip the images horizontally or vertically;

[0067] Adding noise: Add random noise to the images, including Gaussian noise and salt-and-pepper noise;

[0068] Color transformation: Adjust the color attributes of brightness, contrast, and saturation of the image.

[0069] Furthermore, the methods for selecting a data model include the following:

[0070] Cross-validation method: Divide the original data set into a training set and a test set, and evaluate the performance of the model through multiple trainings and tests to select the best model;

[0071] Regularization method: Add a regularization term to the loss function to reduce the complexity of the model and prevent overfitting;

[0072] Bayesian method: Combine the prior probability and the posterior probability using Bayes' formula for model selection;

[0073] Information criterion: Select the best model by comparing the complexity and goodness of fit of different models;

[0074] The selection criteria for choosing a data model include:

[0075] Accuracy: Evaluate the accuracy rate of the model on the test set;

[0076] Training time: Select a model that can be trained within a reasonable time;

[0077] Model complexity: Select a model with moderate complexity;

[0078] Interpretability: According to the requirements of the application scenario, select a model with interpretability to better understand and apply the model.

[0079] Furthermore, model training includes the following steps:

[0080] Input data: Input the image data in the sample image database into the model;

[0081] Optimization algorithm: Select the gradient descent algorithm as the optimization algorithm, and update the parameters of the model iteratively to minimize the loss function;

[0082] Training monitoring: Monitor the changes in the loss function and the accuracy rate of the model to adjust the training strategy in a timely manner.

[0083] Furthermore, evaluating the model using a validation set includes:

[0084] Select evaluation metrics: Select evaluation metrics including accuracy rate, precision, or recall rate according to the requirements of the task;

[0085] Test set evaluation: Evaluate the trained model using the test set;

[0086] Error analysis: Conduct error analysis on the prediction results of the model to identify common error types for further model improvement.

[0087] Furthermore, optimizing the data model includes:

[0088] Parameter tuning: Adjust the parameters of the model according to the evaluation results, including the learning rate and regularization parameters, to further improve the performance of the model.

[0089] Feature engineering: Optimize feature engineering based on model feedback, including feature selection and feature transformation operations, to improve the generalization ability of the model.

[0090] Furthermore, deploy the data model to actual application scenarios including animal and plant recognition systems and agricultural intelligent monitoring systems.

[0091] Furthermore, clean the acquired sample image data to remove duplicate, blurred, and low-quality images; at the same time, label the acquired sample image data by assigning corresponding class labels to each image for subsequent training and testing.

[0092] In a second aspect, the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. The memory is coupled to the processor, and when the processor executes the computer program, the method for identifying animal and plant images based on deep learning described in the first aspect is implemented.

[0093] The beneficial effects of the present invention are as follows:

[0094] The present invention collects data sets through multiple channels, combines public data sets, professional databases, web crawler technology, and on-site shooting to achieve efficient and accurate data annotation. Employ professional annotators and use annotation tools for manual and semi-automatic annotation, and at the same time combine deep learning models for pre-annotation and manual correction to improve annotation accuracy.

[0095] The present invention uses data augmentation techniques to increase data diversity, optimize the model structure, and adjust the training strategy to solve the problem of model underfitting. Increase the complexity of the neural network, select appropriate weight initialization methods and learning rate strategies, and increase the number of training epochs to improve the generalization ability of the model.

[0096] The present invention uses data augmentation techniques to simulate different pose changes, collects multi-pose image data, introduces an attention mechanism and a pose estimation model, and combines multi-scale feature extraction and fusion strategies to achieve differential recognition of animals and plants in different poses. Description of the Drawings

[0097] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0098] Figure 1 is the loss function graph of the prior art of the present invention;

[0099] Figure 2 is the principle step graph of the method for identifying plant and animal images based on deep learning. Detailed implementation manners

[0100] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0101] In one embodiment, referring to Figure 2 as shown, a method for identifying plant and animal images based on deep learning is provided, including the following steps:

[0102] Obtain sample image data, and perform preprocessing and data augmentation on the obtained sample image data to form a sample image database;

[0103] Select and construct a data model, set model training parameters, and input the image data and corresponding labels in the sample image database into the data model for training;

[0104] Use a validation set to evaluate the model, and optimize the data model according to the evaluation results, and output the trained data model;

[0105] Use a test set to evaluate the performance of the data model, and deploy the data model to an actual application scenario to achieve intelligent recognition of the image data in the scenario.

