CN117689646A

CN117689646A - High-precision defect detection method, system and medium for positive and negative sample fusion

Info

Publication number: CN117689646A
Application number: CN202311741054.9A
Authority: CN
Inventors: 胡峥楠; 曾祥瑞; 尹周平
Original assignee: Huazhong University of Science and Technology
Current assignee: Huazhong University of Science and Technology
Priority date: 2023-12-18
Filing date: 2023-12-18
Publication date: 2024-03-12

Abstract

The invention discloses a high-precision defect detection method, a high-precision defect detection system and a high-precision defect detection medium for positive and negative sample fusion, and belongs to the technical field of intelligent manufacturing. The method comprises the following steps: training the supervised model by using a negative sample set formed by the X-ray gray level images with defects to obtain a target detection model; dividing the areas of the defect-free X-ray gray level image to obtain each area block image; respectively training an unsupervised model by using a positive sample set formed by segmented images of the same region to obtain an outlier detection model corresponding to each region; and filtering and fusing the two detection results to obtain a final defect detection result. The invention combines two different detection methods and strategies, utilizes the advantages of supervised learning and outlier detection, further improves the accuracy and reliability of detection through rule screening and fusion, and can help the manufacturing industry to improve the product quality, reduce the cost and improve the competitiveness.

Description

High-precision defect detection method, system and medium for positive and negative sample fusion

Technical Field

The invention belongs to the technical field of intelligent manufacturing, and particularly relates to a high-precision defect detection method, a high-precision defect detection system and a high-precision defect detection medium for positive and negative sample fusion.

Background

Artificial intelligence technology has very important applications in various links of intelligent manufacturing. Defect detection is an important application scenario for intelligent manufacturing technology. The industrial defect detection means that defects in a certain product are detected in the industrial production process, so that the defects are repaired or eliminated in time, and the product quality and the production efficiency are ensured. The artificial intelligence technology has important significance in the application of industrial defect detection, and can reduce the cost of artificial detection. As production scales gradually expand, manual inspection has become difficult to meet. The addition of the artificial intelligence technology can perfect an automatic process and improve the detection efficiency and accuracy.

Among them, X-ray defect detection is a common industrial defect detection method. The X-rays generated in the X-ray detecting device are mainly used for irradiating the inside of the product. Because the X-ray penetrating power is very strong, the X-ray can penetrate the inside of the product to form images, and the defect conditions are clear. Moreover, the X-ray is used for not damaging the product, so that the method is a nondestructive detection mode. After photographing by using the X-ray detection device, a clear gray image reflecting the internal defects can be obtained, so that manual/machine screening can be performed. But the X-ray gray level image is a single-channel image, has fewer characteristics and is difficult to train to obtain a high-precision defect detection model.

Disclosure of Invention

Aiming at the defects or improvement demands of the prior art, the invention provides a high-precision defect detection method, a system and a medium for positive and negative sample fusion, which solve the technical problems that the characteristic of an X-ray gray image is single and a high-precision defect detection model is difficult to train by combining a supervised learning method based on a negative sample and an unsupervised learning method based on a positive sample.

In order to achieve the above object, according to a first aspect of the present invention, there is provided a high-precision defect detection method for positive and negative sample fusion, including:

acquiring a plurality of X-ray gray images with defects and no defects;

training the supervised model by using a negative sample set formed by the X-ray gray level images with defects to obtain a target detection model;

dividing the areas of the defect-free X-ray gray level image to obtain each area block image; respectively training an unsupervised model by using a positive sample set formed by segmented images of the same region to obtain an outlier detection model corresponding to each region;

inputting an X-ray gray level image to be detected into the target detection model to obtain a first detection result;

dividing the X-ray gray level image to be detected into areas, and respectively inputting the segmented images of each area into corresponding outlier detection models to obtain a plurality of second detection results;

and filtering and fusing the first detection result and the plurality of second detection results to obtain a final defect detection result.

According to a second aspect of the present invention, there is provided a high-precision defect detection method for positive and negative sample fusion, comprising:

acquiring a plurality of X-ray gray images with defects and no defects;

screening out a plurality of problem areas of the X-ray gray level image to be detected according to the first detection result; respectively inputting each problem area into a corresponding outlier detection model to obtain a plurality of second detection results;

Further, the screening out a plurality of problem areas of the X-ray gray scale image to be detected according to the first detection result includes:

obtaining a detection result of each area block image of the X-ray gray level image to be detected according to the first detection result; and setting a detection threshold, and screening out areas with indexes not reaching standards according to the detection threshold to obtain a plurality of problem areas of the X-ray gray level image to be detected.

