WO2022236876A1

WO2022236876A1 - Cellophane defect recognition method, system and apparatus, and storage medium

Info

Publication number: WO2022236876A1
Application number: PCT/CN2021/095962
Authority: WO
Inventors: 刘宇迅; 田丰; 罗立浩; 陈小旋; 黄建
Original assignee: 广州广电运通金融电子股份有限公司
Priority date: 2021-05-14
Filing date: 2021-05-26
Publication date: 2022-11-17
Also published as: CN113239930B; CN113239930A

Abstract

Disclosed in the present invention are a cellophane defect recognition method, system and apparatus, and a storage medium. The cellophane defect recognition method specifically comprises: step S1, collecting cellophane surface images to establish test set data; step S2, importing the test set data into a semantic segmentation network model based on an optimized UNET network for semantic segmentation; and step S3: acquiring an output signal of the semantic segmentation network model, and post-processing the output signal to obtain a defect recognition result. According to the present invention, a conventional image processing method and a semantic segmentation network are combined as the basis for automatic classification, thereby effectively improving the robustness of cellophane defect recognition compared with conventional image processing and analysis; moreover, the semantic segmentation model uses the optimized UNET network, such that a cellophane defect can be quickly detected, the time required for actual detection and analysis is reduced, the efficiency and accuracy of cellophane defect detection are improved, and detection costs are reduced.

Description

A cellophane defect identification method, system, device and storage medium

technical field

The invention relates to the technical field of cellophane detection, in particular to a cellophane defect identification method, system, device and storage medium.

Background technique

As an important part of daily life and industrial production, cellophane is an important part of promoting economic development. In the production process of cellophane, there will be different forms of defects on the surface of cellophane. The main method for detecting defects of cellophane is manual visual inspection. It is easy to miss subtle defects in the manual inspection process, and the production speed of cellophane is relatively fast. Using naked eyes The speed of inspection can never keep up with the production speed of cellophane, which makes the detection efficiency very slow and seriously delays the production efficiency of cellophane.

Therefore, in order to improve the detection efficiency of cellophane, some manufacturers will think of using machine vision detection instead of manpower to complete the detection steps in production; however, because cellophane has certain light transmission, and has the characteristics of different patterns and different colors As a result, the traditional image defect recognition algorithm cannot effectively detect defects for different color paper rolls, making the accuracy of the detection results using the traditional image defect recognition method poor, and the detection speed cannot be improved.

Contents of the invention

In order to overcome the deficiencies of the prior art, one of the objects of the present invention is to provide a cellophane defect identification method, which improves the efficiency and accuracy of cellophane defect detection.

The second object of the present invention is to provide a cellophane defect identification system that implements the above method.

The third object of the present invention is to provide a cellophane defect identification device that implements the above method.

The fourth object of the present invention is to provide a storage medium for performing the above method.

One of purpose of the present invention adopts following technical scheme to realize:

A cellophane defect identification method, comprising:

Step S1: collect cellophane surface images, and establish test set data;

Step S2: Import the test set data into the semantic segmentation network model based on the optimized UNET network for semantic segmentation;

Step S3: Obtain the output signal of the semantic segmentation network model, and perform post-processing on the output signal to obtain a defect recognition result.

Further, the construction method of the semantic segmentation network includes:

Establish training set data according to the collected cellophane surface images;

Import the training set data into the optimized UNET network;

Calculate the loss value by using the output data and label data of the UNET network through a loss function, wherein the label data is obtained by segmenting and labeling defect positions in the training set data;

Update the network parameters through backpropagation until the model converges, then save the network parameters of the model to obtain a trained semantic segmentation network model.

Further, the acquisition method of the tag data is:

Import the surface image of the cellophane, select the corresponding defect area according to the requirement, and assign a label to the selected area to obtain the label data.

Further, the UNET network uses the group convolution module to group the input feature map, and then uses the convolution module to convolve each group respectively; one of the convolution modules uses 1*1, 1*3 and 3*1 combined convolution.

Further, the loss function is composed of a cross-entropy loss function and a dice loss measurement function.

Further, the post-processing method is:

Binarize the output signal of the semantic segmentation network model according to a preset fixed threshold;

Then search the image contour and judge whether there is a defect in the image combined with the contour size feature.