[0106] In this embodiment, the collection of image data constitutes a crucial first step in the process of data preparation. This stage usually involves acquiring images from a variety of different sources, which may include the Internet, various databases, and local file systems. In order to ensure the quality of the data and the effectiveness of subsequent processing, the collected image data must be highly representative, which means that these images need to be able to comprehensively and multi-angle reflect the various features and attributes of the target object or scene. Only in this way can it be ensured that in the subsequent image processing and analysis process, useful information can be accurately identified and extracted, thereby providing a solid foundation for the training and verification of the machine learning model.

[0107] In this embodiment, the data collection work involves extensive collection of plant and animal image data from a variety of different sources. These sources include but are not limited to well-known search engines such as Baidu Images, and professional websites focusing on plant and animal information. In addition, in order to obtain more realistic and diverse images, field shooting was also carried out to ensure the richness and diversity of the data.

[0108] The data cleaning process is a crucial step, which involves the careful screening and processing of the collected image data. In this process, duplicate images, as well as those of poor quality and blurry images are eliminated. Through such cleaning work, it can ensure that the final data set used for analysis and training has high accuracy and reliability.

[0109] Data annotation is the process of associating image data with specific category labels. At this stage, each image is carefully labeled with the corresponding animal and plant category label. Such annotation work is crucial for the training and testing of machine learning models because it provides the necessary learning objectives and evaluation criteria for the model, allowing the model to accurately identify and classify different animal and plant images.

[0110] In one embodiment, image data preprocessing involves performing a series of processing steps on the initial image data, aiming to remove irrelevant information in the image, restore real and useful information in the image, and enhance the recognizability of relevant features.

[0111] In this embodiment, the image data preprocessing refers to performing a series of processing on the collected image data to improve the quality and applicability of the image.

[0112] The pre-processing steps in this embodiment include:

[0113] Format conversion: Convert images to a unified format, such as grayscale or RGB, for subsequent processing.

[0114] Data annotation: Annotate the target objects or regions in the image to generate annotation data. The annotation data is an important reference for model training and helps the model learn the features of the target objects.

[0115] Data deduplication: Delete duplicate images to reduce data redundancy and improve training efficiency. It is possible to determine whether images are duplicate by comparing their hash values or feature vectors.

[0116] Data filtering: Delete low-quality, non-standard, and unreasonable images. Low-quality images may include those with problems such as blurriness, high noise, and low contrast.

[0117] Data repair: Repair problems such as missing parts, noise, and artifacts in the image. Image processing algorithms can be used for operations such as image denoising, filling, and repairing missing regions.

[0118] In this embodiment, image data augmentation refers to increasing the quantity and diversity of data by performing a series of random transformations on the original image. Data augmentation can improve the generalization ability and robustness of the model.

[0119] The image data augmentation method in this embodiment includes:

[0120] Rotation: Rotate the image by a certain angle.

[0121] Translation: Perform a translation operation on the image in the image plane.

[0122] Scaling: Change the size of the image, including magnification and reduction.

[0123] Flipping: Flip the image horizontally or vertically.

[0124] Adding noise: Add random noise to the image, such as Gaussian noise, salt-and-pepper noise, etc.

[0125] Color transformation: Adjust color attributes such as brightness, contrast, and saturation of the image.

[0126] In one embodiment, model selection involves selecting the most suitable model from numerous candidate models in the context of specific dataset and task requirements. The quality of this process will directly affect the effectiveness of subsequent model training and the performance of the model in actual applications.