According to a third aspect of the present invention, there is provided a high-precision defect detection method for positive and negative sample fusion, comprising:

acquiring a plurality of X-ray gray images with defects and no defects;

judging whether the X-ray gray level image to be detected has a problem area or not according to the first detection result;

if the problem exists, respectively inputting the problem areas into corresponding outlier detection models to obtain a plurality of second detection results; filtering and fusing the first detection result and the plurality of second detection results to obtain a final defect detection result;

and if the defect detection result does not exist, taking the first detection result as a final defect detection result.

Further, the determining, according to the first detection result, whether the X-ray gray scale image to be detected has a problem area includes:

obtaining a detection result of each area block image of the X-ray gray level image to be detected according to the first detection result; setting a detection threshold, and judging that a problem-free area exists if indexes of all areas reach the standard; otherwise, judging the problem areas, and screening out the areas with indexes not reaching the standard according to the detection threshold value to obtain a plurality of problem areas of the X-ray gray level image to be detected.

According to a fourth aspect of the present invention, there is provided a high-precision defect detection system for positive and negative sample fusion, comprising:

a processor;

a memory storing a computer executable program which, when executed by the processor, causes the processor to perform a high precision defect detection method of positive and negative sample fusion as described in the first to third aspects.

According to a fifth aspect of the present invention, there is provided a computer-readable storage medium having stored thereon a computer program which, when executed by a processor, implements a high-precision defect detection method of positive and negative sample fusion as described in the first to third aspects.

In general, the above technical solutions conceived by the present invention, compared with the prior art, enable the following beneficial effects to be obtained:

(1) Aiming at the problem of single characteristic of an X-ray gray image, the invention combines a supervised learning method based on a negative sample and an unsupervised learning method based on a positive sample, wherein the position of the defect can be detected by using a supervised target mode; by comparing the model with the positive sample pattern obtained by the model, the abnormal pattern which does not appear is detected and positioned, and the actual detection range can be enlarged. Finally, the two detection results are screened and filtered, and then are fused, so that the two detection schemes are effectively complemented, the detection reliability is improved, and the false alarm rate is reduced.

(2) The invention can realize low omission and even zero omission under the condition of permission of false detection rate. This is an extremely important goal in industrial manufacturing, since it ensures product quality while also improving production efficiency. The defect detection system comprehensively designed in the invention combines two different detection methods and strategies, utilizes the advantages of supervised learning and outlier detection, and further improves the accuracy and reliability of detection through rule screening and fusion. The system has wide application prospect in industrial production, can help the manufacturing industry to improve the product quality, reduce the cost and improve the competitiveness. By reducing the occurrence of rejects, the industrial defect detection system can improve the efficiency and reliability of the overall supply chain, thereby positively impacting the manufacturing industry.

Drawings

FIG. 1 is a diagram showing an overall frame of a high-precision defect detection method for positive and negative sample fusion according to an embodiment of the present invention;

fig. 2 is a flowchart of an object detection network according to an embodiment of the present invention.

Detailed Description

The present invention will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention. In addition, the technical features of the embodiments of the present invention described below may be combined with each other as long as they do not collide with each other.

In the present invention, the terms "first," "second," and the like in the description and in the drawings, if any, are used for distinguishing between similar objects and not necessarily for describing a particular sequential or chronological order.

Currently, from the perspective of deep learning, defect detection technologies can be broadly classified into two categories, i.e., supervised learning and unsupervised learning, and the defects in the product (sample) are detected and located by learning on a sample with defects (negative sample) and a normal sample without defects (positive sample). The supervised target detection method based on negative sample learning can obtain more accurate defect detection results when the number of defect labeling samples is large and the labeling quality is high; the unsupervised abnormal detection method based on positive sample learning can obtain more complete defect detection results under the condition that the number of defect labeling samples is small and the number of normal samples is large. For this application, the invention designs a specific solution from the perspective of deep learning. The invention combines two detection schemes of a supervised defect detection method based on a negative sample and an unsupervised learning method based on a positive sample, reasonably fuses respective detection results, and obtains a high-precision detection result.

Referring to fig. 1, the present embodiment provides a high-precision defect detection method for positive and negative sample fusion, which includes operations S1 to S4.

Operation S1, a plurality of X-ray gray level images with defects and without defects are acquired.

The first step in industrial defect detection is data sort. This stage involves collecting, collating and preparing image data for training and testing the defect detection model. Such data typically comes from actual shots taken on the factory line, including images of various products or components, some of which may be defective and some of which are normal, in this embodiment, the X-rays generated in the X-ray inspection apparatus are used to illuminate the interior of the product. The importance of data sorting cannot be underestimated because it directly affects the performance of the model. First, the data collection needs to take into account diversity. This means that images from different working conditions, different angles and different lighting conditions are collected. This helps ensure that the model can be accurately detected in various situations. In addition, the size of the data is also important, and more data generally means better performance. Thus, it is necessary to ensure collection of a sufficiently large-scale data set. Data sorting also involves preprocessing of the image. This includes resizing the image, color normalization, and removing unnecessary noise. These preprocessing steps help to improve training efficiency and performance of the model.

