Semi-supervised learning-based solid waste detection model construction method and solid waste detection method

Abstract

The application discloses a kind of based on semi-supervised learning's solid waste detection model construction method and solid waste detection method, solid waste detection model includes generator and discriminator, generator is based on the input labeled image feature map and unlabeled image feature map and exports prediction result, discriminator is based on the prediction result of generator and the real label of corresponding image feature map and exports consistency discrimination result, based on the sample set combination of labeled image feature map and unlabeled image feature map, through the confrontation training of discriminator and generator makes discriminator learn to the discrimination knowledge of prediction result and real label, generator learns to the ability of identifying solid waste target based on the input image feature map.Extract unmarked data information by semi-supervised learning mechanism comprehensive generator and discriminator, use fusion loss to regulate training and realize pixel-level prediction, the low demand of label makes it have higher application value and popularization, can effectively improve the recall rate of existing method.

Description

Technical Field

[0001] This invention belongs to the field of solid waste detection technology, specifically relating to a method for constructing a solid waste detection model based on semi-supervised learning and a solid waste detection method. Background Technology

[0002] Statistics show that the world generates over 10 billion tons of waste annually, with my country accounting for approximately 1 billion tons, covering 750,000 mu (approximately 56,667 hectares). This figure is increasing at a rate of 10% per year. Scenes of garbage "besieging cities" and solid waste "filling the sea" are commonplace. In particular, the illegal dumping of solid waste by some enterprises for profit is serious, often choosing to secretly discharge pollutants at night in secluded areas. This untreated solid waste, through natural degradation and rainwater erosion, leachate and harmful chemicals damage soil structure and pollute nearby rivers and groundwater systems. Traditional solid waste monitoring relies mainly on manual inspections and video surveillance: surveillance cameras are costly to deploy and have limited monitoring areas; manual inspections are prone to blind spots, are difficult to fully cover due to the large number of locations, and are inefficient. The detection methods urgently need to be changed.

[0003] Deep learning research in the field of solid waste detection is still in its infancy. The shortcomings of existing solid waste detection model construction and detection methods using deep learning are: (1) The spectral features, morphological features and texture features of solid waste are complex and diverse. At the same time, the background complexity of remote sensing image distribution greatly increases the difficulty of learning robust and discriminative representations from scenes and objects in deep learning, resulting in insufficient extraction ability and weak robustness of existing models; (2) The amount of training data is small. The edges of solid waste are complex and the labeling cost is high. Therefore, high-quality training images are generally lacking. Under such circumstances, the effect of using a small amount of training data to apply deep learning methods is poor and it is difficult to obtain relatively accurate prediction results. Summary of the Invention

[0004] The purpose of this invention is to provide a method for constructing a solid waste detection model based on semi-supervised learning;

[0005] Another objective of this invention is to provide a solid waste detection method based on semi-supervised learning.

[0006] This semi-supervised learning-based solid waste detection model construction method integrates generator and discriminator to extract unlabeled data information through a semi-supervised learning mechanism, and uses fusion loss to standardize training to achieve pixel-level prediction;

[0007] The semi-supervised learning-based solid waste detection method has high application value and scalability due to its low demand for tags. The modification method proposed in this invention can effectively improve the recall rate of existing methods.

[0008] To achieve the above-mentioned objectives, the technical solution of this invention is as follows:

[0009] A method for constructing a solid waste detection model based on semi-supervised learning is disclosed. The solid waste detection model includes a generator and a discriminator. The generator outputs a prediction result based on the input labeled and unlabeled image feature maps. The discriminator outputs a consistency discrimination result based on the generator's prediction result and the real labels of the corresponding image feature maps. Based on a combination of sample sets of labeled and unlabeled image feature maps, the discriminator learns the discrimination knowledge between the prediction result and the real label through adversarial training between the discriminator and the generator, while the generator learns the ability to identify solid waste targets based on the input image feature maps.

[0010] This invention is the first to apply semi-supervised learning to solid waste detection. The generator is similar to a Siamese network model, while the discriminator, by accepting segmentation results or image labels as input, performs a higher-order consistency measure between predicted and real samples. Therefore, the solid waste identification results have an accuracy advantage compared to other models, without requiring a large number of high-quality training images. This invention is preferably applicable to urban solid waste detection.

[0011] In the above-mentioned method for constructing a solid waste detection model based on semi-supervised learning, the generator includes an encoder, a decoder, and an attention module. The encoder extracts features from the image feature map, and the attention module remotely transmits the features from the encoder to the decoder. The decoder uses a combination of convolution and deconvolution to locate the features into their original spatial positions and outputs the prediction results.