[0127] The selection method in this embodiment includes the following:

[0128] Cross-validation method: Divide the original dataset into a training set and a test set, and evaluate the performance of the model through multiple trainings and tests to select the best model. The cross-validation method can effectively avoid overfitting and underfitting problems.

[0129] Regularization method: By adding a regularization term to the loss function, the complexity of the model is reduced to prevent overfitting. It includes L1 regularization and L2 regularization.

[0130] Bayesian method: Using Bayes' formula to combine prior probability and posterior probability for model selection. The Bayesian method can consider the prior information of the model, thus more accurately selecting the model.

[0131] Information criteria: Such as Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), which select the best model by comparing the complexity and goodness of fit of different models.

[0132] The selection criteria of this embodiment are as follows:

[0133] Accuracy: Evaluate the accuracy rate of the model on the test set, that is, the proportion of correct predictions.

[0134] Training time: Consider the training time of the model and select a model that can be trained within a reasonable time.

[0135] Model complexity: Select a model with moderate complexity to avoid overfitting and underfitting.

[0136] Interpretability: According to the requirements of the application scenario, select a model with interpretability to better understand and apply the model.

[0137] In one embodiment, a specific model training is provided as follows:

[0138] Input data: Input the preprocessed data into the model.

[0139] Optimization algorithm: Select a suitable optimization algorithm, such as the gradient descent algorithm, and minimize the loss function by iteratively updating the parameters of the model.

[0140] Training monitoring: Monitor the training process of the model, including changes in the loss function, changes in accuracy, etc., in order to adjust the training strategy in a timely manner.

[0141] During the training process of the present invention, the model is regularly evaluated using the validation set to monitor the performance changes of the model; according to the evaluation results, the model is optimized, such as adjusting the learning rate, increasing the regularization term, using data augmentation techniques, etc., to improve the generalization ability and recognition accuracy of the model.

[0142] The model evaluation of this embodiment is as follows:

[0143] Select evaluation metrics: Select appropriate evaluation metrics according to the requirements of the task, such as accuracy rate, precision, recall rate, etc.

[0144] Test set evaluation: Use the test set to evaluate the trained model to understand the performance of the model on unseen data.

[0145] Error analysis: Conduct error analysis on the prediction results of the model to identify common error types for further model improvement.

[0146] The model tuning in this embodiment is as follows:

[0147] Parameter tuning: Adjust the parameters of the model, such as the learning rate, regularization parameter, etc., according to the evaluation results to further improve the performance of the model.

[0148] Feature engineering: Optimize feature engineering based on model feedback, including operations such as feature selection and feature transformation, to improve the generalization ability of the model.

[0149] The present invention uses the test set to test the trained model to evaluate the performance of the model; deploys the trained model to actual application scenarios, such as the animal and plant recognition system, the agricultural intelligent monitoring system, etc.; with the continuous emergence of new animal and plant image data and the continuous development of technology, it is necessary to update and iterate the model to improve its recognition ability and adaptability.

[0150] In one embodiment, the process of model construction includes multiple steps such as data preparation and model design.

[0151] The data preparation is as follows:

[0152] Data collection: Collect task-related data from various sources to ensure the diversity and representativeness of the data.

[0153] Data cleaning: Process missing values, outliers, and duplicate values in the data to ensure the accuracy and consistency of the data.

[0154] Data preprocessing: Perform operations such as normalization and standardization on the data to improve the training efficiency and performance of the model.

[0155] The model design is as follows:

[0156] Select the model architecture: According to the characteristics of the task and data, select a suitable model architecture, such as a multi-layer perceptron (MLP), a convolutional neural network (CNN), a recurrent neural network (RNN), etc.

[0157] Design the model parameters: Determine the hyperparameters of the model, such as the learning rate, regularization parameter, etc., which have an important impact on the performance of the model.

[0158] The model deployment is as follows:

[0159] Select a suitable technology and platform: Select a suitable technology and platform to deploy the model according to the requirements of the application scenario.

[0160] Performance monitoring: Establish a monitoring system to track the performance of the model and ensure the stability and accuracy of the model in actual applications.

