Container small target semi-supervised identification method and system

Abstract

The invention discloses a semi-supervised identification method and system for a small target of a container, and belongs to the technical field of artificial intelligence and computer vision, and the method comprises the steps: carrying out the target detection of a container image through a pre-trained target detection model, intercepting a sub-image, and inputting the sub-image into an initial classification model, and obtaining a classification confidence coefficient; the uncertainty of the model on a sample classification result is quantified through a Monte Carlo Dropout method; a feature space distance filtering and dynamic threshold adjusting mechanism is combined, and samples with high confidence, low uncertainty and consistent feature space are screened out to serve as pseudo label data; pseudo label data and initial synthesis data are mixed, and the generalization ability of the model is gradually improved through semi-supervised iterative training. According to the method, the dependence on manual annotation can be remarkably reduced, meanwhile, the distribution difference between synthetic data and real scene data is gradually reduced, and finally, high-precision recognition and strong generalization ability of a classification model in a real scene are achieved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and computer vision technology, and in particular to a method and system for semi-supervised recognition of small targets in containers. Background Art

[0002] Existing small target recognition systems for container scenarios use a two-stage architecture: target detection and classification. Target detection models (such as YOLO and EfficientDet) are used to locate the target area, while classification models (such as ResNet and EfficientNet) classify the target. In container scenarios, due to the large number of target categories (e.g., 5,000 categories) and the fact that most of them are small-scale, end-to-end detection and recognition using a single model is ineffective. Therefore, a two-stage model is typically used to perform target detection and classification separately. Target detection models already possess strong generalization and robustness through pre-training, and their detection performance does not require further improvement. Classification models, on the other hand, rely on massive amounts of manually annotated data for training, or use a small amount of high-definition images to expand the dataset through data augmentation and 3D reconstruction techniques to improve generalization.

[0003] Traditional supervised learning requires manual labeling of massive samples. In scenarios with over 5,000 container categories, labeling is time-consuming, labor-intensive, and costly. This problem arises because fine-grained classification requires high-precision labeling, but manual labeling efficiency is exponentially outstripping data requirements.

[0004] While augmenting datasets with a small number of high-definition images through data augmentation and 3D reconstruction techniques can reduce labeling reliance, this approach suffers from domain shift from real-world scenarios, resulting in reduced performance of classification models in real-world deployments. This issue arises because synthetic data struggles to accurately simulate the complex optical properties of real environments (such as specular reflections and refraction from transparent materials). Consequently, the resulting distortion and occlusion patterns are inconsistent with the real-world distribution.

[0005] In summary, there are currently two core issues with multi-small object recognition in containerized scenarios: First, the sheer number of target categories (e.g., 5,000) leads to exponentially increasing manual labeling costs, making it difficult to meet the demands of large-scale applications. Second, data generated using data augmentation and 3D reconstruction techniques using a small number of high-definition images differs significantly from real-world data in terms of texture, lighting, and occlusion, making it difficult to align the model's feature space in the real world and resulting in poor recognition results. These issues limit the effectiveness of existing technologies in containerized scenarios, necessitating an innovative solution that balances accuracy, efficiency, and generalization. Summary of the Invention

[0006] The technical task of the present invention is to address the above shortcomings and provide a semi-supervised recognition method and system for small targets in containers, which can significantly reduce the dependence on manual labeling, while gradually narrowing the distribution difference between synthetic data and real scene data, and ultimately achieve high-precision recognition and strong generalization ability of the classification model in real scenes.

[0007] The technical solution adopted by the present invention to solve its technical problem is:

[0008] A semi-supervised recognition method for small container objects uses a joint filtering mechanism integrating confidence screening, uncertainty assessment, and feature space consistency verification to select high-reliability samples from massive unlabeled container sub-images and construct a progressively optimized training dataset.

[0009] Use the pre-trained object detection model to perform object detection on the container image, extract the sub-image and input it into the initial classification model to obtain the classification confidence;

[0010] The Monte Carlo Dropout method is used to quantify the uncertainty of the model's sample classification results. By combining feature space distance filtering with a dynamic threshold adjustment mechanism, samples with high confidence, low uncertainty, and consistent feature space are selected as pseudo-label data.

[0011] Pseudo-labeled data is mixed with the initial synthetic data, and the generalization ability of the model is gradually improved through semi-supervised iterative training.

[0012] This method makes full use of unlabeled container sub-images to automatically generate high-quality pseudo-label data, significantly reducing the dependence on manual labeling. At the same time, it gradually narrows the distribution difference between synthetic data and real scene data, and ultimately achieves high-precision recognition and strong generalization capabilities of the classification model in real scenes.

[0013] Furthermore, the implementation steps of the method include:

[0014] 1) Using the original container image as input, EfficientDet performs object detection and sub-image extraction;

[0015] 2) Training the initial EfficientNet classification model based on enhanced synthetic data:

[0016] 3) Generate multi-dimensional pseudo labels through three parallel modules: confidence screening, Monte Carlo uncertainty assessment, and feature space consistency verification;

[0017] 4) Adaptively filter high-quality data based on sample distribution through a dynamic threshold adjustment module;

[0018] 5) Perform semi-supervised iterative training based on mixed synthetic data and pseudo-labeled data, using layer-wise fine-tuning and weighted loss strategies;

[0019] 6) Form an iterative optimization closed loop and achieve model self-enhancement through performance convergence judgment.

[0020] Furthermore, this method is based on multimodal feature fusion and dynamic domain adaptation technology to improve the model's adaptability to complex scenarios. The specific implementation steps include:

[0021] 1) Add a multimodal feature fusion module to enhance feature discrimination through the RGB+edge feature dual-branch architecture;

[0022] 2) Introducing a dynamic domain adaptation training mechanism, including domain classifiers and adversarial training strategies;

[0023] 3) Adaptive weighted concatenation of cross-modal features is achieved through a feature fusion module;

[0024] 4) Domain adaptation training uses dynamic loss weights to gradually enhance domain-invariant feature learning;

[0025] 5) The semi-supervised training phase simultaneously optimizes classification accuracy and domain confusion;

[0026] 6) Multimodal features participate in pseudo-label generation to improve robustness to illumination changes and occlusion.

[0027] Furthermore, the specific implementation of this method is as follows:

[0028] Step 1: Target detection obtains sub-images of each category,

[0029] During the object detection phase, the original container image is input. The trained EfficientDet detection model is first used to detect the objects in the image and output the target box coordinates and confidence score. A sub-image is then captured from the original image based on the target box coordinates and scaled to a uniform pixel size to meet the input requirements of the subsequent classification model. The target box capture and size standardization provide a unified input format for the subsequent classification model, reducing interference caused by inconsistent preprocessing and laying a solid foundation for subsequent multi-dimensional pseudo-label generation and model iterative optimization.