[0012] This invention designs an attention mechanism (SA) to optimize features and remote information transmission. By enhancing the model's ability to learn deep features of solid waste through feature self-constraint, it achieves more efficient feature representation learning and remote information transmission. Through visualization analysis of intermediate features, compared with the skip connection method of other models, this model significantly improves feature representation ability.

[0013] In the above-mentioned method for constructing a solid waste detection model based on semi-supervised learning, the generator extracts features through an encoder with shared parameters, and then passes the extracted features through the attention module.

[0014] The semi-supervised learning method for building solid waste detection models allows the model to automatically utilize unlabeled samples to improve learning performance without relying on external interactions. Since similar solid waste targets generally have similar features, features can be extracted by using a generator through an encoder with shared parameters.

[0015] In the above-mentioned method for constructing a solid waste detection model based on semi-supervised learning, the attention mechanism includes three branches.

[0016] Branch 1 is used for explicit modeling and feature interaction in the channel dimension to obtain the adjustment results for each channel;

[0017] Branch 2 is used to integrate information from the input values ​​in the local spatial neighborhood;

[0018] Branch 3 is used to focus on the overall spatial correlation based on the comprehensive channel adjustment and the integration of local spatial neighborhood information, and to obtain new scaled features.

[0019] The attention mechanism in this invention provides relatively accurate edge recognition results, making it suitable for extracting solid waste with diverse shapes. It does not confuse the background with the foreground target and has good differentiation of backgrounds such as buildings and green spaces. It also exhibits excellent performance with the same small amount of labeled data.

[0020] In the above-mentioned method for constructing a solid waste detection model based on semi-supervised learning, branch one first convolves the input features, then performs global pooling compression to obtain a vector. This vector represents the global distribution of the response on the feature channels. Weights are generated for each channel through convolution and adjusted using the Gelu activation function. Finally, the adjusted results for each channel are obtained by multiplying the convolutional features with the input features.

[0021] Branch 2 captures the edge gradients of solid waste in relevant directions from horizontally and vertically distributed convolution kernels through asymmetric convolution, and the BN layer pulls the feature distribution output by the convolution module back to a standard normal distribution with a mean between [0, 1], so as to integrate the information of the input values ​​in the local spatial neighborhood.

[0022] Branch 3 first convolves the input features, then performs global average pooling and max pooling to obtain channel descriptions. It then obtains spatial weight coefficients through convolutional layers and the Gelu activation function. Finally, it multiplies the weight coefficients with the superposition of the outputs of Branch 1 and Branch 2 to obtain the scaled new features.

[0023] Preferably, branch one is used to complete explicit modeling and feature interaction in the channel dimension. First, the input features are convolved with 3*3, and then global pooling is performed to compress them along the spatial dimension to obtain a 1*1*C vector. This vector represents the global distribution of the response in the feature channels. Weights are generated for each channel through 1*1 convolution and adjusted using the Gelu activation function. Finally, the adjusted results for each channel are obtained by multiplying the convolutional features with the input features.

[0024] Branch 2 integrates information from the input values ​​in the local spatial neighborhood. Asymmetric convolution, using horizontally and vertically distributed kernels, captures the gradients of solid waste edges in relevant directions, improving edge extraction accuracy. This complements the features of conventional 3x3 convolution. The combination of the three receptive field convolutions, each with different emphases, effectively complements each other. The BN layer pulls the feature distribution output by the convolutional module back to a standard normal distribution with a mean between [0, 1], minimizing the differences in the Gelu activation function outputs and preventing gradient vanishing.

[0025] Branch 3 focuses on overall spatial correlation based on comprehensive channel adjustment and local neighborhood information. First, it performs 3*3 convolution on the input features, and then performs global average pooling and max pooling along the channel dimension to obtain channel descriptions. It then performs a convolutional layer and Gelu activation function to obtain spatial weight coefficients. Finally, it multiplies the weight coefficients with the superposition results of the outputs of Branch 1 and Branch 2 to obtain the scaled new features.

[0026] In the above-mentioned method for constructing a solid waste detection model based on semi-supervised learning, the lowest feature scale of the encoder network is 512*8*8, and the attention module is used for feature transfer in the encoder scales of 8, 16, 32, and 64 respectively.

[0027] The discriminator module introduces the Unet structure and uses a numerical superposition method for remote information transmission.

