A training method, model, system, and application of an improved Inception-v3 network for morel mushroom image recognition are presented.

Abstract

This invention proposes a CNN training method of the Inception-v3 type, characterized by the following steps: S1 Preprocessing of sample images, the preprocessing including combined data augmentation and normalization, to obtain preprocessed sample images; S2 Feature extraction, the preprocessed images are fed into the CNN, where the parameters of the first few layers are frozen, while the parameters of subsequent layers and the classification head remain normal, and the main classification head and auxiliary branch classification heads output the first and second classification results respectively; S3 Parameter update, the parameters of the subsequent layers and the classification head are updated using a joint cross-entropy loss function on the first and second classification results, where the total loss in the joint cross-entropy loss function is a linear combination of the main loss and the weighted auxiliary loss; S4 Parameter optimization, the parameters are optimized using an adaptive first-order optimizer with weight decay, and the learning rate decay is triggered based on the validation set loss; S5 Training termination, in mixed precision mode, the validation set index is used as a signal for achieving the training objective, and the optimal parameter file at the time of training achievement is persistently saved. Compared with the prior art, the model trained by this invention significantly improves accuracy and efficiency.

Description

Technical Field

[0001] This invention belongs to the field of image recognition and artificial intelligence technology, specifically relating to a method, system and storage medium for constructing a morel mushroom image recognition model based on an improved Inception-v3 network. Background Technology

[0002] Morel mushrooms (Morchella spp., hereinafter referred to as "morels") are a type of edible fungus with high economic value. They possess a complex morphology and structure, exhibiting diverse and highly irregular surface textures. Subtle morphological differences exist between different species and strains, making high-precision differentiation difficult using traditional image recognition methods. Furthermore, the phenotypic appearance of morels varies significantly across different growth stages, light conditions, and cultivation environments, further complicating the image recognition task and posing a greater challenge.

[0003] Currently, general-purpose convolutional neural networks (CNNs), such as Inception-v3, ResNet, and VGG, while performing excellently on large-scale natural image datasets (such as ImageNet), generally encounter the following problems when directly applied to morel mushroom image recognition tasks in practical use:

[0004] 1. Small sample sensitivity and overfitting risk: Due to the scarcity of high-quality labeled morel mushroom image data, the model is prone to overfitting under limited training samples, resulting in decreased generalization ability.

[0005] 2. Insufficient fine-grained feature extraction capability: Existing networks have limited ability to model discriminative local features such as micro-texture, edge contours and wrinkled structures on the cap surface, resulting in high inter-class confusion rates;

[0006] 3. Poor inference efficiency and deployment adaptability: The model has a large number of parameters and high computational complexity, making it difficult to meet the real-time requirements of low latency and low power consumption inference for field edge devices (such as embedded vision terminals and mobile sorting devices).

[0007] To address the challenges of overfitting in small sample sizes, difficulty in representing fine-grained textures, and an imbalance between training and deployment efficiency in morel mushroom image recognition, this invention proposes a training and inference method based on a multi-branch convolutional backbone (Inception-type), combined with auxiliary branch joint supervision, hierarchical freezing, data augmentation, AMP, and adaptive learning rate scheduling. Based on this method, an improved neural network optimized for morel mushroom image recognition is constructed. This improves recognition accuracy and model robustness while balancing training efficiency, parameter compression, and edge deployment performance, thus supporting automated sorting and quality grading applications in smart agriculture scenarios. Summary of the Invention

[0008] This invention proposes a CNN training method of the Inception-v3 type, characterized by the following steps:

[0009] S1 preprocesses the sample image, the preprocessing including combined data augmentation and normalization, to obtain a preprocessed image of the sample;

[0010] S2 Feature Extraction: The preprocessed image is fed into a CNN, where the parameters of the first few layers are frozen, while the parameters of subsequent layers and the classification head remain normal. The first and second classification results are output by the main classification head and the auxiliary branch classification head, respectively.

[0011] S3 parameter update: For the first and second classification results, the parameters of the subsequent layers and the classification head layer are updated using the joint cross-entropy loss function. The total loss in the joint cross-entropy loss function is a linear combination of the main loss and the weighted auxiliary loss.

[0012] S4 parameter optimization uses an adaptive first-order optimizer with weight decay to optimize the parameters and triggers learning rate decay based on validation set loss.