[0030] Step 2: Train the initial classification model,

[0031] An initial classification and recognition model is trained based on high-definition images of various categories. The training dataset contains multiple categories, and each category contains no more than 10 high-definition images with no background or a single background. These images are obtained by crawling the Internet or taking pictures using mobile phones, cameras, and other devices. To improve the generalization ability of the model, the dataset is enhanced in multiple dimensions, including spatial transformation (such as random rotation of ±30°, affine shear amplitude of 0.2, perspective transformation distortion of 0.1) and photometric transformation (such as adjusting brightness by ±20%, contrast by ±15%, and adding Gaussian noise σ = 0.05) to simulate the diverse changes in real scenes. Then, the EfficientNet architecture model is used for training, using the cross-entropy loss function, the initial learning rate is set to 0.01 and decayed based on the cosine annealing strategy, the batch size is 128, and training is carried out for 100 epochs until the model converges.

[0032] By adopting data augmentation strategies for the dataset, the model can initially learn the feature representation of the target under different scales, postures and lighting conditions, thus having basic feature extraction capabilities;

[0033] Step 3: Multi-dimensional pseudo-label generation, including confidence screening, uncertainty assessment, feature space distance filtering, and dynamic threshold adjustment.

[0034] Step 4: Semi-supervised iterative training,

[0035] First, the generated pseudo-labeled dataset is mixed with the initial synthetic dataset in a 1:1 ratio to form a new training dataset. Then, a layered fine-tuning strategy is adopted to freeze the underlying convolutional layers of EfficientNet to retain the basic feature extraction capabilities, and only fine-tune the fully connected layers to adapt to the new data distribution. At the same time, a weighted cross-entropy loss function (with a loss weight of 0.7 for synthetic data and a loss weight of 0.3 for pseudo-labeled data) is used for optimization. The learning rate is set to 0.01, the batch size is 64, and the training is carried out for 100 epochs. Finally, by continuously monitoring the accuracy of the validation set, the iterative training is terminated when the accuracy improvement in three consecutive rounds of training is less than 0.5%.

[0036] Through a hybrid training strategy, the model can gradually adapt to the distribution of real-world data while retaining the clear features of synthetic data and alleviating domain differences. The layered fine-tuning strategy ensures the stability of feature extraction while improving the adaptability of the classification head and avoiding overfitting. Dynamically amplified pseudo-label data and model optimization form a positive feedback loop, achieving a self-reinforcing mechanism of "data quality improvement → model optimization → more accurate pseudo-labeling", thereby continuously improving the model's generalization ability and recognition accuracy.

[0037] Step 5: Iterative optimization and self-enhancement,

[0038] After completing steps 1 to 4, use the currently trained optimized classification model as the new base model and repeat steps 3 and 4 to continuously update the pseudo-label data and iteratively optimize the semi-supervised training. The details are as follows:

[0039] The new basic model makes predictions on a large number of unlabeled container sub-images, re-executes the multi-dimensional pseudo-label generation process (including confidence screening, uncertainty assessment, and feature space distance filtering), and dynamically adjusts the threshold according to the latest sample distribution to generate updated high-quality pseudo-label data; then, the updated pseudo-label data is re-mixed with the initial synthetic data in a 1:1 ratio, and further semi-supervised iterative training is performed using a layered fine-tuning strategy and weighted cross-entropy loss to gradually improve the model's recognition accuracy and generalization ability in real scenarios; after multiple iterations, the quality and quantity of pseudo-label data are significantly improved, forming a positive feedback loop of "the more data is used, the more accurate it is, and the more trained the model, the stronger it is", until the model performance reaches the expected indicators and tends to converge.

[0040] Furthermore, the multi-dimensional pseudo-label generation is specifically implemented as follows:

[0041] (1) Confidence screening:

[0042] The intercepted container sub-image is input into the initial classification model, which outputs a multi-dimensional probability vector, where each element represents the predicted probability that the sub-image belongs to the current category. The maximum confidence in the probability vector is calculated and its corresponding category label is determined.

[0043] (2) Uncertainty assessment:

[0044] Using the Monte Carlo Dropout method, the Dropout layer is enabled during classification model inference (with a retention rate set to 0.5). Ten forward propagations are performed on the same subgraph to obtain 10 probability vectors, each containing the predicted probability values of multiple categories. Based on the results of these 10 forward propagations, the average predicted probability of each category is calculated, and the predicted entropy is further calculated. The entropy value is used to quantify the uncertainty of the model's sample classification results.

[0045] By using the Monte Carlo Dropout method, the Dropout layer is enabled during model inference and multiple forward propagations are performed on the same subgraph to quantify the uncertainty of the model's sample classification results.

[0046] (3) Feature space distance filtering:

[0047] First, based on the initial training set (synthetic data), we extract the feature vectors of all samples (with a dimension of 2048, output by the global pooling of the last layer of ResNet-50) and calculate the feature mean of each category as the feature center of that category. Then, we calculate the cosine similarity between the feature vector of the container subimage and the feature centers of all categories, and take the maximum similarity value as the match between the subimage and the closest category.

[0048] By filtering the feature space distance, we ensure that pseudo-labeled samples are highly consistent with known category centers in the feature space, thus avoiding mislabeling caused by distribution differences between synthetic data and real-world data. This design effectively reduces interference from out-of-distribution samples, improves the reliability of pseudo-labels, and provides the model with higher-quality training data, enhancing its generalization ability in real-world scenarios.

[0049] (4) Dynamic threshold adjustment:

[0050] First, the distribution of confidence, entropy and feature similarity of all candidate samples in the current round is statistically analyzed; then, the threshold is dynamically updated according to the quantile, and the confidence threshold is set to the top 90% quantile of the current sample confidence distribution, the entropy threshold is set to the top 10% quantile of the current sample entropy distribution, and the feature similarity threshold is set to the top 90% quantile of the current sample similarity distribution; finally, only the samples that simultaneously meet the conditions of confidence greater than or equal to the updated confidence threshold, entropy less than the updated entropy threshold, and feature similarity greater than or equal to the updated similarity threshold are retained as the final pseudo-label data.

[0051] A dynamic threshold adjustment mechanism adaptively balances the quantity and quality of pseudo-labeled data. In the early stages of model training, a high screening threshold is set to ensure the reliability of pseudo-labels. As model performance improves, the threshold is gradually relaxed to utilize more difficult examples, thereby expanding the coverage of training data. This design not only effectively reduces the introduction of noisy labels but also gradually improves the model's generalization capabilities.

