Data denoising method based on multi-scale reliable clustering

Abstract

The present invention discloses a data denoising method based on multi-scale reliable clustering. The present invention proposes a single-category denoising algorithm for the early processing of data in the commodity recognition model, aiming to solve the problem of dirty and messy training data for deep learning. The algorithm clusters each category to find out the discrete points outside the cluster, that is, the noise data and remove them, to achieve an end-to-end data automation cleaning process, save the time of manual data cleaning, and greatly improve the work efficiency of data processing. The experimental results show that the accuracy of the network trained with the denoised data set is higher than the accuracy obtained by training the network with the original data, which also reflects the good performance of the denoising method from the side.

Description

Technical Field

[0001] The invention belongs to the field of image processing and computer vision, and in particular relates to a data denoising method based on a scale-reliable clustering method. Background Art

[0002] In recent years, the DBSCAN-based clustering method has been widely used in the pedestrian re-identification data processing step. The purpose is to achieve unsupervised learning through clustering. In addition, it can also be used for data cleaning. However, due to the uneven distribution of sample density during clustering and the large difference in distance between classes, the clustering effect is often very poor. The most widely used density-based clustering method is DBSCAN, among which the Heterogencous DBSCAN (HDBSCAN) algorithm is used to overcome the shortcomings of the original DBSCAN algorithm in the case of unbalanced density, and the adaptability of the HDBSCAN algorithm to unbalanced data is enhanced by defining two parameters, LongRegionQuery and SHortRegionQuery; the Enhanced DBSCAN (EDBSCAN) algorithm is also used to solve the problem of low clustering accuracy in the case of uneven data density. By calculating the distance density between different center points and all points, if its density is less than the preset threshold, the center point is used for the next clustering, otherwise the point is assigned to the boundary point. In order to improve the operating efficiency of the algorithm, L-DBSCAN is proposed as a comprehensive and efficient algorithm. This algorithm selects a suitable clustering method based on the characteristics of the data set to improve the efficiency of the algorithm. This method is suitable for clustering large-scale data sets. Summary of the invention

[0003] The problem to be solved by the present invention is that in the data processing before deep learning training, since the training data is dirty and there is a large amount of noise data, it will seriously affect the accuracy of network training, and the existing methods can only screen the data by manual cleaning, which cannot efficiently realize the data processing. It is necessary to propose a new data denoising method based on clustering, which uses an algorithm to clean the data on a large scale.

[0004] In order to solve the above problems, the present invention provides a data denoising method based on multi-scale reliable clustering. By clustering all data, the noise data at discrete points are found and removed, saving the time of manual data selection, thereby achieving the purpose of data cleaning. Experimental results show that the network trained with denoised data can obtain higher recognition accuracy and improve the network's expression ability. Figure 1 It can be seen that the data denoising method based on multi-scale reliable clustering includes the following steps:

[0005] Step 1, data feature extraction: The data selected for this experiment is the product recognition data used in the CVPR 2020Al i Products Challenge: Large-scale Product Recogn it ion competition, with a total of more than 50,000 SKU categories and 3 million data. First, we pass all the data through our pre-trained product recognition network model, extract features for each data, save its features before the FC fully connected layer, convert them into 1*2048 tensors, generate corresponding labels at the same time, and save them as .npy format data files.

[0006] Step 2, algorithm clustering. We use the DBSCAN algorithm as the clustering algorithm and make improvements on it. Based on step 1, perform a DBSCAN clustering algorithm on the data in each SKU category. In the returned results, the labels of non-discrete point data are between 0 and positive infinity, representing the category number of each cluster, while the label of discrete point noise data is -1.

[0007] Step 3: Initial noise detection. Through the clustering method in step 2, we obtain the data index with label -1, and then obtain the storage path of each data. According to the path, each discrete data is deleted, thus completing the data denoising process.

[0008] Step 4, deep refinement. In the process of initial noise detection, if the data base of this category is too large, the clustering effect will decrease, and 40% of the data may be judged as noise data, which contains a large number of correct samples, thus affecting the distribution of training data. Therefore, based on the initial noise detection, we make an accurate judgment on each piece of data again, in order to correct the misclassified noise data. First, we calculate the average of the features of all data in each small category after clustering the category, as a feature of a query library, and then compare each piece of noise data that has been judged as noise data with the features in the query library. If the similarity between the two reaches more than 80%, it can be determined that the noise data is misjudged and will not be deleted during the denoising process. By making such a secondary judgment on each category of noise data, we can greatly reduce the misjudgment rate.