[0028] The discriminator network of this invention accepts segmentation results or image labels as input and performs a higher-order consistency measurement between predicted and real samples. By introducing a Unet structure, the channel stacking method of the original model is changed to a numerical stacking method during long-distance information transmission, making the feature decoding process closer to residual learning. The discriminator generates pixel-level uncertainty maps / confidence maps to measure the quality of the generator-generated samples, implicitly forcing the semantic segmentation network to generate prediction results that are closer to the desired mask overall. After a certain training process, the discriminator will learn the discrimination knowledge between the generator's prediction results and the real labels.

[0029] In the aforementioned method for constructing a solid waste detection model based on semi-supervised learning, the solid waste detection model further includes an edge extraction module located at the output of the prediction results. This module uses the Canny operator to convert the areal results of the generator's predictions and the true label map into an edge map, and then extracts the edge loss function L. line Describe the differences between the two.

[0030] Due to the extreme imbalance between the edge lines and the background, this module does not perform parameter updates during backpropagation.

[0031] In the above-mentioned method for constructing a solid waste detection model based on semi-supervised learning, the loss function L for generator training is...model as follows:

[0032] L model =λ sup L sup +λ semi L semi +λ line L line +L res ,

[0033] L sup For generator supervision loss, L semi To combat learning loss, L line To extract the edge loss of the results, L res This is for the loss of the extracted result area.

[0034] The detection model of this invention uses a fusion loss to standardize training and achieve pixel-level prediction. Preferably, λ sup , λ semi , λ line The values ​​were set to 0.2, 0.2, and 0.1 to train the solid waste detection model.

[0035] In the above-mentioned method for constructing a solid waste detection model based on semi-supervised learning, the loss function for training this model includes the generator supervision loss L. sup Adversity learning loss L semi Extraction result region loss L res The edge loss of the extraction result L line And the discriminator discrimination loss L dis ,in,

[0036]

[0037]

[0038] This represents the prediction result for class c of unlabeled samples, where the superscript u indicates no label, c represents the number of classification categories, and Y is the true label.

[0039]

[0040] These are confidence graphs of the predicted result R and the true label Y in the discriminator network at positions i and j, respectively, where y represents the judgment result of whether it was generated by the generator.

[0041]

[0042] L line FocalLoss with reshaped standard cross-entropy loss:

[0043] FL(pt )=-(1-p t ) γ log(p t ),

[0044] P t It is the predicted probability of being classified as solid waste, and γ is the adjustment factor.

[0045] L sup This represents the supervision loss term applied by the generator based on labeled samples, which is calculated from the cross-entropy of the labeled mask and the generator's prediction data.

[0046] L semi The adversarial learning loss consists of a discriminator loss term for both labeled and unlabeled samples. In this case, the unlabeled samples' predictions in the generator can be treated as pseudo-labels in the discriminator. By combining the adversarial loss and the supervised loss, the generator is jointly trained to maximize the confidence of the predicted samples learned from the object, thereby deceiving the discriminator.

[0047] A solid waste detection method based on semi-supervised learning includes a solid waste detection model constructed by any of the methods described above, and a trained solid waste detection model obtained by training the solid waste detection model using a sample set including labeled image feature maps and unlabeled image feature maps based on the semi-supervised learning method. The trained solid waste detection model identifies solid waste targets based on the input image feature maps to be identified.

[0048] This invention employs a semi-supervised learning method to train models when high-quality training images are lacking and only a small amount of training data is available for deep learning, thereby effectively improving the recall rate of the model.

[0049] Preferably, the data is a 256*256 image. The generator module extracts features by inputting two feature maps of different regions. One of the two feature maps is labeled data, and the other is unlabeled data. The remote information transmission uses an attention module for feature transmission and / or uses a numerical superposition method for feature decoding during the remote information transmission process.

[0050] Furthermore, remote sensing images were used as the data source with a spatial resolution of 0.038m, and the solid waste targets selected in the images should be larger than 0.5 square meters.

[0051] The solid waste detection method provided by this invention improves the ability of deep learning to learn robust and discriminative representations from scenes and objects. The solid waste identification results have an advantage in accuracy compared with other models, and the method has a low requirement for labels, thus having high application value and scalability.

[0052] Compared with the prior art, the beneficial effects of the present invention are reflected in:

[0053] (1) This invention is the first to apply semi-supervised learning to solid waste detection. The generator is similar to the Siamese Net model, and the discriminator takes the segmentation results or image labels as input and performs a higher-order consistency measurement between the predicted samples and the real samples. Therefore, the solid waste identification results have an advantage in accuracy compared with other models, without the need for a large number of high-quality training images.