[0052] Furthermore, the model's adaptability to complex scenarios is improved through multimodal feature fusion and dynamic domain adaptation technology. Feature discrimination is enhanced through multimodal feature fusion. In the feature space distance filtering stage, a multimodal feature joint alignment strategy is adopted. The specific implementation is as follows:

[0053] Dual-branch feature extraction: A dual-branch network structure is constructed. The main branch extracts RGB image features based on EfficientNet, and the auxiliary branch extracts edge features based on the pre-trained EdgeNet model to capture the contour and structural information of the target.

[0054] Feature fusion and weighted splicing: The feature fusion module performs adaptive weighted fusion on two types of features to generate a fused multimodal feature vector;

[0055] Cross-modal similarity metric: This metric calculates the cosine similarity between the container sub-image and the category center based on fused features, replacing the single RGB feature metric. This design effectively enhances robustness to lighting changes and partial occlusion. For example, in scenes with strong reflections or partial occlusions, edge features can supplement the discriminative information of RGB features to avoid mismatches.

[0056] Furthermore, the multimodal feature fusion and dynamic domain adaptation technologies are used to improve the model's adaptability to complex scenarios. Dynamic domain adaptation is used to optimize cross-domain generalization capabilities. Dynamic domain adaptation is introduced into classification model training to narrow the distribution gap between synthetic data and real-world scene data, including:

[0057] Domain adversarial training architecture: A domain classifier is inserted before the fully connected layer of the initial classification model. Through an adversarial learning strategy, the feature extractor is forced to generate domain-invariant features. The feature extractor must simultaneously optimize classification accuracy (distinguishing target categories) and confuse the domain classifier (making it unable to distinguish between synthetic data and real data).

[0058] Dynamic loss weight adjustment: The total loss function is defined as the weighted sum of classification loss and domain adversarial loss, where the weight of the domain adversarial loss increases linearly with the training process. In the initial stage, classification loss is mainly used to quickly establish basic feature representation. In the later stage, the domain alignment strength is gradually enhanced to achieve progressive distribution alignment between synthetic data and real data.

[0059] The present invention also claims protection for a semi-supervised recognition system for small objects in a container, comprising:

[0060] Target detection and sub-image acquisition module, used to complete target detection and sub-image interception;

[0061] Initial classification model training module, which trains the initial classification model based on enhanced synthetic data;

[0062] Multi-dimensional pseudo-label generation module generates multi-dimensional pseudo-labels based on confidence screening, uncertainty assessment, and feature space consistency verification;

[0063] Dynamic threshold adjustment module, used to adaptively filter high-quality data based on sample distribution;

[0064] Semi-supervised iterative training, which is used to implement semi-supervised iterative training based on mixed synthetic data and pseudo-labeled data;

[0065] Iterative optimization module forms an iterative optimization closed loop and realizes model self-enhancement through performance convergence judgment;

[0066] The system specifically realizes semi-supervised recognition of small targets in containers through the above method.

[0067] The present invention also claims protection for a semi-supervised identification device for small objects in a container, comprising: at least one memory and at least one processor;

[0068] The at least one memory is configured to store a machine-readable program;

[0069] The at least one processor is configured to call the machine-readable program to implement the above method.

[0070] The present invention also claims protection for a computer-readable medium having computer instructions stored thereon, which are capable of implementing the above method when executed by a processor.

[0071] Compared with the prior art, the semi-supervised recognition method and system for small container targets of the present invention have the following beneficial effects:

[0072] This semi-supervised framework combines multi-dimensional pseudo-label generation (confidence + uncertainty + feature space) with dynamic domain adaptation optimization. This framework enhances cross-domain discrimination capabilities through multimodal feature fusion. Combined with dynamic threshold adjustment and adversarial training, it enables efficient automatic labeling of unlabeled container sub-images. This improves the model's recognition accuracy and generalization capabilities in real-world scenarios. Furthermore, through dynamic domain adversarial training and multimodal feature alignment, it effectively narrows the distribution gap between high-definition synthetic data and real-world data, providing a high-precision, low-cost solution for object recognition in container scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 This is a diagram illustrating the implementation structure of a basic solution for a semi-supervised recognition method for small objects in a container provided by an embodiment of the present invention;

[0074] Figure 2 This is a diagram illustrating the structure of a further optimized solution for the semi-supervised recognition method for small targets in containers provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0075] The present invention will be further described below with reference to specific embodiments.

[0076] The present invention provides a semi-supervised recognition method for small container objects. By integrating a three-dimensional joint filtering mechanism, namely confidence screening, uncertainty assessment, and feature space consistency verification, the method selects high-reliability samples from a large number of unlabeled container sub-images and constructs a progressively optimized training dataset.

[0077] First, we use a pre-trained object detection model to perform object detection on the container image, extract sub-images, and input them into the initial classification model to obtain the classification confidence.

[0078] Then, the uncertainty of the model's sample classification results is quantified through the Monte Carlo Dropout method. Combined with feature space distance filtering and dynamic threshold adjustment mechanism, samples with high confidence, low uncertainty and consistent feature space are screened out as pseudo-label data.

[0079] Finally, the pseudo-labeled data is mixed with the initial synthetic data, and the generalization ability of the model is gradually improved through semi-supervised iterative training.

[0080] In this way, we can make full use of unlabeled container sub-images to automatically generate high-quality pseudo-label data, significantly reducing the dependence on manual labeling. At the same time, we can gradually narrow the distribution difference between synthetic data and real scene data, and ultimately achieve high-precision recognition and strong generalization capabilities of the classification model in real scenes.

[0081] To address the issues raised in the background art, a company's container product detection and recognition solution uses a small number of high-definition images for 3D reconstruction. Using a 3D engine, it renders 3D representations under varying lighting, angles, occlusions, and deformations to generate 2D images. A neural network model is then trained based on a large, evenly distributed collection of synthetic 2D images and annotation information. This approach reduces annotation costs by automatically generating annotated data, reducing manual annotation workload and improving efficiency. It also reduces reliance on real samples: a large amount of training data can be synthesized from only a small number of images, addressing sample collection challenges. It enhances data diversity: the 3D engine simulates varying lighting, angles, occlusions, and deformations, improving model generalization. However, it has the following drawbacks: Synthetic-real data discrepancies: rendered images may not fully reproduce complex details such as real-world lighting and material reflections, affecting recognition accuracy in real scenes. Limited cross-scenario transferability: model performance may decline if the appearance, material, or placement of the real containers differ significantly from the synthetic data. Reliance on 3D reconstruction quality: model accuracy is limited by the 3D reconstruction and rendering techniques used. Loss of detail or texture distortion can compromise data reliability.