Further, in the step S1, an industrial camera is used to photograph the surface of the cellophane to obtain an image of the cellophane surface.

Two of the purpose of the present invention adopts following technical scheme to realize:

A cellophane defect identification system based on the UNET network, including:

The collection module is used to collect cellophane surface images and establish test set data;

The model analysis module is used to import the test set data into the constructed semantic segmentation network model for semantic segmentation;

The post-processing module is used to obtain the output signal of the model analysis module, and perform post-processing on the output signal to obtain the defect identification result.

Three of the purpose of the present invention adopts following technical scheme to realize:

A cellophane defect identification device, characterized in that it comprises:

program;

a memory for storing the program;

The processor is configured to load the program to execute the method for identifying cellophane defects as described above.

Four of the purpose of the present invention adopts following technical scheme to realize:

A storage medium stores a program, and when the program is executed by a processor, the above cellophane defect identification method is implemented.

Compared with the prior art, the beneficial effects of the present invention are:

The present invention uses the traditional image processing method combined with the semantic segmentation network as the basis for automatic classification, which effectively improves the robustness of cellophane defect recognition compared to relying solely on traditional image processing and analysis; at the same time, the semantic segmentation model uses the existing The optimized UNET network can quickly detect cellophane defects, reduce the time required for actual detection and analysis, improve the efficiency and accuracy of cellophane defect detection, and reduce detection costs.

Description of drawings

Fig. 1 is the schematic flow chart of cellophane defect identification method of the present invention;

Fig. 2 is the overall flowchart of the testing and training steps of cellophane defect recognition method of the present invention;

Fig. 3 is the structural diagram of the UNET model of the present invention;

Fig. 4 is a schematic diagram of group convolution in the present invention;

Fig. 5 is a schematic diagram of 1x1, 3x1, 1x3 combined convolution of the present invention;

Fig. 6 is a diagram of the defect prediction result of a solid-color paper roll according to the present invention;

Fig. 7 is a diagram of the defect prediction results of colored paper rolls of the present invention;

Fig. 8 is a schematic block diagram of modules of the cellophane defect identification system of the present invention.

Detailed ways

Below, the present invention will be further described in conjunction with the accompanying drawings and specific implementation methods. It should be noted that, under the premise of not conflicting, the various embodiments described below or the technical features can be combined arbitrarily to form new embodiments. .

Embodiment one

This embodiment provides a cellophane defect identification method, which can replace the existing manual detection and improve the efficiency and accuracy of cellophane defect detection.

As shown in Figure 1 and Figure 2, the cellophane defect identification method of this embodiment specifically includes the following steps:

Step S1: collect cellophane surface images, and establish test set data;

In this embodiment, an industrial camera is used to photograph the surface of cellophane after production and manufacture, thereby obtaining an image of the surface of cellophane, and establishing a data set of related original images, wherein the data set is divided into training set data and test set data, and the training set is photographed Several photographed images obtained on the surface of cellophane, use the training set data as the training basis of the semantic segmentation network model, thereby constructing a complete semantic segmentation network model; The goal is to import the test set into the established semantic segmentation network model, detect whether there are defects in the pictures in the test set, and obtain the cellophane defect recognition result.

In this embodiment, after the training set data is obtained, it is input into the lightweight UNET network for training, and finally a trained semantic segmentation network model is obtained. In this embodiment, the lightweight UNET network is optimized on the basis of the existing UNET network; the network structure of the existing UNET network includes two parts, an encoder and a decoder, wherein the encoder performs a down-sampling process, and The decoder performs the upsampling process; the downsampling part consists of multiple Down Blocks, which function as feature extractors. Each Down Block contains two convolutional and pooling layers, and the Max Pooling layer used by the pooling layer The pooling operation, every time a Down Block structure is passed, the spatial size of the feature map will be halved. The upsampling part of the existing UNET network is composed of Up Block. In the Up Block structure, it contains two convolutions and one upsampling layer, which is used to restore the size of the feature map layer by layer.

What Fig. 3 shows is the structural representation of the lightweight UNET network of the present embodiment, as shown in Fig. 3, the lightweight UNET network in the present embodiment is to improve on the structure of the existing UNET network, the present embodiment To optimize the original traditional convolution structure, each Down Block structure in the existing UNET network uses two 3*3 convolutions to complete the down-sampling process, and this embodiment improves the Down Block structure to A 3*3 convolution and a combination convolution, and adopts the group convolution method, and its combination convolution is composed of 1*1, 1*3 and 3*1, and this embodiment compresses the times of downsampling and upsampling, and then Reduce the amount of network parameters.