[0054] (2) This invention designs an attention mechanism SA to optimize features and remote information transmission. By enhancing the model’s ability to learn deep features of solid waste through feature self-constraint, it achieves more efficient feature representation learning and remote information transmission. Through the visualization analysis of intermediate features, compared with the skip connection method of other models, this model significantly improves the feature representation ability.

[0055] (3) The attention mechanism in this invention has relatively accurate edge recognition results, which is suitable for solid waste extraction with diverse shapes. It will not confuse the background with the foreground target, and has good differentiation of backgrounds such as buildings and green spaces. It has excellent performance under the same small amount of labeled dataset.

[0056] (4) The discriminator of this invention accepts segmentation results or image labels as input and performs a higher-order consistency measurement between predicted samples and real samples. By introducing the Unet structure, the channel stacking method of the original model is changed to a numerical stacking method during remote information transmission, making the feature decoding process closer to residual learning. The discriminator generates pixel-level uncertainty maps / confidence maps to measure the quality of the samples generated by the generator, implicitly forcing the semantic segmentation network to generate prediction results that are closer to the desired mask. After a certain training process, the discriminator will learn the discrimination knowledge between the generator's prediction results and the real labels.

[0057] (5) In the case of lack of high-quality training images and only a small amount of training data for deep learning, the present invention adopts a semi-supervised learning method to train the model, thereby effectively improving the recall rate of the model.

[0058] (6) The solid waste detection method provided by this invention improves the ability of deep learning to learn robust and discriminative representations from scenes and objects. The solid waste identification results have an advantage in accuracy compared with other models. Moreover, this method has a low requirement for labels, so it has high application value and promotion potential. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the Semi-SD network structure;

[0060] Figure 2This is a schematic diagram of the attention mechanism.

[0061] Figure 3 This is a schematic diagram illustrating the relationship of the loss function;

[0062] Figure 4 Example image for the training set;

[0063] Figure 5 A schematic diagram illustrating the results of remote feature extraction for different models;

[0064] Figure 6 Example figures showing solid waste identification results for different models and detection methods. Detailed Implementation

[0065] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0066] like Figure 1 This embodiment provides a method for constructing a solid waste detection model based on semi-supervised learning and a solid waste detection model based on semi-supervised learning (Semi-supervised Slide Detection, semi-SD). The model includes a generator and a discriminator, employing ResNet34 as the feature extractor. The lowest feature scale of the network is 512*8*8. An attention mechanism is designed to achieve efficient feature transfer at scales of 8, 16, 32, and 64. In the decoder, a combination of convolutional and deconvolutional layers locates the features to their original spatial positions, outputting pixel-by-pixel prediction results. Figure 1 The purple block represents the output of the SA structure, which is summed with the orange block to serve as remote transmission information. In this invention, the discriminator network uses a Unet structure, as shown in the second part of the network diagram. After a certain training process, the discriminator will learn the knowledge to distinguish between the generator's predictions and the true labels.

[0067] like Figure 2 As shown, the generator uses a Self-Attention module (SA) for remote information transmission. First, a 3*3 convolution is performed on the input features, followed by global pooling to compress them along the spatial dimension to obtain a 1*1*C vector. This vector represents the global distribution of the response on the feature channels. Weights are generated for each channel through 1*1 convolution and adjusted using the Gelu activation function. Finally, the adjusted results for each channel are obtained by multiplying the convolutional features with the input features.

[0068] Branch 2 uses asymmetric convolution to capture the gradients of solid waste edges in relevant directions through horizontally and vertically distributed convolution kernels, improving edge extraction accuracy. It complements the features of conventional 3x3 convolution. The combination of the three receptive field convolutions with different focuses can effectively complement each other. The BN layer can pull the feature distribution output by the convolution module back to a standard normal distribution with a mean between [0, 1], making the output differences of the Gelu activation function smaller and avoiding gradient vanishing.

[0069] Branch 3 first performs a 3*3 convolution on the input features, then performs global average pooling and max pooling along the channel dimension to obtain channel descriptions, performs a convolutional layer and Gelu activation function to obtain spatial weight coefficients, and finally multiplies the weight coefficients with the superposition of the outputs of Branch 1 and Branch 2 to obtain the scaled new features.