[0082] This method comprehensively considers the current state of the container image recognition industry and addresses core issues in container scenarios, such as the large number of target categories, high labeling costs, and data augmentation, which create domain differences between generated data and real-world scenarios. It proposes a semi-supervised learning framework based on multi-dimensional pseudo-label quality assessment and dynamic threshold adaptive adjustment. By integrating a three-dimensional joint filtering mechanism—confidence screening, uncertainty assessment, and feature space consistency verification—it selects high-reliability samples from a massive collection of unlabeled container subimages, constructing a progressively optimized training dataset, and gradually narrowing the distribution gap between training data and real-world data, ultimately achieving continuous improvement in the generalization capability of the classification model.

[0083] The specific implementation of this method is as follows:

[0084] Step 1: Target detection obtains sub-images of each category.

[0085] During the object detection phase, the original container image is input. The trained Efficientdet detection model is first used to detect the objects in the image and output the target bounding box coordinates and confidence scores. A sub-image is then captured from the original image based on the target bounding box coordinates and uniformly scaled to 224×224 pixels to meet the input requirements of the subsequent classification model. This target bounding box capture and resizing standardization provides a unified input format for the subsequent classification model, reducing interference caused by inconsistent preprocessing and laying a solid foundation for subsequent multi-dimensional pseudo-label generation and iterative model optimization.

[0086] Step 2: Train the initial classification model.

[0087] An initial classification and recognition model is trained based on high-definition images of various categories. The training dataset contains multiple categories, and each category contains no more than 10 high-definition images with no background or a single background. These images are obtained by crawling the Internet or taking pictures using mobile phones, cameras and other devices. In order to improve the generalization ability of the model, the dataset is enhanced in multiple dimensions, including spatial transformation (such as random rotation of ±30°, affine shear amplitude of 0.2, perspective transformation distortion of 0.1) and photometric transformation (such as adjusting brightness by ±20%, contrast by ±15%, and adding Gaussian noise σ=0.05) to simulate the diverse changes in real scenes. Then, the EfficientNet architecture model is used for training, with a cross-entropy loss function, an initial learning rate set to 0.01 and decayed based on a cosine annealing strategy, a batch size of 128, and training for 100 epochs until the model converges.

[0088] By adopting data augmentation strategies for the dataset, the model can initially learn the feature representations of the target under different scales, postures and lighting conditions, thereby having basic feature extraction capabilities.

[0089] Step 3: Multi-dimensional pseudo-label generation.

[0090] (1) Confidence screening:

[0091] The captured container subimage is fed into the initial classification model, which outputs a multidimensional probability vector, where each element represents the predicted probability that the subimage belongs to the current category. The maximum confidence level in the probability vector is calculated, and its corresponding category label is determined.

[0092] (2) Uncertainty assessment:

[0093] The Monte Carlo Dropout method is used. The Dropout layer is enabled during classification model inference (the retention rate is set to 0.5). Ten forward propagations are performed on the same subgraph to obtain 10 probability vectors, each of which contains the predicted probability values of multiple categories. Based on the results of these 10 forward propagations, the average predicted probability of each category is calculated, and the predicted entropy is further calculated. The entropy value is used to quantify the uncertainty of the model's sample classification results.

[0094] By using the Monte Carlo Dropout method, the Dropout layer is enabled during model inference and multiple forward propagations are performed on the same subgraph to quantify the uncertainty of the model's sample classification results.

[0095] (3) Feature space distance filtering:

[0096] First, based on the initial training set (synthetic data), the feature vectors of all samples (with a dimension of 2048, output by the global pooling of the last layer of ResNet-50) are extracted, and the feature mean of each category is calculated as the feature center of that category. Then, the cosine similarity between the feature vector of the container subimage and the feature centers of all categories is calculated, and the maximum similarity value is taken as the matching degree between the subimage and the closest category.

[0097] By filtering the feature space distance, we ensure that pseudo-labeled samples are highly consistent with known category centers in the feature space, thus avoiding mislabeling caused by distribution differences between synthetic and real-world data. This design effectively reduces interference from out-of-distribution samples, improves the reliability of pseudo-labels, and provides the model with higher-quality training data, enhancing its generalization ability in real-world scenarios.

[0098] (4) Dynamic threshold adjustment:

[0099] First, the distribution of confidence, entropy and feature similarity of all candidate samples in the current round is statistically analyzed; then, the threshold is dynamically updated according to the quantile, and the confidence threshold is set to the top 90% quantile of the current sample confidence distribution, the entropy threshold is set to the top 10% quantile of the current sample entropy distribution, and the feature similarity threshold is set to the top 90% quantile of the current sample similarity distribution; finally, only the samples that simultaneously meet the conditions of confidence greater than or equal to the updated confidence threshold, entropy less than the updated entropy threshold, and feature similarity greater than or equal to the updated similarity threshold are retained as the final pseudo-label data.

[0100] A dynamic threshold adjustment mechanism adaptively balances the quantity and quality of pseudo-labeled data. In the early stages of model training, a high screening threshold is set to ensure the reliability of pseudo-labels. As model performance improves, the threshold is gradually relaxed to utilize more difficult examples, thereby expanding the coverage of training data. This design not only effectively reduces the introduction of noisy labels but also gradually improves the model's generalization capabilities.

[0101] Step 4: Semi-supervised iterative training.

[0102] First, the generated pseudo-labeled dataset is mixed with the initial synthetic dataset in a 1:1 ratio to form a new training dataset; then, a hierarchical fine-tuning strategy is adopted to freeze the underlying convolutional layers of EfficientNet to retain the basic feature extraction capability, and only fine-tune the fully connected layers to adapt to the new data distribution. At the same time, a weighted cross-entropy loss function (the loss weight of synthetic data is 0.7, and the loss weight of pseudo-labeled data is 0.3) is used for optimization, the learning rate is set to 0.01, the batch size is 64, and 100 epochs are trained; finally, through continuous monitoring of the validation set accuracy, the iterative training is terminated when the accuracy improvement in three consecutive rounds of training is less than 0.5%.

[0103] Through the hybrid training strategy, the model can gradually adapt to the data distribution of real scenes while retaining the clear features of synthetic data and alleviating the domain difference problem; the layered fine-tuning strategy improves the adaptability of the classification head while ensuring the stability of feature extraction and avoids overfitting; the dynamically expanded pseudo-label data and model optimization form a positive feedback loop, realizing the self-enhancement mechanism of "data quality improvement → model optimization → more accurate pseudo-labeling", thereby continuously improving the model's generalization ability and recognition accuracy.

[0104] Step 5: Iterative optimization and self-enhancement.