[0009] As a further preferred embodiment, the specific steps of step 2) are:

[0010] In the clustering process, we use the dbscan clustering algorithm. Compared with kmeans clustering, dbscan does not require any parameters and is applicable to convex and non-convex samples. It can cluster dense data sets of any shape. dbscan has four important parameters: density direct, density reach, eps and minsample. Figure 2 As shown, density direct access means that object A is within the domain of object B; density reachable means that there is an object chain ABC, AB has direct density access, BC has direct density access, and AC has density reachable; eps is the domain radius when defining density; minsample is the threshold of the core point. When clustering, first traverse the data, starting from a core object p, find all density-connected points of the core object, add the density-connected points to the cluster where p is located, find the core objects in the direct density-reachable points of p, and add the density-connected points of these objects to p, and repeat this recursive operation until it can no longer be expanded. Finally, the result returned by the clustering algorithm is a string of numbers between -1 and positive infinity, which is the same length as the original data. The number at each position represents the category into which the data corresponding to the index is classified after clustering. We only need to find the pictures with the corresponding clustering value of -1 in these data to complete the denoising process. As shown in Figure 2 As shown in the figure, A represents the core object, BC represents the boundary point, and N represents the outlier point. Point N is the noise data we found through clustering.

[0011] In order to verify the denoising performance of the clustering algorithm, we selected 20 types of header data and 20 types of tail data from the Ali Products ChallengeDatasets commodity identification dataset, a total of 40 categories and 8,632 images for denoising experiments. There is 10% noise data in each category of these data. We use the dbscan clustering algorithm to cluster these data, and by adjusting the two thresholds of eps and min_sample, we can find and remove the noise data, thereby cleaning the data. The following is the detailed information of the experiment.

[0012]

[0013] It can be seen from the table that when the dbscan clustering algorithm is used to experiment on the data, about 90% of the noise data can be removed. However, at the same time, due to the complexity of the joint adjustment of the two parameters eps and min_sample in the clustering process of the clustering algorithm, and the limitations of the algorithm itself, some noise data will be missed in the clustering process. Therefore, we explore algorithms with better and more reliable clustering effects.

[0014] We propose a reliable clustering method based on multi-scales. First, we define two parameters: cluster independence and cluster compactness.

[0015] Independence of clusters. A reliable clustering should be independent of other clusters and individual samples. Intuitively, if a cluster is far away from other samples, it can be considered to be highly independent. However, due to the uneven density of the latent space, we cannot use the distance between the cluster centroid and the samples outside the cluster to measure cluster independence. In general, the clustering results can be adjusted by changing some hyperparameters of the clustering criterion. The clustering criterion can be relaxed so that each cluster may contain more samples, or it can be tightened so that each cluster may contain fewer samples. We use Represents a cluster In the samples clustered using the dbscan clustering algorithm, we propose the following metric to measure cluster independence, which is expressed as the intersection-over-union score.

[0016]

[0017] From formula 1, we can see that When the clustering criterion becomes loose, The cluster set of The bigger, The more independent the clusters, the more independent the clusters; that is, if one relaxes the clustering criteria, no more new samples will be included in the new clusters. Figure 3 From the upper part, we can see that if the independence of a cluster is not good, after relaxing the clustering criteria, some clusters that do not belong to this category will be mistakenly classified into this category, which will change the clustering shape of the cluster and affect the clustering performance.

[0018] Compactness of clusters. A reliable cluster should also be compact, that is, the samples in the same cluster should have a small distance between samples. In extreme cases, when a cluster is the most compact, the distance between all samples in the cluster is zero, and even if the clustering criteria are tightened, its samples will not be divided into different clusters. Based on this assumption, we can define the following metric to determine the compactness within a cluster:

[0019]

[0020] When the clustering criterion becomes strict The cluster set of The larger the After clustering, the distance between samples around is smaller. Figure 3 From the lower part, we can see that when stricter standards are adopted, clusters with larger distances between samples contain fewer points, thus dividing clusters that originally belonged to one category into two or more categories. We call this poor compactness of clusters.

[0021] Based on the above indicators for measuring cluster reliability, we can calculate the independence and compactness scores of each data point in the cluster, so as to make more detailed judgments on each data point and find out the missed noise data.

[0022] In order to verify the denoising performance of the multi-scale reliable clustering algorithm, we set up a comparative experiment using the same data set. From the experimental results in the table below, we can see that for the head data, we are able to remove all the noise data, and we also have a good performance in the denoising process of the tail data.