[0070] This model is constructed using fusion loss for standardized training. The model loss includes 5 classes, with the discriminator discrimination loss L. dis Generator supervision loss L sup Adversity learning loss L semi Extraction result region loss L res The edge loss of the extraction result L line The relationship between the various loss functions is shown in the graph below. Figure 3 As shown,

[0071] L dis For discriminator loss, Here, R represents the prediction graph and Y represents the confidence graphs of the discriminator network at positions i and j, respectively, and y represents the judgment result of whether it was generated by the generator.

[0072]

[0073] L sup This represents the supervision loss term applied by the generator based on the labeled samples:

[0074]

[0075] L semi For adversarial learning loss, This represents the prediction result for class c of unlabeled samples. The superscript u indicates no label, and c indicates the number of classification categories. In this example, there are 2 categories: solid waste and non-solid waste.

[0076]

[0077] L res It is the cross-entropy loss between the model segmentation result and the true label, R i，j，c Y represents the prediction result for class c of the labeled sample. i，j，c This represents the true value of class c for the labeled sample.

[0078]

[0079] L line For edge loss, L line The Focal Loss, which reshapes the standard cross-entropy loss, is used as the calculation formula, P t γ is the predicted probability of being classified as solid waste, and γ is an adjustment factor, which is set to 0.5 in this embodiment.

[0080] L line =-(1-P t ) γ log(P t )

[0081] The overall loss function L for generator training model for:

[0082] L model =λ sup L sup +λ semi L semi +λ line L line +L res

[0083] λ sup , λ semi , λ line These are the hyperparameter weights of each loss function, which are set to 0.2, 0.2, and 0.1 in this embodiment, and the model is trained using the above 5 types of losses.

[0084] SEnet and Unet are typical models of remote information transmission, namely attention mechanism and skip connection, respectively. Solid waste detection experiments were conducted using SEnet, Unet and the model provided in this embodiment.

[0085] Using local UAV imagery of a certain area collected in June 2021 as the data source, with a spatial resolution of 0.038m, the imagery was cropped into 256*256 blocks. Solid waste targets larger than 0.5 square meters in the imagery were sampled to finely delineate the distribution boundaries of the solid waste. After allocation, the training set consisted of 100 labeled images and 867 unlabeled images, while the test and validation sets each contained 150 images. An example of the dataset is shown below. Figure 4 .

[0086] The three methods described above were used to recover 64*64 intermediate layer feature maps of remote and reconstructed information during feature decoding. Nine channels were randomly selected from the multi-channel feature maps of four images for feature comparison. The results are as follows. Figure 5 As shown.

[0087] Depend on Figure 5As can be seen, regarding feature sparsity, UNet has pure black, informationless feature channels, while the features extracted in this embodiment contain information in each channel. Regarding feature representation capability, FCHarDNet contains a lot of irregular noise, making it difficult to use for feature learning. Overall, the model and method provided in this embodiment have higher feature representation capability, which is helpful for long-distance information transmission.

[0088] Furthermore, this embodiment compares the recognition performance of our proposed method (abbreviated as Semi-CD) with Unet, PSPNet, FCHARDNet, SegFormer, DeepLabV3+, Res34, and their corresponding semi-supervised modifications (name prefixed with Semi-). The experimental results are analyzed from both quantitative and qualitative perspectives below:

[0089] Quantitative analysis

[0090] Table 1 shows the performance data of the aforementioned models. As can be seen from Table 1, all the compared methods achieve an accuracy of over 80%. The main difference in model performance lies in recall. SegFormer has the lowest recall at only 60%, while Semi-Res34 has the highest at 79.3%. Looking at the comprehensive index F1, DeepLabV3+ and Res34 perform better, with F1 values ​​around 80%. The generator proposed in this invention has an F1 score more than 4 percentage points higher than other methods, achieving a recall of 84.8%, thus outperforming other existing models. A comparison of semi-supervised modifications to various methods shows that the proposed modification method can effectively improve the performance of each method.

[0091] Table 1. Performance data of recognition results for different models and detection methods.

[0092]

[0093] Qualitative analysis

[0094] like Figure 6 The image shows example images of solid waste identification results using different models and detection methods. Region 1 shows scattered solid waste piles. It can be seen that Unet and PSPNet exhibit significant over-extraction, and the extraction results of each comparison method differ greatly from the actual labels. Only the method of this invention provides relatively accurate edge identification results, indicating that the method of this invention is more suitable for the extraction of solid waste with diverse forms.

[0095] Regions 2 and 3 contain solid waste surrounding buildings. SegFormer and Unet tend to confuse background with foreground objects, while the method of this invention can better distinguish between buildings, green areas, and other background elements. Overall, the method proposed in this invention outperforms the comparative methods on the same small amount of labeled dataset.