[0105] After completing steps 1 to 4, use the currently trained optimized classification model as the new basic model and repeat steps 3 and 4 to continuously update the pseudo-label data and iteratively optimize the semi-supervised training. The details are as follows:

[0106] The new basic model predicts a large number of unlabeled container sub-images, re-executes the multi-dimensional pseudo-label generation process (including confidence screening, uncertainty assessment, and feature space distance filtering), and dynamically adjusts the threshold according to the latest sample distribution to generate updated high-quality pseudo-label data. The updated pseudo-label data is then re-mixed with the initial synthetic data in a 1:1 ratio, and further semi-supervised iterative training is performed using a layered fine-tuning strategy and weighted cross-entropy loss to gradually improve the model's recognition accuracy and generalization ability in real-world scenarios. After multiple iterations, the quality and quantity of pseudo-label data have been significantly improved, forming a positive feedback loop of "the more data is used, the more accurate it is, and the more the model is trained, the stronger it is", until the model performance reaches the expected indicators and converges.

[0107] This method further proposes a multimodal feature fusion strategy, which improves the robustness to illumination changes and occlusion by weighted fusion of RGB features and edge features, combined with cross-modal similarity measurement. Secondly, a dynamic domain adversarial training architecture is designed to embed a learnable domain adapter in the classification model, and achieve feature space alignment between synthetic data and real data through progressive adversarial loss weight adjustment. Finally, a multidimensional dynamic threshold mechanism is developed to dynamically adjust the confidence, prediction entropy, and feature similarity quantiles to adaptively balance the quality and quantity of pseudo-labels.

[0108] 1. Enhance feature discrimination through multimodal feature fusion.

[0109] In the feature space distance filtering stage, a multimodal feature joint alignment strategy is adopted. The specific implementation is as follows:

[0110] Dual-branch feature extraction: A dual-branch network structure is constructed. The main branch extracts RGB image features based on EfficientNet, and the auxiliary branch extracts edge features based on the pre-trained EdgeNet model to capture the contour and structural information of the target.

[0111] Feature fusion and weighted splicing: The feature fusion module performs adaptive weighted fusion on two types of features to generate a fused multimodal feature vector;

[0112] Cross-modal similarity metric: This metric calculates the cosine similarity between the container sub-image and the category center based on fused features, replacing the single RGB feature metric. This design effectively enhances robustness to lighting changes and partial occlusion. For example, in scenes with strong reflections or partial occlusions, edge features can supplement the discriminative information of RGB features to avoid mismatches.

[0113] 2. Optimize cross-domain generalization capabilities through dynamic domain adaptation.

[0114] In order to narrow the distribution gap between synthetic data and real scene data, a dynamic domain adaptation mechanism is introduced in the classification model training:

[0115] Domain adversarial training architecture: A domain classifier is inserted before the fully connected layer of the initial classification model. Through adversarial learning, the feature extractor is forced to generate domain-invariant features. Specifically, the feature extractor needs to simultaneously optimize classification accuracy (distinguishing target classes) and confuse the domain classifier (making it unable to distinguish between synthetic data and real data).

[0116] Dynamic loss weight adjustment: The total loss function is defined as the weighted sum of classification loss and domain adversarial loss, where the weight of the domain adversarial loss increases linearly with the training process. Initially, the classification loss is used to quickly establish basic feature representations. Later, the domain alignment strength is gradually increased to achieve progressive distribution alignment between synthetic and real data.

[0117] like Figure 1 As shown, the basic implementation steps of this method are:

[0118] 1. Using the original container image as input, perform object detection and sub-image extraction using EfficientDet.

[0119] 2. Train the initial EfficientNet classification model based on enhanced synthetic data.

[0120] 3. Generate multi-dimensional pseudo labels through three parallel modules: confidence screening, Monte Carlo uncertainty assessment, and feature space consistency verification.

[0121] 4. Adaptively filter high-quality data based on sample distribution through the dynamic threshold adjustment module.

[0122] 5. Perform semi-supervised iterative training based on mixed synthetic data and pseudo-labeled data, using layered fine-tuning and weighted loss strategies.

[0123] 6. Form an iterative optimization closed loop and achieve model self-enhancement through performance convergence judgment.

[0124] like Figure 2 As shown in the figure, based on the above basic solution and based on multimodal feature fusion and dynamic domain adaptation technology, the model's adaptability to complex scenarios is improved; the specific implementation steps are:

[0125] 1. Add a multimodal feature fusion module to enhance feature discrimination through the RGB+edge feature dual-branch architecture.

[0126] 2. Introduce a dynamic domain adaptation training mechanism, including domain classifiers and adversarial training strategies.

[0127] 3. Adaptive weighted splicing of cross-modal features is achieved through the feature fusion module.

[0128] 4. Domain adaptation training uses dynamic loss weights to gradually enhance domain-invariant feature learning.

[0129] 5. The semi-supervised training stage optimizes both classification accuracy and domain confusion.

[0130] 6. Multimodal features participate in pseudo-label generation to improve robustness to lighting changes and occlusion.

[0131] The two structural diagrams illustrate the technological evolution from foundational to enhanced approaches. The optimized solution significantly enhances the model's cross-domain generalization capabilities in complex scenarios through multimodal feature fusion and dynamic domain adaptation. The closed-loop design in the structural diagrams reflects the self-enhancement mechanism proposed in this method: "Data quality improvement → model optimization → more accurate pseudo-labeling."

[0132] An embodiment of the present invention further provides a semi-supervised recognition system for small objects in a container, comprising:

[0133] Target detection and sub-image acquisition module, used to complete target detection and sub-image interception;

[0134] Initial classification model training module, which trains the initial classification model based on enhanced synthetic data;

[0135] Multi-dimensional pseudo-label generation module generates multi-dimensional pseudo-labels based on confidence screening, uncertainty assessment, and feature space consistency verification;

[0136] Dynamic threshold adjustment module, used to adaptively filter high-quality data based on sample distribution;

[0137] Semi-supervised iterative training, which is used to implement semi-supervised iterative training based on mixed synthetic data and pseudo-labeled data;

[0138] Iterative optimization module forms an iterative optimization closed loop and realizes model self-enhancement through performance convergence judgment;

[0139] The system implements semi-supervised recognition of small container objects through the semi-supervised recognition method for small container objects described in the above embodiment. The details are as follows:

[0140] Step 1: Target detection obtains sub-images of each category.

[0141] During the object detection phase, the original container image is input. The trained Efficientdet detection model is first used to detect the objects in the image and output the target bounding box coordinates and confidence scores. A sub-image is then captured from the original image based on the target bounding box coordinates and uniformly scaled to 224×224 pixels to meet the input requirements of the subsequent classification model. This target bounding box capture and resizing standardization provides a unified input format for the subsequent classification model, reducing interference caused by inconsistent preprocessing and laying a solid foundation for subsequent multi-dimensional pseudo-label generation and iterative model optimization.