[0013] S5 training terminates, using validation set metrics as a signal that the training objective has been achieved in mixed precision mode, and persistently saves the optimal parameter file at the time of training achievement.

[0014] The method described above is characterized in that the data augmentation includes at least one or more of the following operations: random scale cropping, preferably cropping to 299×299 pixels;

[0015] Random horizontal and / or vertical flipping;

[0016] Random rotation, preferably with a rotation angle not exceeding 45°;

[0017] Random affine translation, preferably with an affine translation amplitude not exceeding 15%;

[0018] Random perspective perturbation, preferably with an amplitude not exceeding 0.2;

[0019] Color fluctuation;

[0020] Gaussian blur and / or random erasure; preferably, the area of ​​random erasure accounts for 1%-10%.

[0021] The method described above is characterized in that the standardization process uses channel mean and variance to process the data-enhanced image.

[0022] The method described above is characterized in that the freezing of the first few layers of parameters is to freeze any parameters of the first 120-220 layers in parameter order; preferably, it is to freeze any parameters of the first 150 layers.

[0023] The method described above is characterized in that the Inception-v3 type CNN contains multi-branch convolutional units.

[0024] The method described in any of the above methods is characterized by:

[0025] The main classification head includes the following structure: Dropout → Fully connected layer (2048→512) → ReLU activation layer → BN layer → Dropout → Fully connected layer (512→C).

[0026] The method described in any of the above methods is characterized by:

[0027] The auxiliary branch classification head includes the following structure: Dropout → Fully Connected Layer (in_features→C).

[0028] The method described in any of the above methods is characterized in that the joint cross-entropy loss function is defined as follows:

[0029] L = Lmain + αLaux, where Lmain is the main loss, Laux is the auxiliary loss, and α is the weighting coefficient, preferably 0.2-0.6, and more preferably 0.4.

[0030] The method described above is characterized in that the adaptive first-order optimizer is the AdamW optimizer, which combines the learning rate scheduling strategy of ReduceLROnPlateau and dynamically decays the learning rate based on the validation set loss.

[0031] The method described above is characterized in that the initial learning rate of the AdamW optimizer is 1×10⁻⁶. -5 -5×10 -4 The preferred value is 1×10 -4 The initial value of the weight decay coefficient is 1×10. -6 -1×10 -3 The preferred value is 1×10 -4 .

[0032] The method described above is characterized in that the attenuation factor of ReduceLROnPlateau is adjusted in the range of 0.3–0.7, preferably 0.5; and the attenuation value is adjusted in the range of 2–6, preferably 3.

[0033] The method described above is characterized in that the mixed precision mode is an AMP mixed precision mechanism; at the same time, a dual-index early stopping mechanism is used to verify the loss and accuracy, and training is terminated when the verification performance no longer improves within a set range.

[0034] The method described above is characterized in that,

[0035] In the aforementioned AMP hybrid precision mechanism, the tolerance is δ≈1×10⁻⁶. -3 Patience p≈10.

[0036] During training, an AMP hybrid precision mechanism (GradScaler + autocast) is employed to improve throughput and memory utilization. Simultaneously, a dual-metric early stopping mechanism is used, combining validation loss and accuracy, with training epochs controlled by setting tolerance and patience. Training terminates when validation performance no longer improves within a set range; and the optimal weight file is automatically persisted and saved when performance improves.

[0037] After training, the model is evaluated using validation and test sets: multi-dimensional evaluation metrics are calculated on the validation and test sets, including accuracy, recall, F1 score, weighted multi-class AUC-ROC, log loss, and confusion matrix.

[0038] After evaluation, save the complete optimal weight file, category mapping relationship and complete training log to support subsequent model inference and deployment.

[0039] Compared with existing technologies, the model trained by this invention has the following significant technical effects in practical applications:

[0040] 1. Accuracy: On the three-class classification task (M. eximia / M. importuna / M. sextelata), the validation set accuracy was 95.31%, which is an improvement of +1.56 percentage points compared to the original model's 93.75%;

[0041] 2. Convergence: Optimal result achieved in 24 epochs, reducing the number of epochs by 50% compared to the original 48 epochs;

[0042] 3. Efficiency: The total training time decreased from 23.75 min to 9.47 min, resulting in an efficiency improvement of approximately 60.1%;

[0043] 4. Robustness: More robust to pose, lighting, and local occlusion;

[0044] 5. Deployability: The model size and computational overhead are moderate, and it can achieve edge-accelerated inference when used with ONNX / TensorRT and other technologies. Attached Figure Description

[0045] Figure 1 Morel mushroom image recognition model training and deployment application process

[0046] Figure 2 Training process curve (loss / accuracy changes with epochs)

[0047] Figure 3 Bar chart comparing the original and improved models in terms of training time, convergence speed, and accuracy.