In the data sort process, it is necessary to mark whether each image contains defects and the type of defects so that the model can learn how to identify different types of defects when training. In the steps of defining defect categories and manually labeling, the present invention has intensively studied different types of defects and assigned specific labels. First, a defect classification system needs to be established. This includes identifying and defining different types of defects, such as cracks, flaws, scratches, etc. These categories are established to ensure that the model can accurately identify and classify different types of defects, thereby improving the accuracy of the detection. Then, for the defined defect class, manual labeling is required. The practitioner will scrutinize the image and mark the location and type of defect on the image. These annotation data will be key information required to train the supervised learning model. At this stage, the accuracy and consistency of the labels is critical, as they directly affect the performance of the model.

And S2, training the supervised model by using a negative sample set formed by the X-ray gray level image with the defects to obtain a target detection model.

With the noted data, training of the defect detection model can be undertaken. Typically, this model is a deep learning based object detection network. The goal of such a network is to automatically detect defects in the image and locate their position. The key to this step is to select the appropriate deep learning architecture. The model will understand the visual characteristics of the different defect types by learning the marked data and establish a pattern of defect identification. Training of the model is an iterative process that requires constant adjustments to network parameters to improve performance. In the training process, using the marked positive sample data, the model learns how to accurately identify and locate defects. In addition, some negative sample data is needed, with no defects in the images, to help the model distinguish between normal and abnormal situations. The invention adopts a target detection network (YOLO) for defect detection, and the specific steps are shown in fig. 2.

1. Defects, i.e., the morphology of the defects and their corresponding categories, are predefined.

2. Image data with defects are collected according to given defect definition and marked manually.

3. After the data set is acquired, the data set is sent to a target detection network for defect detection. Typically, to achieve fast detection, an end-to-end target detection network such as YOLO may be used.

YOLO has end-to-end detection capability, enabling real-time detection. In the model training process, image data with defects is used as a training set, and the model is supervised by using manually marked information. The model will learn to identify the different classes of defects and be able to locate them accurately in the image. The training of the model requires a lot of computing resources and time, but once the training is completed, it can be run efficiently in an actual production environment, enabling fast and accurate defect detection.

S3, carrying out region division on the defect-free X-ray gray level image to obtain each region block image; and training the unsupervised model by using a positive sample set formed by the segmented images of the same region to obtain an outlier detection model corresponding to each region.

Unlike object detection, the task of the outlier detection network is to identify abnormal defects in the image, i.e. those that do not belong to the previously defined defect class, thereby increasing the actual detection range. This step requires the use of a positive sample learning based approach. This network will learn the normal characteristics of the industrial product or component. These features may relate to aspects of the shape, color, texture, etc. of the product. Once the network learns of these normal features, it can detect abnormal situations that do not coincide with normal situations. In this process, positive sample data that does not contain defects is used to train the network. The network will strive to capture patterns under normal conditions and identify anomalies at the time of testing.

By using the current advanced unsupervised anomaly detection method, a normal mode different from the defect sample is established by modeling the feature distribution of the normal sample. When an industrial image to be detected passes through the detection system, the following processes are carried out:

1. extracting image features by using a deep neural network, and obtaining an image feature map;

2. comparing the Marshall distance between the characteristic vector of the current sample and the corresponding positive sample distribution for each spatial position of the characteristic image to obtain an abnormal score map consistent with the size of the input image;

3. normalizing the anomaly score graph and setting a threshold value. The specific mode of normalization is as follows:

wherein Xmin is the score minimum of the score map and Xmax is the maximum score value in the score map. The score plot after normalization is between 0 and 1. After the normalized anomaly graphs are obtained, setting a threshold value super-parameter. The anomaly score for each location in the anomaly graph is compared to the threshold value, and an anomaly is defined as exceeding the threshold value. In order to expand the number of positive samples and reduce potential problems such as small target missed detection caused by super-large resolution image input when positive sample data are acquired, the invention divides an original industrial image into image blocks with uniform sizes, and models the distribution of the respective positive samples and the detection of abnormal points are carried out on a single image block.

S4, inputting an X-ray gray level image to be detected into the target detection model to obtain a first detection result; dividing the X-ray gray level image to be detected into areas, and respectively inputting the segmented images of each area into corresponding outlier detection models to obtain a plurality of second detection results; and filtering and fusing the first detection result and the plurality of second detection results to obtain a final defect detection result.