[0142] Step 2: Train the initial classification model.

[0143] An initial classification and recognition model is trained based on high-definition images of various categories. The training dataset contains multiple categories, and each category contains no more than 10 high-definition images with no background or a single background. These images are obtained by crawling the Internet or taking pictures using mobile phones, cameras and other devices. In order to improve the generalization ability of the model, the dataset is enhanced in multiple dimensions, including spatial transformation (such as random rotation of ±30°, affine shear amplitude of 0.2, perspective transformation distortion of 0.1) and photometric transformation (such as adjusting brightness by ±20%, contrast by ±15%, and adding Gaussian noise σ=0.05) to simulate the diverse changes in real scenes. Then, the EfficientNet architecture model is used for training, with a cross-entropy loss function, an initial learning rate set to 0.01 and decayed based on a cosine annealing strategy, a batch size of 128, and training for 100 epochs until the model converges.

[0144] By adopting data augmentation strategies for the dataset, the model can initially learn the feature representations of the target under different scales, postures and lighting conditions, thereby having basic feature extraction capabilities.

[0145] Step 3: Multi-dimensional pseudo-label generation.

[0146] (1) Confidence screening:

[0147] The captured container subimage is fed into the initial classification model, which outputs a multidimensional probability vector, where each element represents the predicted probability that the subimage belongs to the current category. The maximum confidence level in the probability vector is calculated, and its corresponding category label is determined.

[0148] (2) Uncertainty assessment:

[0149] The Monte Carlo Dropout method is used. The Dropout layer is enabled during classification model inference (the retention rate is set to 0.5). Ten forward propagations are performed on the same subgraph to obtain 10 probability vectors, each of which contains the predicted probability values of multiple categories. Based on the results of these 10 forward propagations, the average predicted probability of each category is calculated, and the predicted entropy is further calculated. The entropy value is used to quantify the uncertainty of the model's sample classification results.

[0150] By using the Monte Carlo Dropout method, the Dropout layer is enabled during model inference and multiple forward propagations are performed on the same subgraph to quantify the uncertainty of the model's sample classification results.

[0151] (3) Feature space distance filtering:

[0152] First, based on the initial training set (synthetic data), the feature vectors of all samples (with a dimension of 2048, output by the global pooling of the last layer of ResNet-50) are extracted, and the feature mean of each category is calculated as the feature center of that category. Then, the cosine similarity between the feature vector of the container subimage and the feature centers of all categories is calculated, and the maximum similarity value is taken as the matching degree between the subimage and the closest category.

[0153] By filtering the feature space distance, we ensure that pseudo-labeled samples are highly consistent with known category centers in the feature space, thus avoiding mislabeling caused by distribution differences between synthetic and real-world data. This design effectively reduces interference from out-of-distribution samples, improves the reliability of pseudo-labels, and provides the model with higher-quality training data, enhancing its generalization ability in real-world scenarios.

[0154] (4) Dynamic threshold adjustment:

[0155] First, the distribution of confidence, entropy and feature similarity of all candidate samples in the current round is statistically analyzed; then, the threshold is dynamically updated according to the quantile, and the confidence threshold is set to the top 90% quantile of the current sample confidence distribution, the entropy threshold is set to the top 10% quantile of the current sample entropy distribution, and the feature similarity threshold is set to the top 90% quantile of the current sample similarity distribution; finally, only the samples that simultaneously meet the conditions of confidence greater than or equal to the updated confidence threshold, entropy less than the updated entropy threshold, and feature similarity greater than or equal to the updated similarity threshold are retained as the final pseudo-label data.

[0156] A dynamic threshold adjustment mechanism adaptively balances the quantity and quality of pseudo-labeled data. In the early stages of model training, a high screening threshold is set to ensure the reliability of pseudo-labels. As model performance improves, the threshold is gradually relaxed to utilize more difficult examples, thereby expanding the coverage of training data. This design not only effectively reduces the introduction of noisy labels but also gradually improves the model's generalization capabilities.

[0157] Step 4: Semi-supervised iterative training.

[0158] First, the generated pseudo-labeled dataset is mixed with the initial synthetic dataset in a 1:1 ratio to form a new training dataset; then, a hierarchical fine-tuning strategy is adopted to freeze the underlying convolutional layers of EfficientNet to retain the basic feature extraction capability, and only fine-tune the fully connected layers to adapt to the new data distribution. At the same time, a weighted cross-entropy loss function (the loss weight of synthetic data is 0.7, and the loss weight of pseudo-labeled data is 0.3) is used for optimization, the learning rate is set to 0.01, the batch size is 64, and 100 epochs are trained; finally, through continuous monitoring of the validation set accuracy, the iterative training is terminated when the accuracy improvement in three consecutive rounds of training is less than 0.5%.

[0159] Through the hybrid training strategy, the model can gradually adapt to the data distribution of real scenes while retaining the clear features of synthetic data and alleviating the domain difference problem; the layered fine-tuning strategy improves the adaptability of the classification head while ensuring the stability of feature extraction and avoids overfitting; the dynamically expanded pseudo-label data and model optimization form a positive feedback loop, realizing the self-enhancement mechanism of "data quality improvement → model optimization → more accurate pseudo-labeling", thereby continuously improving the model's generalization ability and recognition accuracy.

[0160] Step 5: Iterative optimization and self-enhancement.

[0161] After completing steps 1 to 4, use the currently trained optimized classification model as the new basic model and repeat steps 3 and 4 to continuously update the pseudo-label data and iteratively optimize the semi-supervised training. The details are as follows:

[0162] The new basic model predicts a large number of unlabeled container sub-images, re-executes the multi-dimensional pseudo-label generation process (including confidence screening, uncertainty assessment, and feature space distance filtering), and dynamically adjusts the threshold according to the latest sample distribution to generate updated high-quality pseudo-label data. The updated pseudo-label data is then re-mixed with the initial synthetic data in a 1:1 ratio, and further semi-supervised iterative training is performed using a layered fine-tuning strategy and weighted cross-entropy loss to gradually improve the model's recognition accuracy and generalization ability in real-world scenarios. After multiple iterations, the quality and quantity of pseudo-label data have been significantly improved, forming a positive feedback loop of "the more data is used, the more accurate it is, and the more the model is trained, the stronger it is", until the model performance reaches the expected indicators and converges.

[0163] This system further includes a multimodal feature fusion strategy, which improves robustness to illumination changes and occlusion by weighted fusion of RGB features and edge features, combined with cross-modal similarity measurement; secondly, it designs a dynamic domain adversarial training architecture, embeds a learnable domain adapter in the classification model, and achieves feature space alignment between synthetic data and real data through progressive adversarial loss weight adjustment; finally, it develops a multi-dimensional dynamic threshold mechanism, which dynamically adjusts the confidence, prediction entropy, and feature similarity quantiles to adaptively balance the quality and quantity of pseudo labels. The specific implementation is as follows:

[0164] 1. Enhance feature discrimination through multimodal feature fusion.