[0048] Figure 4 Morel mushroom image prediction visualization (ground truth / prediction and confidence level display) Detailed Implementation

[0049] Example 1

[0050] The improved Inception-v3 CNN model was trained based on images of three morel mushroom categories.

[0051] The training steps are as follows:

[0052] Organization and loading of the S1 dataset

[0053] The dataset was constructed by collecting images of any three species of morel mushrooms, totaling approximately 4,000 images, with about 1,300 images for each morel mushroom species. The images were then divided into a training set and a test set in a 7:3 ratio, with approximately 15% of the training set further divided into a validation set.

[0054] Morel mushroom images are organized and stored according to category directories; batch reading is performed using DataLoader with batch_size=32, num_workers=4, and pin_memory=True to improve loading efficiency.

[0055] S2 Preprocessing

[0056] First, combined data augmentation is performed during the input phase. Data augmentation involves manipulating the images in the dataset using one or more of the following methods:

[0057] Random scale cropping, preferably cropped to 299×299 pixels;

[0058] Random horizontal and vertical flipping;

[0059] Random rotation, preferably with an angle not exceeding 45°).

[0060] Random affine translation, preferably with an amplitude not exceeding 15%;

[0061] Random perspective perturbation, preferably with an amplitude not exceeding 0.2;

[0062] Color fluctuation;

[0063] Optional Gaussian blur and random erase.

[0064] After performing any one or more of the above image enhancement operations, the enhanced image is further standardized based on the channel mean and variance.

[0065] After geometric translation, flipping, rotation, coloring, occlusion and other operations, the diversity of the dataset is enriched to the maximum extent, so as to simulate real morel mushroom images under different forms, improve the training efficiency and recognition accuracy of the model, and significantly enhance the generalization ability of the model.

[0066] Enhanced standardization ensures that the input distribution matches the pre-trained weights.

[0067] In this embodiment, the illustrative dataset random augmentation and normalization transformation operations based on the PyTorch framework are as follows:

[0068] RandomResizedCrop(299, scale=(0.8,1.0));

[0069] Operation Instructions: Randomly crop a region of the image and resize it to a fixed size of 299x299 pixels. The scale=(0.8, 1.0) parameter specifies the ratio of the cropped area to the original image area, i.e., 80% to 100%.

[0070] RandomHorizontalFlip(p=0.5) and RandomVerticalFlip(0.3);

[0071] Instructions: Flip the image horizontally with a 50% probability and flip the image vertically with a 30% probability.

[0072] Random Rotation (45°);

[0073] Operation instructions: Randomly rotate the image, with a maximum rotation angle of ±45 degrees.

[0074] ColorJitter(brightness=0.3, contrast=0.3, saturation=0.3, hue=0.1);

[0075] Instructions: Randomly change the image's brightness (±30%), contrast (±30%), saturation (±30%), and hue (±10%).

[0076] RandomAffine(degrees=0, translate=(0.15,0.15));

[0077] Instructions: Apply affine transformation. degrees=0 means no rotation, and translate=(0.15, 0.15) means translate the image size by a maximum of 15% in the X and Y directions.

[0078] RandomPerspective (distortion_scale=0.2, p=0.5);

[0079] Instructions: Apply perspective transformation with a 50% probability. Distortion_scale=0.2 controls the intensity of the perspective transformation.

[0080] RandomApply([GaussianBlur(kernel_size=(3,5))], p=0.3);

[0081] Instructions: Apply Gaussian blur with a 30% probability, and randomly select the blur kernel size as 3x3 or 5x5 pixels.

[0082] ToTensor();

[0083] Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]);

[0084] Instructions: Standardize the image by using the standard mean and standard deviation of the ImageNet pre-trained model to match the input data distribution with the pre-trained weights.

