Noise label learning method in visual target detection

Abstract

The embodiment of the invention discloses a noise label learning method in visual target detection, which mainly utilizes random sampling and a Bayesian function to obtain the posterior probability of corresponding classification of an image, and selects an optimal parameter for training according to the size of a loss value. Compared with a traditional method, the label noise filtering method based on probability distribution estimation has the advantages that under the condition that the accuracy is basically kept, the algorithm execution time is greatly shortened, and meanwhile the problems that under the condition that a large amount of data containing noise labels are used for training, the model performance can be greatly reduced, and misjudgment is caused are solved.

Description

Technical Field

[0001] The present invention relates to a data classification task learning method, and in particular to a noise label learning method in visual target detection. Background Art

[0002] In today's computer vision field, many tasks such as image classification, object detection, semantic segmentation, etc. show excellent performance in deep learning. The fundamental reasons are divided into three aspects: First, the continuous improvement of computer hardware system equipment, especially in graphics processing units and parallel processing. Second, the technology of network parameter initialization and training methods has made great progress. However, the above two situations need data support to show their good performance. The third reason is that there are large-scale datasets with accurate annotations such as ImageNet, which can more conveniently train network parameters efficiently and relatively accurately.

[0003] In the real world, it is relatively easy to collect datasets with noisy labels, but collecting large-scale clean datasets is expensive and time-consuming. There are two reasons for this: On the one hand, expert-labeled data is necessary for certain research, such as medical data, which must be accurately labeled by experts due to its professionalism; on the other hand, it is impossible to manually label a large dataset like JFT300M, which contains 300 million images.

[0004] In the field of machine learning, it is very common to use labeled data to train neural networks to achieve classification, regression or other learning purposes. However, in related model learning, if data containing a lot of erroneous information is used, it will affect the accuracy of the final prediction model.

[0005] Therefore, more and more scholars have begun to study noisy data, which is usually divided into two categories: feature noise and noise labels.

[0006] Specifically, feature noise refers to the deviation between the training sample instance feature itself and its true feature, for example: artificially adding Gaussian noise to the existing feature; noisy label refers to the deviation between the target label used for training and the true label of the corresponding instance itself, such as setting the label incorrectly when setting the label. Both types of noise will lead to a decline in system performance, but relevant studies have shown that noisy labels are more harmful. For example, in expression recognition, the uncertainty from high-quality facial expressions to low-quality micro-expressions will lead to inconsistent labels or even wrong labels, which will ultimately reduce the final classification performance of the system; in the field of medical image analysis, the degree of lesions and the corresponding locations have different causes, and slightly deviated labels will affect the final diagnosis and treatment results. In semantic segmentation, due to unreasonable labeling rules, the training model is not effective, etc. These all reflect the importance of dealing with noisy labels.

[0007] Noisy labels are a complex phenomenon that can be caused by a variety of reasons, which can be divided into four categories: (1) Insufficient information during the annotation process to fully describe the category data; (2) Due to poor data quality, the recognition is reduced, making it impossible for experts to correctly label the corresponding tags during the annotation process; (3) Due to different subjectivities, that is, not all labels are annotated by experts. This is because reliable labeling is a time-consuming and expensive task, so some labels are obtained from non-experts, and different annotators label labels from different perspectives and obtain inconsistent results; (4) Sample labels may also be incorrect due to data encoding or communication problems.

[0008] However, in large data sets, when the data sources are complex and the data volume is large, the above situation is difficult to completely avoid. To this end, the present invention provides a training method that can reduce the impact of noise labels on training results and has high efficiency and higher classification accuracy in response to the above defects. Summary of the invention

[0009] In order to solve the above technical problems, an embodiment of the present invention provides a noise label learning method in visual target detection, which specifically includes the following steps: Step 1: Input image data, extract image features through a deep neural network, input it into the classifier to get the classification result, and get the classification probability through the Softmax function layer.

[0010] Step 2: Extract a certain number of samples and their corresponding labels from the data set, and have experts manually label these samples. The re-labeled labels are considered correct labels, and the correct labels are compared with the original labels to obtain a probability matrix. Step 3: Input the probability matrix and classification probability as parameters into the Bayesian network adaptive layer to calculate the probability of a certain sample category being mislabeled, and use this to estimate the posterior probability of any sample in the data set being mislabeled; Step 4: Use the posterior probability obtained in step 3 together with the image data as input parameters of the loss value calculation function to calculate the loss value LOSS; Step 5: Select the loss value corresponding to the combination with the smallest loss value to update the neural network (including the Bayesian network adaptive layer), thereby achieving robust training with higher classification accuracy.

[0011] Among them, the loss value calculation function described in step 4 is:

[0012] in, Indicates the number of channels, and Represents the height and width of the image respectively, and The predicted value and the true value are Channel No. The value of pixels. Indicates Channel No. The posterior probability of pixels is is a positive number, Indicates Gaussian kernel function at pixel The weight at can be defined as:

[0013]

[0014] in, represents the number of Gaussian kernel functions, Indicates Gaussian kernel function at pixel The weight at can be defined as:

[0015]

[0016] in, Represents pixels With The distance between the center points of the Gaussian kernel functions can be calculated using different distance measurement methods as needed. Indicates The standard deviation of the Gaussian kernel function controls the size of the Gaussian kernel function.