After two stages of detection, namely target detection and outlier detection, two groups of defect detection results are obtained. These results may overlap or may be complementary. In this step, a set of rules needs to be formulated to filter and fuse the results to obtain the final defect detection output. Rules may include threshold settings, area requirements, shape matching, etc. to determine whether a defect is considered a true defect, and may also take into account the trustworthiness of the different inspection methods. By applying these rules, false positives or uncertain results can be eliminated, thereby improving overall detection accuracy. In addition, result fusion is required to combine two different detection results into one comprehensive result. The fusion aims to ensure that the two detection schemes can be effectively complemented and improve the accuracy of the whole detection. The comprehensive method not only increases the reliability of detection, but also reduces the false alarm rate. In addition, the filtering and fusion process typically needs to be tuned and optimized for the specific application scenario.

Since the defect detection method based on positive sample learning and the method based on negative sample learning are to predict defects from two different angles. The reasonable fusion rule is set, so that the respective advantages of the two methods can be combined, and the defect detection with higher precision is achieved.

In consideration of the fact that the supervised learning can obtain better defect positioning effect under the condition of sufficient marking data and high marking quality, the invention can also adopt the following fusion scheme:

1. target detection network prediction using negative sample learning;

2. screening out the common problem area of the network according to the prediction, and setting the common problem area as a key area;

3. separately training an outlier detection model based on positive sample learning in a key area;

4. the filtering and fusing rules are set manually, the advantage intervals of different methods are played, and the defect omission rate is reduced.

Based on this, operation S4 may be replaced with:

s4', inputting an X-ray gray level image to be detected into the target detection model to obtain a first detection result; screening out a plurality of problem areas of the X-ray gray level image to be detected according to the first detection result; respectively inputting each problem area into a corresponding outlier detection model to obtain a plurality of second detection results; and filtering and fusing the first detection result and the plurality of second detection results to obtain a final defect detection result.

According to the first detection result, a plurality of problem areas of the X-ray gray level image to be detected are screened out, specifically:

Or, operation S4' ″ inputs the X-ray gray scale image to be detected into the target detection model to obtain a first detection result; judging whether the X-ray gray level image to be detected has a problem area or not according to the first detection result; if the problem exists, respectively inputting the problem areas into corresponding outlier detection models to obtain a plurality of second detection results; filtering and fusing the first detection result and the plurality of second detection results to obtain a final defect detection result; and if the defect detection result does not exist, taking the first detection result as a final defect detection result.

Judging whether the X-ray gray level image to be detected has a problem area according to the first detection result, wherein the problem area is specifically as follows:

Through the steps, the method can obtain the high-precision defect labeling image. These images show the location and type of defects detected on the industrial product or component. These annotation images are important for subsequent quality control, product improvement and production optimization. Summarizing, industrial defect detection is a complex and critical process that requires careful data preparation, training of deep learning models, and filtering fusion of results. High-precision defect detection can be realized only when the details and the accuracy are fully considered in each step, so that the quality and the efficiency of industrial production are improved.

The high-precision defect detection system for positive and negative sample fusion provided by the invention can realize low omission ratio under the condition of permission of false detection rate. This means that the system can efficiently identify and locate defects while minimizing false alarms. This is critical to quality control in an industrial process because it helps to improve product quality, reduce reject rate, and thus improve production efficiency and reliability.

It will be readily appreciated by those skilled in the art that the foregoing description is merely a preferred embodiment of the invention and is not intended to limit the invention, but any modifications, equivalents, improvements or alternatives falling within the spirit and principles of the invention are intended to be included within the scope of the invention.

Claims

1. The high-precision defect detection method for positive and negative sample fusion is characterized by comprising the following steps of:

acquiring a plurality of X-ray gray images with defects and no defects;

2. The high-precision defect detection method for positive and negative sample fusion is characterized by comprising the following steps of:

acquiring a plurality of X-ray gray images with defects and no defects;

3. The method for detecting defects with high precision by fusion of positive and negative samples according to claim 2, wherein the step of screening out a plurality of problem areas of the X-ray gray scale image to be detected according to the first detection result comprises the steps of:

4. The high-precision defect detection method for positive and negative sample fusion is characterized by comprising the following steps of:

acquiring a plurality of X-ray gray images with defects and no defects;

5. The method for detecting a defect with high precision by fusion of positive and negative samples according to claim 4, wherein the step of determining whether the X-ray gray scale image to be detected has a problem area according to the first detection result comprises:

6. A high-precision defect detection system for positive and negative sample fusion is characterized by comprising:

a processor;

a memory storing a computer executable program which, when executed by the processor, causes the processor to perform a high precision defect detection method of positive and negative sample fusion as claimed in any one of claims 1 to 5.

7. A computer readable storage medium having stored thereon a computer program, wherein the program when executed by a processor implements a high precision defect detection method of positive and negative sample fusion according to any of claims 1-5.