Specifically, as shown in Figure 4, in the combined convolution module of the lightweight UNET network of this embodiment, the cellophane image can be processed by the convolution kernel to obtain the feature map of the image, and the input feature map is grouped , and then each group is convolved separately. Assume that the size of the input feature map is still C*H*W, and the number of output feature maps is N. If it is set to be divided into G groups, the number of input feature maps for each group is C/G, and the output feature map for each group is The number of maps is N/G, the size of each convolution kernel is C/G*K*K, the total number of convolution kernels is still N, and the number of convolution kernels in each group is N/G, and the convolution kernel is only related to The input map of the same group is convolved, and the total parameter quantity of the convolution kernel is N*C/G*K*K, so it can be seen that the total parameter quantity is reduced to the original 1/G.

Specifically, as shown in Figure 5, in this embodiment, the existing 3*3 convolution is decomposed into a combined convolution of 1*1, 1*3 and 3*1, and the convolution parameters after decomposition are undecomposed The previous 45% can greatly reduce the amount of network parameters, and carry out lightweight processing on the existing UNET network, thereby improving the efficiency of cellophane defect identification. In addition, in this embodiment, three more activation functions are used after decomposition, which can increase the linearization capability.

In this embodiment, the training set data is imported into the optimized lightweight UNET network of the above structure, and the semantic segmentation network model can be constructed using the training set data. The construction method of the semantic segmentation network of the present embodiment includes: importing the training set data into the optimized UNET network; after the images in the training set are processed by the lightweight UNET network, the semantically segmented image can be obtained, from the semantic All the features of the cellophane surface can be seen in the image obtained after segmentation, including the pattern on the cellophane and the defects on the cellophane surface; then, the output data and label data of the UNET network are calculated through the loss function to calculate the loss value, update the network parameters through backpropagation, if the model converges, that is, the loss value stabilizes to the lowest value, then test the model and save the latest network parameters for subsequent testing, otherwise, continue to input the training set into the network, The network parameters are not saved until the network converges, and finally a trained semantic segmentation network model is obtained.

Wherein the label data is obtained by segmenting and labeling the defect positions in the training set data, that is, when the training set data is obtained, the pre-imported cellophane defect image is calibrated for the defect area, and the selected area is assigned a label , repeat the above process until the labeling is completed, and save the corresponding label to obtain the label data. In this embodiment, the labelme tool can be used to perform image labeling, that is, import the cellophane image into the labelme tool, select a defective region in the cellophane image, and assign values to image pixels in the defective region to obtain label data.

In this embodiment, the label data and the signal output by the lightweight UNET network are used to calculate the loss value through the loss function. In order to better evaluate the matching degree between the predicted value of the model and the real label, the loss function of this embodiment is composed of the cross entropy loss function It is combined with the dice loss metric function.

Among them, the cross-entropy is derived from the Kullback-Leibler (KL) divergence, which is a measure of the difference between two distributions. For general machine learning tasks, the distribution of data is given by the training set. Therefore minimizing the KL divergence is equivalent to minimizing the cross-entropy. Cross entropy is defined as:

where N is the number of samples, if label c is the correct classification for pixel i, then

binary index;

is the corresponding predicted probability.

The Dice loss metric function aims to minimize the area where the ground truth G and the predicted segmentation area S do not match, or to maximize the area where G and S overlap.

where, if label c is the correct classification for pixel i, then

binary index;

is the corresponding predicted probability.

After obtaining the trained semantic segmentation network model through the above method, use the trained model to perform semantic segmentation on the cellophane surface image in the test set, and then use traditional methods to post-process the network output signal to determine whether the signal has defects. To further explain, the method of post-processing the network output signal is: set a fixed threshold to binarize the output signal, perform contour search and combine the contour size features to identify the defect position on the cellophane surface, thereby judging the cellophane Whether the surface image is flawed.

As shown in Figures 6 and 7, after the method of this embodiment is used to identify defects on the surface of cellophane in different colors, no matter whether it is a pure-color paper roll or a colored paper roll, the position of the defect on the surface of the cellophane can be accurately identified. detection, the method improves the efficiency and accuracy of cellophane defect detection, and reduces the detection cost.