[0096] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for constructing a solid waste detection model based on semi-supervised learning, characterized in that, The solid waste detection model includes a generator and a discriminator. The generator outputs a prediction result based on the input labeled and unlabeled image feature maps. The discriminator outputs a consistency discrimination result based on the generator's prediction result and the real labels of the corresponding image feature maps. Based on the combination of sample sets of labeled and unlabeled image feature maps, the discriminator learns the discrimination knowledge between the prediction result and the real label through adversarial training between the discriminator and the generator, and the generator learns the ability to identify solid waste targets based on the input image feature maps. The discriminator module introduces a Unet structure and uses a numerical superposition method for remote information transmission; The generator includes an encoder, a decoder, and an attention module. The encoder extracts features from the image feature map. The attention module remotely transmits the features from the encoder to the decoder. The decoder uses a combination of convolution and deconvolution to locate the features into their original spatial positions and outputs the prediction results. The attention module includes three branches: Branch 1 is used for explicit modeling and feature interaction in the channel dimension to obtain the adjustment results for each channel; Branch 2 is used to integrate information from the input values ​​in the local spatial neighborhood; Branch 3 is used to focus on the overall spatial correlation based on the comprehensive channel adjustment and the integration of local spatial neighborhood information, and to obtain new scaled features.

2. The method for constructing a solid waste detection model based on semi-supervised learning as described in claim 1, characterized in that, The generator extracts features through an encoder with shared parameters, and then passes the extracted features through the attention module.

3. The method for constructing a solid waste detection model based on semi-supervised learning as described in claim 1, characterized in that, Branch 1 first convolves the input features, then performs global pooling to compress them into a vector. This vector represents the global distribution of the response on the feature channels. Weights are generated for each channel through convolution and adjusted using the Gelu activation function. Finally, the adjusted results for each channel are obtained by multiplying the convolutional features with the input features. Branch 2 captures the edge gradients of solid waste in relevant directions from horizontally and vertically distributed convolution kernels through asymmetric convolution, and the BN layer pulls the feature distribution output by the convolution module back to a standard normal distribution with a mean between [0,1], so as to integrate the information of the input values ​​in the local spatial neighborhood. Branch 3 first convolves the input features, then performs global average pooling and max pooling to obtain channel descriptions. It then obtains spatial weight coefficients through convolutional layers and the Gelu activation function. Finally, it multiplies the weight coefficients with the superposition of the outputs of Branch 1 and Branch 2 to obtain the scaled new features.

4. The method for constructing a solid waste detection model based on semi-supervised learning as described in claim 1, characterized in that, The lowest feature scale of the encoder network is 512*8*8, and the attention module is used for feature transfer in the encoder scales of 8, 16, 32, and 64 respectively.

5. The method for constructing a solid waste detection model based on semi-supervised learning as described in claim 1, characterized in that, The solid waste detection model also includes an edge extraction module located at the output of the prediction results. The Canny operator is used to convert the areal results of the generator's prediction results and the real label map into edge maps, and the difference between the two is described by the edge loss function of the extracted results.

6. The method for constructing a solid waste detection model based on semi-supervised learning as described in claim 5, characterized in that, Loss function for generator training as follows: ， , , These are the hyperparameter weights of each loss function; For generator supervision loss, To combat learning loss, To extract the edge loss of the results, This is for the loss of the extracted result area.

7. The method for constructing a solid waste detection model based on semi-supervised learning as described in claim 6, characterized in that, The loss function for training this model includes the generator supervision loss mentioned above. Combating learning loss Extraction result region loss Edge loss of extraction results And the discriminator identifies the loss. ,in, ， ， This represents the prediction result for class c of unlabeled samples, where the superscript u indicates no label, c indicates the number of classification categories, and Y is the true label. ， , The graphs are the prediction result R and the true label Y in the discriminator network at positions i and j, respectively, and y represents the judgment result of whether it was generated by the generator. ， FocalLoss with reshaped standard cross-entropy loss: ， It is the predicted probability of being classified as solid waste, and γ is the adjustment factor.

8. A solid waste detection method based on semi-supervised learning, characterized in that, The method includes a solid waste detection model constructed by any one of claims 1-7, and a trained solid waste detection model is obtained by training the solid waste detection model with a sample set including labeled image feature maps and unlabeled image feature maps based on a semi-supervised learning method. The trained solid waste detection model identifies solid waste targets based on the input image feature maps to be identified.

Images