[0165] In the feature space distance filtering stage, a multimodal feature joint alignment strategy is adopted. The specific implementation is as follows:

[0166] Dual-branch feature extraction: A dual-branch network structure is constructed. The main branch extracts RGB image features based on EfficientNet, and the auxiliary branch extracts edge features based on the pre-trained EdgeNet model to capture the contour and structural information of the target.

[0167] Feature fusion and weighted splicing: The feature fusion module performs adaptive weighted fusion on two types of features to generate a fused multimodal feature vector;

[0168] Cross-modal similarity metric: This metric calculates the cosine similarity between the container sub-image and the category center based on fused features, replacing the single RGB feature metric. This design effectively enhances robustness to lighting changes and partial occlusion. For example, in scenes with strong reflections or partial occlusions, edge features can supplement the discriminative information of RGB features to avoid mismatches.

[0169] 2. Optimize cross-domain generalization capabilities through dynamic domain adaptation.

[0170] In order to narrow the distribution gap between synthetic data and real scene data, a dynamic domain adaptation mechanism is introduced in the classification model training:

[0171] Domain adversarial training architecture: A domain classifier is inserted before the fully connected layer of the initial classification model. Through adversarial learning, the feature extractor is forced to generate domain-invariant features. Specifically, the feature extractor needs to simultaneously optimize classification accuracy (distinguishing target classes) and confuse the domain classifier (making it unable to distinguish between synthetic data and real data).

[0172] Dynamic loss weight adjustment: The total loss function is defined as the weighted sum of classification loss and domain adversarial loss, where the weight of the domain adversarial loss increases linearly with the training process. Initially, the classification loss is used to quickly establish basic feature representations. Later, the domain alignment strength is gradually increased to achieve progressive distribution alignment between synthetic and real data.

[0173] An embodiment of the present invention further provides a semi-supervised recognition device for small objects in a container, comprising: at least one memory and at least one processor;

[0174] The at least one memory is configured to store a machine-readable program;

[0175] The at least one processor is used to call the machine-readable program to implement the semi-supervised recognition method for small container targets described in the above embodiment.

[0176] Embodiments of the present invention further provide a computer-readable medium storing computer instructions that, when executed by a processor, cause the processor to perform the semi-supervised recognition method for small container objects described in the aforementioned embodiments. Specifically, a system or device equipped with a storage medium storing software program code that implements the functions of any of the aforementioned embodiments can be provided, and a computer (or CPU or MPU) of the system or device can be configured to read and execute the program code stored in the storage medium.

[0177] In this case, the program code itself read from the storage medium can realize the function of any one of the above-mentioned embodiments, and thus the program code and the storage medium storing the program code constitute part of the present invention.

[0178] Examples of storage media for providing program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, the program code can be downloaded from a server computer via a communication network.

[0179] In addition, it should be clear that the functions of any of the above embodiments can be achieved not only by executing the program code read by the computer, but also by enabling the operating system operating on the computer to complete part or all of the actual operations based on the instructions of the program code.

[0180] In addition, it can be understood that the program code read out from the storage medium is written into the memory provided in the expansion board inserted into the computer or into the memory provided in the expansion unit connected to the computer, and then based on the instructions of the program code, the CPU installed on the expansion board or expansion unit is enabled to perform part or all of the actual operations, thereby realizing the functions of any of the above embodiments.

[0181] The present invention has been shown and described in detail above through the accompanying drawings and preferred embodiments. However, the present invention is not limited to these disclosed embodiments. Based on the above multiple embodiments, those skilled in the art can know that the code review methods in the above different embodiments can be combined to obtain more embodiments of the present invention, and these embodiments are also within the scope of protection of the present invention.

Claims

1. A semi-supervised recognition method for small container targets, characterized by: By integrating a three-dimensional joint filtering mechanism, including confidence screening, uncertainty assessment, and feature space consistency verification, we can filter high-reliability samples from massive amounts of unlabeled container sub-images and construct a progressively optimized training dataset. Use the pre-trained object detection model to perform object detection on the container image, extract the sub-image and input it into the initial classification model to obtain the classification confidence; The Monte Carlo Dropout method is used to quantify the uncertainty of the model's sample classification results. By combining feature space distance filtering with a dynamic threshold adjustment mechanism, samples with high confidence, low uncertainty, and consistent feature space are selected as pseudo-label data. Pseudo-labeled data is mixed with the initial synthetic data, and the generalization ability of the model is gradually improved through semi-supervised iterative training.

2. A semi-supervised recognition method for small container targets according to claim 1, characterized in that: The implementation steps of this method include: 1) Using the original container image as input, EfficientDet performs object detection and sub-image extraction; 2) Training the initial EfficientNet classification model based on enhanced synthetic data: 3) Generate multi-dimensional pseudo labels through three parallel modules: confidence screening, Monte Carlo uncertainty assessment, and feature space consistency verification; 4) Adaptively filter high-quality data based on sample distribution through a dynamic threshold adjustment module; 5) Perform semi-supervised iterative training based on mixed synthetic data and pseudo-labeled data, using layer-wise fine-tuning and weighted loss strategies; 6) Form an iterative optimization closed loop and achieve model self-enhancement through performance convergence judgment.

3. A semi-supervised recognition method for small container targets according to claim 1 or 2, characterized in that: This method is based on multimodal feature fusion and dynamic domain adaptation technology to improve the model's adaptability to complex scenarios; The specific implementation steps include: 1) Add a multimodal feature fusion module to enhance feature discrimination through the RGB+edge feature dual-branch architecture; 2) Introducing a dynamic domain adaptation training mechanism, including domain classifiers and adversarial training strategies; 3) Adaptive weighted concatenation of cross-modal features is achieved through a feature fusion module; 4) Domain adaptation training uses dynamic loss weights to gradually enhance domain-invariant feature learning; 5) The semi-supervised training phase simultaneously optimizes classification accuracy and domain confusion; 6) Multimodal features participate in pseudo-label generation to improve robustness to illumination changes and occlusion.