[0161] Model maintenance: Regularly evaluate and update the model to adapt to new data and maintain the performance of the model.

[0162] In one embodiment, the present invention is applied to the test of the animal and plant image recognition model as follows:

[0163] Test purpose:

[0164] Verify the accuracy and robustness of the model in the animal and plant image recognition task.

[0165] Discover potential problems and deficiencies in the model during the recognition process for optimization and improvement.

[0166] Test dataset:

[0167] Collect an image dataset containing various animal and plant categories to ensure the diversity and representativeness of the dataset.

[0168] Preprocess the dataset, including operations such as image scaling and normalization, to improve the recognition effect of the model.

[0169] Test method:

[0170] Cross-validation: Divide the dataset into multiple subsets and test the model using cross-validation to evaluate the stability and generalization ability of the model.

[0171] Accuracy evaluation: Calculate the accuracy of the model on the test dataset to evaluate the recognition performance of the model.

[0172] Robustness test: Test the recognition performance of the model under different conditions such as lighting, angle, and occlusion to evaluate the anti-interference ability of the model.

[0173] Analysis of test results:

[0174] Statistically analyze the test results and calculate indicators such as the accuracy and recall rate of the model.

[0175] Analyze the error types and causes in the model during the recognition process and propose targeted optimization suggestions.

[0176] In one embodiment, the present invention is applied to the application of the animal and plant image recognition model as follows:

[0177] Agriculture: Utilize animal and plant image recognition technology to achieve pest and disease detection and growth status monitoring of crops, improving agricultural production efficiency and quality.

[0178] Forestry: By identifying tree species, growth conditions, etc., effective management and protection of forest resources are achieved.

[0179] Ecological protection: Using image recognition technology to monitor the population quantity and distribution of animals and plants, providing data support for ecological protection.

[0180] Education: In biology teaching, using animal and plant image recognition technology to assist students in identifying and understanding animals and plants, improving the teaching effect.

[0181] In this embodiment, the animal and plant image data can be obtained from multiple channels, such as professional databases, online image platforms, research institutions, etc. Ensure that the data is representative, covering animal and plant images under different species, different postures, and different lighting conditions.

[0182] In this embodiment, CNN is one of the most commonly used image feature extraction models in the field of deep learning. Through structures such as convolutional layers, pooling layers, and fully connected layers, high-dimensional features of images are extracted layer by layer to achieve accurate recognition of animals and plants.

[0183] In this embodiment, the model is trained using the labeled image data, and the model parameters are optimized through the backpropagation algorithm, enabling the model to accurately recognize animal and plant images. During the training process, issues such as overfitting and underfitting need to be concerned about and corresponding measures are taken to solve them.

[0184] In this embodiment, the recognition results of the model are verified to ensure the accuracy and reliability of the recognition results. This can be achieved through methods such as manual review and multi-model fusion.

[0185] In this embodiment, with the discovery of new animals and plants and the increase in image data, the model needs to be continuously updated and optimized to adapt to new recognition tasks and data characteristics.

[0186] In one embodiment, the present invention provides an electronic device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. The memory is coupled to the processor, and when the processor executes the computer program, the method for identifying animal and plant images based on deep learning described in the first aspect is implemented.

[0187] The present invention uses publicly available datasets, such as the plant and animal image datasets on platforms like Kaggle. These datasets have been preliminarily sorted and annotated and are suitable as training data for deep learning models. High-quality plant and animal image data is obtained by accessing professional databases, such as biodiversity databases and herbariums. The data in these databases usually has high accuracy and representativeness. Image data is crawled from relevant websites (such as animal and plant protection organizations, research institutions, etc.) using web crawler technology. However, relevant laws, regulations, and the usage agreements of the websites need to be complied with. For specific species or plants and animals in specific environments, image data can be obtained through on-site shooting. This method can ensure the authenticity and diversity of the data.