[0017] The Bayesian network in step 3 is composed of nodes representing variables and directed edges connecting these nodes. The update of the Bayesian network parameters adopts random variational inference, and its formula is:

[0018] is the marginal likelihood function of the observed data x, represents KL divergence, which is used to measure the similarity between two probability distributions, and ELBO is the lower bound of evidence. is the posterior distribution, is a hidden variable, is the approximate posterior distribution.

[0019] The implementation of the embodiments of the present invention has the following beneficial effects: ① This training method combines random sampling and Bayesian networks, and directly performs robust training on a target detection data set containing noise labels to obtain a deep learning network model with good detection performance, thereby achieving efficient and high-accuracy training. ② This solution improves the traditional loss value function, reduces the impact of noise labels on training effects through parameters such as posterior probability, and improves training efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a schematic block diagram of the steps of the present invention; Figure 2 It is the network structure diagram and the overall network flow chart of the present invention; Figure 3 is a composition diagram of the Bayesian network adaptive layer of the present invention; Figure 4 This is a comparison chart of the running time of this solution and the traditional method. DETAILED DESCRIPTION

[0021] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings.

[0022] Example 1 A noise label learning method for visual object detection specifically includes the following steps: Step 1: Input image data, extract image features through a deep neural network, input it into the classifier to get the classification result, and get the classification probability through the Softmax function layer.

[0023] Step 2: Extract a certain number of samples and their corresponding labels from the data set, and have experts manually label these samples. The re-labeled labels are considered correct labels, and the correct labels are compared with the original labels to obtain a probability matrix. Step 3: Input the probability matrix and classification probability as parameters into the Bayesian network adaptive layer to calculate the probability of a certain sample category being mislabeled, and use this to estimate the posterior probability of any sample in the data set being mislabeled; Step 4: Use the posterior probability obtained in step 3 together with the image data as input parameters of the loss value calculation function to calculate the loss value LOSS; Step 5: Select the loss value corresponding to the combination with the smallest loss value to update the weight parameters of the deep neural network and the Bayesian network adaptive layer, thereby achieving robust training with higher classification accuracy.

[0024] like Figure 1 As shown, due to the existence of noisy labels, the labels in the dataset are unreliable. , according to its unreliable label Select the posterior probability matrix corresponding to The output prediction of DNN is obtained, and the loss value calculation function is:

[0025]

[0026] Network flow chart Figure 2 As shown, the overall operation process is described as follows: input image , extract features through deep neural network, and then input them into the classifier to get the results. The function obtains the posterior classification probability.

[0027] After obtaining the classification probability of each category, it is input into an adaptive layer based on a Bayesian network. The adaptive layer outputs the probability corresponding to the category of the image being incorrectly labeled. At this time, it is used together with the noise label as the input of the loss function to calculate the loss value of the sample, and finally random variational inference is used to update the model parameters.

[0028] The Bayesian network adaptive layer is as follows Figure 3 As shown, the Bayesian network in step 3 is composed of nodes representing variables and directed edges connecting these nodes. At the same time, the update of the Bayesian network parameters adopts random variational inference, and its formula is:

[0029] in, is the marginal likelihood function of the observed data x, represents KL divergence, which is used to measure the similarity between two probability distributions, and ELBO is the lower bound of evidence. is the posterior distribution, is a hidden variable, is the approximate posterior distribution.

[0030] The parameter updating scheme of this embodiment does not use the EM algorithm, thus avoiding the shortcomings of the EM algorithm. , the probability is Compared with the previously proposed research scheme using nonlinear layers, this scheme takes advantage of the graph data structure, reduces the number of parameters in the adaptive layer, alleviates the overfitting problem to a certain extent, and the network structure is easy to implement and can be embedded in other computer vision tasks, thus improving versatility.

[0031] Example 2 This embodiment discloses a method for optimizing a loss function based on Embodiment 1, and describes the loss function in detail. In the traditional weighted Dice loss function, the weight of a pixel can be set according to custom rules such as the position of the pixel, the attributes of the pixel, etc., so as to achieve different weighting methods. For example, weighting can be performed based on factors such as the distance from the pixel to the boundary, the probability distribution of the pixel, and the complexity of the pixel.

[0032] Since the weighted Dice loss function can better focus on important areas in the segmentation task while maintaining the edge information of the image, it has broad application prospects in image segmentation tasks.

[0033] The improved weighted Dice loss function uses a weighted Gaussian function based on the inverse proportional function of the distance to dynamically adjust the weight of the pixel. Its mathematical description is as follows:

[0034] in, Indicates the number of channels, and Represents the height and width of the image respectively, and The predicted value and the true value are Channel No. The value of pixels. Indicates Channel No. The posterior probability of pixels.

[0035] The formula derivation process is as follows. When only a single channel is considered, the basic form of the formula is:

[0036] in, and represent the predicted value and the true value respectively.