[0085] RandomErasing(p=0.5, scale=(0.02,0.1)).

[0086] Instructions: Randomly erase a rectangular area from the image with a 50% probability. The erased area will be between 2% and 10% of the original image's total area.

[0087] Structural design of the S3 model

[0088] This embodiment uses the classic Inception-v3 network as the backbone feature extractor. Inception-v3's own multi-branch convolutional architecture (Inception modules) can capture features at different receptive fields and scales in parallel, which is crucial for processing morel mushroom images with complex morphology and diverse surface textures.

[0089] To fully utilize the general features learned by the pre-trained model on large-scale image data and to prevent overfitting on the limited morel mushroom dataset, a transfer learning strategy is used to freeze several layers with the highest parameter ranking in the Inception-v3 backbone network. The specific number of layers can be determined according to the actual model structure. In this embodiment, 150 layers are preferred, and parameter updates are only allowed for subsequent layers and newly designed classification heads.

[0090] To address the core issue of difficulty in distinguishing subtle morphological differences among morel species and to improve the stability and efficiency of model training, this embodiment designs a dual-output structure including a main classification head and an auxiliary classification head:

[0091] Auxiliary Classifier

[0092] Structure: The auxiliary classification head is placed after an intermediate feature layer in the Inception-v3 network (corresponding to the default position when aux_logits=True), and its structure is Dropout → fully connected layer (in_features→C). Here, in_features is the dimension of the flattened feature map of this intermediate layer, and C is the number of target categories.

[0093] The auxiliary classification head provides direct supervision signals to the front and middle parts of the network. During backpropagation, the gradients from this part can bypass the deep network and directly optimize the convolutional layers closer to the input, enabling these layers to learn more discriminative local features (such as the microtexture of the cap, edge contours, and fold structures), which is particularly crucial for distinguishing morel species that are extremely similar in appearance.

[0094] In deep networks, gradients can gradually diminish during backpropagation, leading to slow learning in preceding layers. The auxiliary classification head provides a "shortcut" for gradient flow, enhancing the overall trainability of the network, indirectly supporting effective learning of subtle features, effectively mitigating gradient vanishing, helping the model converge faster, and shortening training time.

[0095] The presence of the auxiliary classification head imposes additional constraints on the intermediate layer features, requiring the intermediate representation itself to have a certain classification ability. This helps prevent the model from overfitting to noisy or irrelevant features in the training set, thereby improving generalization ability and regularization effect.

[0096] Main Classifier

[0097] Structure: The main classification head receives the output of the global average pooling of the Inception-v3 backbone network (usually a 2048-dimensional feature vector). Its specific design is Dropout → fully connected layer (2048→512) → ReLU activation layer → BN layer → Dropout → fully connected layer (512→C).

[0098] Design considerations:

[0099] A nonlinear transformation layer and hidden space are introduced: High-dimensional features are projected into a 512-dimensional hidden representation space using FC(2048→512) + ReLU, and nonlinear activation is introduced. This increases the model's expressive power, enabling it to learn more complex feature combinations for distinguishing subtle differences, significantly improving the extraction and recognition of subtle feature differences, thus addressing the challenge of accurate identification caused by the small appearance differences among different morel species.

[0100] Batch Normalization (BN): A BN layer is added after the first fully connected layer and before the activation function to normalize the input, accelerate training convergence, and help improve the stability of the model.

[0101] Double Dropout: Two Dropout events are implemented in the architecture. The first occurs before the features are input to the 512-dimensional fully connected layer, and the second occurs after the 512-dimensional features undergo ReLU and BN. Double Dropout serves as an effective regularization technique, significantly improving the model's robustness in small sample sizes and reducing the risk of overfitting.

[0102] S4 Training

[0103] This embodiment employs the following training strategy and hyperparameter configuration to ensure model accuracy while also considering training efficiency, convergence stability, and generalization ability.

[0104] Loss function: The standard cross-entropy loss is used as the basic loss function. For the main output and auxiliary output, a joint loss function L = Lmain + αLaux is constructed.

[0105] The weighting coefficient α = 0.4.

[0106] This value has been verified through ablation experiments to effectively balance the contributions of backbone supervision and intermediate layer supervision in this task—too small a value results in weak auxiliary supervision, while too large a value may interfere with backbone feature learning. In practice, α can be adjusted within the range of [0.2, 0.6], and results superior to the baseline can be obtained.