[0188] Hire professional annotators to manually annotate the collected image data. The annotation content includes the species of plants and animals, key features, etc. Manual annotation has high accuracy, but the cost is relatively high. Use annotation tools (such as LabelImg, VIA, etc.) for semi-automatic annotation. These tools can help annotators quickly and accurately annotate image data. At the same time, deep learning models can also be combined for pre-annotation, and then manually corrected and improved. In the case of a large amount of data and relatively simple annotation tasks, automated annotation methods can be tried. For example, use existing deep learning models to initially classify and annotate images, and then combine manual review and correction to improve the accuracy of annotation. However, it should be noted that the accuracy of automated annotation may not be as high as that of manual annotation.

[0189] The present invention performs operations on images such as rotation, scaling, translation, flipping, and cropping to increase the diversity of the data. Using data augmentation techniques can simulate the changes of data in the real world, such as lighting changes and occlusions, which helps to improve the generalization ability of the model. Collect plant and animal image data from more channels, including professional databases, online image platforms, research institutions, etc. Take on-site photos of specific species or plants and animals in specific environments to supplement the diversity of the dataset.

[0190] Increase the number of hidden layers and hidden units of the neural network to make the model have stronger fitting ability. For convolutional neural networks (CNNs), the number and depth of convolutional layers can be increased, and larger convolutional kernels can be used. Try using more complex deep learning models, such as ResNet, VGG, etc., which have stronger feature extraction and classification abilities.

[0191] Select appropriate weight initialization methods, such as He initialization, Xavier initialization, etc., to avoid the problems of gradient vanishing or gradient explosion. Use a smaller learning rate for training to prevent the model from converging to a local optimum prematurely during training. You can try using learning rate decay strategies, such as exponential decay, cosine decay, etc., to gradually reduce the learning rate. Increase the number of training epochs of the model to allow the model sufficient time to learn the complex patterns in the data.

[0192] The present invention performs operations such as rotation, scaling, and translation on images to simulate the pose changes of animals and plants from different perspectives. Use data augmentation techniques to generate more animal and plant images with different poses to increase the diversity of the dataset. Collect animal and plant images with different poses from multiple channels, including professional databases, online image platforms, research institutions, etc. Take on-site photos of animals and plants with different poses to supplement the diversity of the dataset.

[0193] The attention mechanism can help the model focus on important regions in the image and improve the recognition performance of animals and plants with different poses. Introduce attention modules, such as the SE (Squeeze-and-Excitation) module, CBAM (Convolutional Block Attention Module), etc., into the CNN model.

[0194] The pose estimation model can predict the pose information of animals and plants, such as joint positions, pose angles, etc. Combine the pose estimation model with the animal and plant recognition model and use the pose information to assist in recognition.

[0195] Use convolutional kernels and pooling layers of different scales to extract multi-scale features of the image. Fuse the multi-scale features to improve the model's recognition ability for animals and plants with different poses. Pay attention to the local features of animals and plants, such as the heads, limbs, tails of animals, and the leaves, flowers, fruits of plants. Use local feature extraction methods, such as ROI (Region of Interest) pooling, local convolution, etc., to extract the features of these key regions. Fuse the global features and local features to improve the recognition performance of the model. Strategies such as concatenation, weighted summation, etc., can be used for feature fusion.

[0196] In summary, the present invention collects the dataset through multiple channels, combines public datasets, professional databases, web crawler technology, and on-site shooting to achieve efficient and accurate data annotation. Hire professional annotators and use annotation tools for manual and semi-automatic annotation, and at the same time combine deep learning models for pre-annotation and manual correction to improve the annotation accuracy.

[0197] The present invention increases data diversity through data augmentation techniques, optimizes the model structure, and adjusts the training strategy to solve the problem of model underfitting. The complexity of the neural network is increased, a suitable weight initialization method and learning rate strategy are selected, and the number of training epochs is increased to improve the generalization ability of the model.

[0198] The present invention uses data augmentation techniques to simulate different pose changes, collects multi-pose image data, introduces an attention mechanism and a pose estimation model, and combines multi-scale feature extraction and fusion strategies to achieve differential recognition of animals and plants in different poses.