[0037] Considering the importance of each pixel, a weight is introduced into the basic form of the formula to obtain the form of the weighted coefficient:

[0038] Considering the existence of noisy labels, in order to achieve robust training, Multiply by the corresponding posterior probability , reducing the impact of noisy labels on model robustness through uncertainty:

[0039]

[0040] In the process, in order to prevent calculation errors caused by the denominator being 0, a small positive number is added to both the numerator and the denominator:

[0041] Finally, we get the loss function of a single channel, for all channels , then:

[0042]

[0043] The improved weighted Dice loss function can better capture the important pixel areas in the image segmentation task and pay more attention to the detailed features of the target area while retaining the image edge information.

[0044] Example 3 This embodiment mainly illustrates the beneficial effects of the solution of the present invention through training on benchmark datasets: CIFAR-10, CIFAR-100, CIFAR-10N and CIFAR-100N.

[0045] The CIFAR-10 and CIFAR-100 datasets are the most widely used in the literature for learning with noisy labels. The labels of these two datasets are very good and there are no noisy labels, so we add synthetic noise to the training set using the settings described in CurriculumNet, Co-teaching, and Topological, and the test dataset does not have any noisy labels added. The synthetic noise settings are as follows: Symmetric noise setting: With probability p, each sample in the training set is randomly assigned to a uniform random label other than its true label. In the experiments, p is set to 10%, 20%, and 40%.

[0046] Pairwise noise setting: Samples in one class can only be mislabeled to the same one in other classes. This experiment used the same noise transition matrix in Co-teachin and set the probability of samples in one class being mislabeled to 10%, 20%, and 40%.

[0047] In CIFAR-10N, 60.27% of the training images get consistent labels. The noise rates of the five prepared noisy label sets are 9.03% (aggregate), 17.23% (random 1), 18.12% (random 2), 17.64% (random 3), and 40.21% (worst).

[0048] In addition, for the CIFAR-100N dataset, each image contains a coarse and fine label. The overall noise level of the coarse and fine labels is 25.60% and 40.20%.

[0049] The specific experimental results are shown in Table 1 and Table 2: Table 1 Comparison of classification accuracy on CIFAR-10

[0050] Table 2 Comparison of classification accuracy on CIFAR-100

[0051] Table 3 Comparison of classification accuracy on CIFAR-10N

[0052] As shown in Table 1, Table 2, Table 3 and Figure 4 As shown in the figure, the experimental results show that the model obtained by robust training after adding synthetic noise to the CIFAR-10 and CIFAR-100 datasets has a roughly 25% reduction in algorithm execution time and an improvement in classification accuracy of about 2% compared to traditional methods, while maintaining basically the same accuracy.

[0053] After achieving robust training on CIFAR-10N and CIFAR-100N, the algorithm’s classification accuracy is improved by about 1% compared to other methods.

[0054] Of course, the above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any modifications made according to the spirit of the main technical solution of the present invention should be included in the protection scope of the present invention.

Claims

1. A noise label learning method for visual object detection, characterized in that: The specific steps include: Step 1: Input image data, extract image features through a deep neural network, input it into the classifier to get the classification result, and get the classification probability through the Softmax function layer. Step 2: Extract a certain number of samples and their corresponding labels from the data set, and have experts manually label these samples. The re-labeled labels are considered correct labels, and the correct labels are compared with the original labels to obtain a probability matrix. Step 3: Input the probability matrix and classification probability as parameters into the Bayesian network adaptive layer, calculate the probability of a certain sample category being mislabeled, and use this to estimate the posterior probability of any sample in the data set being mislabeled; Step 4: Use the posterior probability obtained in step 3 together with the image data as input parameters of the loss value calculation function to calculate the loss value LOSS; Step 5: Select the loss value corresponding to the combination with the smallest loss value to update the weight parameters of the deep neural network and the Bayesian network adaptive layer, thereby achieving robust training with higher classification accuracy.

2. The noise label learning method in visual target detection according to claim 1, characterized in that: The loss value calculation function described in step 4 is: in, Indicates the number of channels, and Represents the height and width of the image respectively, and The predicted value and the true value are Channel No. The value of pixels. Indicates Channel No. The posterior probability of pixels is is a positive number, Indicates Gaussian kernel function at pixel The weight at can be defined as: in, represents the number of Gaussian kernel functions, Indicates Gaussian kernel function at pixel The weight at can be defined as: in, Represents pixels With The distance between the center points of the Gaussian kernel functions can be calculated using different distance measurement methods as needed. Indicates The standard deviation of the Gaussian kernel function controls the size of the Gaussian kernel function.

3. The noise label learning method in visual target detection according to claim 1, characterized in that: The Bayesian network in step 3 is composed of nodes representing variables and directed edges connecting these nodes.

4. The noise label learning method in visual target detection according to claim 1, characterized in that: The update of the Bayesian network parameters adopts stochastic variational inference, and its formula is: in, is the marginal likelihood function of the observed data x, represents KL divergence, which is used to measure the similarity between two probability distributions, and ELBO is the lower bound of evidence. is the posterior distribution, is a hidden variable, is the approximate posterior distribution.

Images