[0107] Optimizer: AdamW (lr=1×10) was selected as the optimizer. -4 weight_decay=1×10 -4 ).

[0108] AdamW decouples weight decay and gradient update from Adam, which can more effectively control model complexity and suppress overfitting. It is particularly suitable for fine-tuning large pre-trained models in small sample scenarios.

[0109] Initial learning rate 1×10 -4 It is a common starting point in transfer learning, which avoids destroying pre-trained weights while ensuring effective learning; the weight decay coefficient is 1×10. -4 It provides appropriate regularization.

[0110] Other first-order adaptive optimizers (such as Adam, RMSprop) or momentum-driven SGD can also be used, but they usually require more refined learning rate scheduling and their convergence speed and stability are slightly inferior to AdamW.

[0111] Learning rate scheduling: ReduceLROnPlateau(factor=0.5, patience=3) is used, which means that if the validation set loss does not decrease significantly within 3 consecutive epochs, the learning rate is multiplied by 0.5. This "on-demand decay" mechanism is more adaptable to the actual training dynamics than fixed step decay, and helps the model escape local optima and converge more finely.

[0112] Other scheduling strategies (such as CosineAnnealing and StepLR) are also feasible, but they require manual adjustment of the period or step size based on the performance verification, resulting in lower flexibility.

[0113] Mixed-precision training: Enable the Automatic Mixed Precision (AMP) training mechanism, combining autocast and GradScaler, i.e., AMP (GradScaler + autocast). This technique significantly reduces memory usage and increases training throughput (as shown in the example, training time is reduced by approximately 60%) while maintaining almost the same model accuracy, which is especially important in resource-constrained environments. This is an optional but strongly recommended engineering optimization method.

[0114] Early stopping mechanism: An early stopping strategy based on validation loss is adopted (patience=10, delta=1×10). -3 That is, when the validation loss is within the tolerance of 1×10 -3 Training is terminated if there is no improvement after 10 consecutive rounds. This effectively prevents overfitting and automatically determines the optimal number of training rounds.

[0115] In practice, a dual-index early stop logic can be designed by combining verification accuracy to further improve robustness.

[0116] Model saving: After each epoch, if the validation performance (such as accuracy or loss) reaches its historical best, the current model weights are automatically saved. Ultimately, the model that performs best throughout the entire training process is retained, not the model from the last epoch.

[0117] The parameters listed in this embodiment are only illustrative of the training of the model in this invention and should not be construed as limiting the scope of protection of this invention.

[0118] S5 Evaluation and Persistence

[0119] After training, the model is comprehensively evaluated using a validation set and a separate test set. Specifically, this includes:

[0120] Calculate and output a classification report (including precision, recall, and F1 score for each category);

[0121] Generate a confusion matrix and analyze the misclassification among different morel species;

[0122] Calculate the weighted multi-class AUC-ROC and log loss to comprehensively measure the model's discriminative ability and prediction confidence.

[0123] At the same time, the system automatically saves the following content:

[0124] The model weight file for optimal performance (e.g., best_model_epoch_*.pth);

[0125] Complete training log (including metrics such as loss and accuracy for each round, saved as training_metrics.csv);

[0126] Training process curves (loss and accuracy as a function of epochs);

[0127] The final confusion matrix and a visualization of typical prediction results.

[0128] Training process graphs and prediction results visualization, such as Figure 2 As shown.

[0129] Example 2

[0130] To verify the effectiveness of the method of the present invention, the settings such as "AuxLogits", "layered freezing", and data augmentation were turned off one by one, and the model in the present invention was compared with the conventional Inception-v3 model in an ablation experiment.

[0131] Based on the same dataset (training set:test 7:3), the following head-to-head comparison shows the number of training rounds, time, and accuracy:

[0132] Standard model: convergence takes about 48 epochs, training time is about 23.75 minutes, and final accuracy is 93.75%.

[0133] Improved model: converged in about 24 epochs, training time was about 9.47 minutes, and final accuracy was 95.31%.

[0134] like Figure 3 As shown.

[0135] Clearly, the training efficiency of this invention is doubled. Compared to 48 epochs in the conventional model, this invention only requires 24 epochs to reach convergence, reducing the training time by nearly two-thirds. Correspondingly, the accuracy is improved by nearly 2%.