[0199] 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 foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for animal and plant image recognition based on deep learning, characterized in that: The following steps are involved: Acquire sample image data, and perform preprocessing and data enhancement on the acquired sample image data to form a sample image database; Select and build a data model, set model training parameters, and input image data and corresponding labels from the sample image database into the data model for training; Use the validation set to evaluate the model, and optimize the data model based on the evaluation results, and output the trained data model; Use the test set to evaluate the performance of the data model, deploy the data model to actual application scenarios, and realize intelligent recognition of image data in the scene.

2. The method for recognizing plant and animal images based on deep learning according to claim 1, characterized in that: In the method, the data preprocessing step includes: Format conversion: convert images into a unified format; Data annotation: Annotate the target objects or areas in the image to generate annotation data; Data deduplication: remove duplicate images; Data filtering: remove low-quality, non-standard and unreasonable images; Data restoration: Repair missing images, noise, and artifacts.

3. The method for animal and plant image recognition based on deep learning according to claim 1, characterized in that: In the method, image data enhancement includes: Rotate: rotate the image at a set angle; Translation: translate the image on the image plane; Scaling: changing the size of an image, including zooming in and out; Flip: flip the image horizontally or vertically; Add noise: Add random noise to the image, including Gaussian noise and salt and pepper noise; Color Transformation: Adjust the color properties of an image such as brightness, contrast, and saturation.

4. The method for recognizing plant and animal images based on deep learning according to claim 1, characterized in that: In the method, the method of selecting a data model includes the following: Cross-validation method: divide the original data set into a training set and a test set, evaluate the performance of the model through multiple training and testing, and select the best model; Regularization method: By adding regularization terms to the loss function, the complexity of the model is reduced to prevent overfitting; Bayesian method: Use the Bayesian formula to combine prior probability and posterior probability for model selection; Information criterion: Select the best model by comparing the complexity and fit of different models; The selection criteria for choosing a data model include: Accuracy: Evaluate the accuracy of the model on the test set; Training time: Choose a model that can be trained in a reasonable amount of time; Model complexity: Choose a model with moderate complexity; Interpretability: According to the requirements of the application scenario, select a model with interpretability to better understand and apply the model.

5. The method for animal and plant image recognition based on deep learning according to claim 1, characterized in that: In the method, model training includes the following steps: Input data: Input image data from the sample image database into the model; Optimization algorithm: Select the gradient descent optimization algorithm to minimize the loss function by iteratively updating the model parameters; Training monitoring: Monitor changes in the model’s loss function and accuracy so that the training strategy can be adjusted in a timely manner.

6. The method for animal and plant image recognition based on deep learning according to claim 1, characterized in that: In the method, evaluating the model using a validation set includes: Select evaluation metrics: Select evaluation metrics including accuracy, precision or recall according to the requirements of the task; Test set evaluation: Use the test set to evaluate the trained model; Error Analysis: Perform error analysis on the model’s prediction results to identify common types of errors so that the model can be further improved.

7. The method for recognizing plant and animal images based on deep learning according to claim 1, characterized in that: In the method, optimizing the data model includes: Parameter tuning: Adjust the model parameters, including learning rate and regularization parameters, based on the evaluation results to further improve the performance of the model. Feature Engineering: Optimize feature engineering based on model feedback, including feature selection and feature transformation operations to improve the generalization ability of the model.

8. The method for recognizing plant and animal images based on deep learning according to claim 1, characterized in that: In the method, the data model is deployed to actual application scenarios including animal and plant identification systems and agricultural intelligent monitoring systems.

9. The method for recognizing plant and animal images based on deep learning according to claim 1, characterized in that: In the method, the acquired sample image data is cleaned to remove duplicate, blurred, and low-quality images; at the same time, the acquired sample image data is labeled to give each image a corresponding category label for subsequent training and testing.

10. An electronic device, characterized in that: It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, the memory is coupled to the processor, and when the processor executes the computer program, the deep learning-based plant and animal image recognition method as described in any one of claims 1 to 9 is implemented.

Images