[0023]

[0024] Through this comparative experiment, it can be seen that the method based on multi-scale reliable clustering is much better than other clustering methods. Finally, we decided to use the improved dbscan clustering algorithm for denoising.

[0025] The core technologies of this patent include:

[0026] (1) Adopt a new clustering idea and use the multi-scale reliable clustering method to improve the accuracy of data clustering.

[0027] (2) An end-to-end data automation cleaning model was built to achieve automated data cleaning, which can remove 90% of the noise data.

[0028] (3) The data was pre-processed using data cleaning methods and then the model was trained, which improved the recognition accuracy of the model by one percent. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a flow chart of the data denoising method based on multi-scale reliable clustering of the present invention.

[0030] Figure 2 It is a schematic diagram of the dbscan clustering algorithm principle of the present invention

[0031] Figure 3 This is a schematic diagram of the principle of the multi-scale reliable clustering algorithm of the present invention. DETAILED DESCRIPTION

[0032] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0033] We downloaded more than 50,000 categories of data sets from the CVPR 2020 All Products Challenge: Large-scale Product Recognition competition, and randomly selected 20 categories of head data and 20 categories of tail data. We extracted features from these 40 categories of 8,632 data on our recognition model and saved the labels and paths of these data in .npy format.

[0034] After obtaining these features, we use a multi-scale reliable clustering algorithm to cluster them. Based on the jitter of the independence and compactness of each data after clustering, we determine whether the stability of the clustering is good, thereby filtering out more noise data and obtaining very clean training data.

[0035] Without changing the training method, we used this method to perform overall denoising on the training data of our product recognition model. As can be seen from the chart, the accuracy of model recognition was finally improved by one percentage point.

[0036]

Claims

1. Data denoising algorithm based on multi-scale reliable clustering, characterized by: It is divided into the following four steps: Step 1: Data feature extraction: All data is passed through a pre-trained product recognition network. The model extracts features from each piece of data, saves its features before the FC fully connected layer, converts them into 1*2048 tensors, and generates corresponding labels at the same time, saving them. Save as a data file in .npy format; Step 2: Algorithm clustering: Use the dbscan algorithm as the clustering algorithm for each sku The data within the category is clustered using the dbscan algorithm. In the returned results, the non-discrete The labels of the point data are between 0 and positive infinity, representing the category number of each cluster, while the label of the noise data of the discrete point is -1; Step 3: Initial noise detection; using the clustering method in step 2, obtain the data with label -1. Then obtain the storage path of each data and store each discrete data according to the path. The data is deleted, thus completing the process of data denoising; Step 4, depth refinement; On the basis of the initial noise inspection, each piece of data is accurately judged again in order to correct the misclassified noise data; first, the features of all the data in each sub-category after the category is clustered are averaged as the features of a query library, and then each piece of noise data that has been judged as noise data is compared with the features in the query library. If the similarity between the two reaches more than 80%, it is judged as normal data and will not be deleted in the denoising process; otherwise, it is judged as noise data and deleted; Specifically, we first define two parameters: cluster independence and cluster compactness; Independence of clusters; Represents a cluster In the samples clustered using the dbscan clustering algorithm, the following metric is proposed to measure cluster independence, which is expressed as the score of the intersection-over-union ratio; When the clustering criterion becomes loose, The cluster set of The bigger, The more independent the clusters are; that is, if one relaxes the clustering criteria, no more new samples will be included in the new clusters; Compactness of clusters; define the following metric to determine compactness within a cluster: When the clustering criterion becomes strict The cluster set of The larger the The smaller the distance between samples around clustering; Based on the above indicators for measuring cluster reliability, the independence and compactness scores of each data point in the cluster are calculated, so that each data point can be judged more carefully and the missed noise data can be found.

2. The data denoising method based on multi-scale reliable clustering according to claim 1, characterized in that: In step 2, when clustering, first traverse the data, starting from a core object p, find all the density-connected points of the core object, add the density-connected points to the cluster where p is located, find the core objects among the directly density-reachable points of p, and add the density-connected points of these objects to p. Repeat this recursive operation until it can no longer be expanded. Finally, the result returned by the clustering algorithm is a string of numbers between -1 and positive infinity, which is the same length as the original data. The number at each position represents the category into which the data corresponding to the index is classified after clustering. As long as the images with the corresponding clustering value of -1 in these data are found, the denoising process is completed.

Images