[0136] Ablation experiments show that, based on the complete scheme of this invention, if the auxiliary classification head (AuxLogits) is removed alone or the random erasure enhancement strategy is disabled, the accuracy of the model on the validation set will decrease by about 0.6–1.2 percentage points, while the number of training rounds required to reach convergence will increase by 20%–40%, and the training efficiency will be significantly reduced.

[0137] The results above demonstrate that key design features such as AuxLogits and random erasure make substantial contributions to improving model accuracy and accelerating convergence. Overall, this invention, through multi-strategy collaborative optimization, significantly outperforms the conventional Inception-v3 baseline model in terms of recognition accuracy (improved by 1.56%), convergence speed (reduced by 50% in epochs), and training efficiency (reduced by nearly 60% in training time).

[0138] Example 3

[0139] In practical applications, the optimal model weights obtained in Example 1 are exported as ONNX format and deployed on a general-purpose edge GPU / CPU using TensorRT or ONNX Runtime to achieve online sorting and quality detection of morel mushroom images.

[0140] Deployment performance tests show that:

[0141] A throughput of 50–120 FPS can be achieved on GPU platforms;

[0142] A throughput of 15–40 FPS can be achieved on CPU platforms;

[0143] End-to-end inference latency is in the range of 10–80 ms (depending on the hardware environment).

[0144] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A CNN training method of Inception-v3 type, characterized in that, Includes the following steps: S1 preprocesses the sample image, the preprocessing including combined data augmentation and normalization, to obtain a preprocessed image of the sample; S2 Feature Extraction: The preprocessed image is fed into a CNN, where the parameters of the first few layers are frozen, while the parameters of subsequent layers and the classification head remain normal. The first and second classification results are output by the main classification head and the auxiliary branch classification head, respectively. S3 parameter update: For the first and second classification results, the parameters of the subsequent layers and the classification head layer are updated using the joint cross-entropy loss function. The total loss in the joint cross-entropy loss function is a linear combination of the main loss and the weighted auxiliary loss. S4 parameter optimization uses an adaptive first-order optimizer with weight decay to optimize the parameters and triggers learning rate decay based on validation set loss. S5 training terminates, using validation set metrics as a signal that the training objective has been achieved in mixed precision mode, and persistently saves the optimal parameter file at the time of training achievement.

2. The method of claim 1, wherein The data augmentation includes at least one or more of the following operations: random scale cropping, preferably cropping to 299×299 pixels; Random horizontal and / or vertical flipping; Random rotation, preferably with a rotation angle not exceeding 45°; Random affine translation, preferably with an affine translation amplitude not exceeding 15%; Random perspective perturbation, preferably with an amplitude not exceeding 0.2; Color fluctuation; Gaussian blur and / or random erasure; preferably, the area of ​​random erasure accounts for 1%-10%.

3. The method of any of claims 1-2, wherein, The Inception-v3 type CNN contains multi-branch convolutional units.

4. The method according to any one of claims 1-3, characterized in that: The main classification head includes the following structure: Dropout → Fully connected layer (2048→512) → ReLU activation layer → BN layer → Dropout → Fully connected layer (512→C).

5. The method according to any one of claims 1-4, characterized in that: The auxiliary branch classification head includes the following structure: Dropout → Fully Connected Layer (in_features→C).

6. The method of any one of claims 1-5, wherein, The joint cross-entropy loss function is defined as follows: L=Lmain+αLauxL = L_{\text{main}} + \alpha L_{\text{aux}}L=Lmain+αLaux The weighting coefficient α is preferably 0.2-0.6, and more preferably 0.

4.

7. A morel mushroom identification system, the system comprising an Inception-v3 type CNN trained by any one of the methods described in claims 1-6.

8. A method of identifying a morel, characterized by, The morel mushroom image to be identified is fed into an Inception-v3 type CNN trained by the method described in claims 1-6.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-6 or the morel mushroom identification method as described in claim 8.

10. A computer apparatus comprising a memory, a processor, and a computer program stored on the memory, wherein the computer program, when executed by the processor, causes the processor to perform the method of any one of claims 1 to 9. The processor executes the computer program to implement the method of any one of claims 1-6 or the morel mushroom identification method of claim 8.

Images