4. The method for semi-supervised recognition of small objects in a container according to claim 1, characterized in that: The specific implementation of this method is as follows: Step 1: Target detection obtains sub-images of each category, In the object detection phase, the original container image is input. The trained EfficientDet detection model is first used to detect the objects in the image and output the target box coordinates and confidence level. Then, a sub-image is captured from the original image based on the target box coordinates and scaled to a uniform pixel size to meet the input requirements of the subsequent classification model. Step 2: Train the initial classification model, An initial classification and recognition model was trained based on high-definition images of various categories. The training dataset contained multiple categories, with each category containing no more than 10 high-definition images with no or a single background. This dataset was augmented with multi-dimensional data, including spatial and photometric transformations, to simulate the diverse variations found in real-world scenes. The model was then trained using the EfficientNet architecture, employing a cross-entropy loss function and a cosine annealing strategy for decay until the model converged. Step 3: Multi-dimensional pseudo-label generation, including confidence screening, uncertainty assessment, feature space distance filtering, and dynamic threshold adjustment. Step 4: Semi-supervised iterative training, First, the generated pseudo-labeled dataset is mixed with the initial synthetic dataset in a 1:1 ratio to form a new training dataset. Then, a layered fine-tuning strategy is adopted to freeze the underlying convolutional layers of EfficientNet to retain the basic feature extraction capabilities, and only fine-tune the fully connected layers to adapt to the new data distribution, while using a weighted cross-entropy loss function for optimization. Finally, by continuously monitoring the accuracy of the validation set, iterative training is terminated when the accuracy improvement in three consecutive rounds of training is less than 0.5%. Step 5: Iterative optimization and self-enhancement, After completing steps 1 to 4, use the currently trained optimized classification model as the new base model and repeat steps 3 and 4 to continuously update the pseudo-label data and iteratively optimize the semi-supervised training. The details are as follows: The new basic model makes predictions on a large number of unlabeled container sub-images, re-executes the multi-dimensional pseudo-label generation process, and dynamically adjusts the threshold according to the latest sample distribution to generate updated high-quality pseudo-label data; then, the updated pseudo-label data is re-mixed with the initial synthetic data in a 1:1 ratio, and further semi-supervised iterative training is performed using a hierarchical fine-tuning strategy and weighted cross-entropy loss.

5. The method for semi-supervised recognition of small objects in a container according to claim 4, characterized in that: The multi-dimensional pseudo label generation is specifically implemented as follows: (1) Confidence screening: The intercepted container sub-image is input into the initial classification model, and the model outputs a multi-dimensional probability vector, where each element represents the predicted probability that the sub-image belongs to the current category; Calculate the maximum confidence in the probability vector and determine its corresponding category label; (2) Uncertainty assessment: Using the Monte Carlo Dropout method, the Dropout layer is enabled during classification model inference. Ten forward propagations are performed on the same subgraph to obtain 10 probability vectors, each containing the predicted probability values of multiple categories. Based on the results of these 10 forward propagations, the average predicted probability of each category is calculated, and the predicted entropy is further calculated. The entropy value is used to quantify the uncertainty of the model's sample classification results. (3) Feature space distance filtering: First, we extract the feature vectors of all samples based on the initial training set and calculate the feature mean of each category as the feature center of that category. Then, we calculate the cosine similarity between the feature vector of the container sub-image and the feature centers of all categories, and take the maximum similarity value as the matching degree between the sub-image and the closest category. (4) Dynamic threshold adjustment: First, the distribution of confidence, entropy and feature similarity of all candidate samples in the current round is statistically analyzed; then, the threshold is dynamically updated according to the quantile, and the confidence threshold is set to the top 90% quantile of the current sample confidence distribution, the entropy threshold is set to the top 10% quantile of the current sample entropy distribution, and the feature similarity threshold is set to the top 90% quantile of the current sample similarity distribution; finally, only the samples that simultaneously meet the conditions of confidence greater than or equal to the updated confidence threshold, entropy less than the updated entropy threshold, and feature similarity greater than or equal to the updated similarity threshold are retained as the final pseudo-label data.

6. The method for semi-supervised recognition of small objects in a container according to claim 3, characterized in that: The multimodal feature fusion and dynamic domain adaptation technology are used to improve the model's adaptability to complex scenarios. The multimodal feature fusion is used to enhance feature discrimination. In the feature space distance filtering stage, a multimodal feature joint alignment strategy is adopted. The specific implementation is as follows: Dual-branch feature extraction: A dual-branch network structure is constructed. The main branch extracts RGB image features based on EfficientNet, and the auxiliary branch extracts edge features based on the pre-trained EdgeNet model to capture the contour and structural information of the target. Feature fusion and weighted splicing: The feature fusion module performs adaptive weighted fusion on two types of features to generate a fused multimodal feature vector; Cross-modal similarity measurement: Calculate the cosine similarity between the container sub-image and the category center based on the fused features, replacing the single RGB feature measurement.

7. The method for semi-supervised recognition of small objects in a container according to claim 3, characterized in that: The multimodal feature fusion and dynamic domain adaptation technologies are used to improve the model's adaptability to complex scenarios. Dynamic domain adaptation is used to optimize cross-domain generalization capabilities. Dynamic domain adaptation is introduced into classification model training to narrow the distribution gap between synthetic data and real-world scene data. These include: Domain adversarial training architecture: A domain classifier is inserted before the fully connected layer of the initial classification model. Through an adversarial learning strategy, the feature extractor is forced to generate domain-invariant features. The feature extractor needs to simultaneously optimize classification accuracy and confuse the domain classifier. Dynamic loss weight adjustment: The total loss function is defined as the weighted sum of classification loss and domain adversarial loss, where the weight of the domain adversarial loss increases linearly with the training process. In the initial stage, classification loss is mainly used to quickly establish basic feature representation. In the later stage, the domain alignment strength is gradually enhanced to achieve progressive distribution alignment between synthetic data and real data.

8. A semi-supervised recognition system for small objects in containers, characterized by: include: Target detection and sub-image acquisition module, used to complete target detection and sub-image interception; Initial classification model training module, which trains the initial classification model based on enhanced synthetic data; Multi-dimensional pseudo-label generation module generates multi-dimensional pseudo-labels based on confidence screening, uncertainty assessment, and feature space consistency verification; Dynamic threshold adjustment module, used to adaptively filter high-quality data based on sample distribution; Semi-supervised iterative training, which is used to implement semi-supervised iterative training based on mixed synthetic data and pseudo-labeled data; Iterative optimization module forms an iterative optimization closed loop and realizes model self-enhancement through performance convergence judgment; The system specifically implements semi-supervised recognition of small targets in containers through the method described in any one of claims 1 to 7.

9. A semi-supervised recognition device for small objects in a container, characterized in that: include: at least one memory and at least one processor; The at least one memory is configured to store a machine-readable program; The at least one processor is configured to call the machine-readable program to implement the method according to any one of claims 1 to 7.

10. A computer-readable medium, characterized in that The computer readable medium stores computer instructions, which, when executed by a processor, can implement the method according to any one of claims 1 to 7.

Images