Embodiment two

This embodiment provides a cellophane defect recognition system based on the UNET network, which implements the cellophane defect recognition method described in Embodiment 1; as shown in Figure 8, the recognition system of this embodiment specifically includes the following modules:

This embodiment optimizes the existing UNET network, which can reduce the amount of network parameters, thereby achieving the effect of a lightweight UNET network, thereby increasing the image processing speed, quickly detecting cellophane defects, and significantly reducing the time required for actual detection and analysis; Secondly, the semantic segmentation network is used to replace the original manual detection method, which eliminates the subjectivity of manual detection and manual analysis, improves the efficiency and accuracy of cellophane defect detection, and reduces the detection cost; in addition, this embodiment uses traditional image processing The combination of the method and the semantic segmentation network is used as the basis for automatic classification. Compared with relying solely on traditional image processing and analysis, the accuracy of the detection method in this embodiment can be increased to 98%. Stickiness.

Embodiment Three

This embodiment discloses a cellophane defect identification device, including:

program;

a memory for storing the program;

The processor is configured to load the program to execute the cellophane defect identification method described in Embodiment 1.

This embodiment discloses a storage medium, which stores a program, and when the program is executed by a processor, the method for identifying cellophane defects is realized.

The device and storage medium in this embodiment are based on the two aspects of the same inventive concept as the method in the previous embodiment. The implementation process of the method has been described in detail above, so those skilled in the art can understand clearly from the foregoing description To understand the structure and implementation process of the device in this implementation, for the sake of brevity of the description, details will not be repeated here. The above-mentioned embodiment is only a preferred embodiment of the present invention, and cannot be used to limit the protection scope of the present invention. Any insubstantial changes and substitutions made by those skilled in the art on the basis of the present invention belong to the scope of the present invention. Scope of protection claimed.

The above-mentioned embodiment is only a preferred embodiment of the present invention, and cannot be used to limit the protection scope of the present invention. Any insubstantial changes and substitutions made by those skilled in the art on the basis of the present invention belong to the scope of the present invention. Scope of protection claimed.

Claims

A cellophane defect identification method, characterized in that, comprising:

Step S1: collect cellophane surface images, and establish test set data;

Step S2: Import the test set data into the semantic segmentation network model based on the optimized UNET network for semantic segmentation;

Step S3: Obtain the output signal of the semantic segmentation network model, and perform post-processing on the output signal to obtain a defect recognition result.
The cellophane defect recognition method according to claim 1, wherein the construction method of the semantic segmentation network comprises:

Establish training set data according to the collected cellophane surface images;

Import the training set data into the optimized UNET network;

Calculate the loss value by using the output data and label data of the UNET network through a loss function, wherein the label data is obtained by segmenting and labeling defect positions in the training set data;

Update the network parameters through backpropagation until the model converges, then save the network parameters of the model to obtain a trained semantic segmentation network model.
The cellophane defect identification method according to claim 2, wherein the method for obtaining the label data is:

Import the surface image of the cellophane, select the corresponding defect area according to the requirement, and assign a label to the selected area to obtain the label data.
The cellophane defect identification method according to claim 2, wherein the UNET network uses a group convolution module to group the input feature map, and then uses the convolution module to convolve each group respectively; one of the convolution modules The product module uses a combined convolution of 1*1, 1*3 and 3*1.
The cellophane defect recognition method according to claim 2, wherein the loss function is formed by combining a cross-entropy loss function and a dice loss measurement function.
Cellophane defect identification method according to claim 1, is characterized in that, the method for described aftertreatment is:

Binarize the output signal of the semantic segmentation network model according to a preset fixed threshold;

Then search the image contour and judge whether there is a defect in the image combined with the contour size feature.
The cellophane defect identification method according to claim 1, characterized in that, in the step S1, an industrial camera is used to photograph the surface of the cellophane to obtain an image of the cellophane surface.
A cellophane defect identification system based on the UNET network, characterized in that it comprises:

The collection module is used to collect cellophane surface images and establish test set data;

The model analysis module is used to import the test set data into the constructed semantic segmentation network model for semantic segmentation;

The post-processing module is used to obtain the output signal of the model analysis module, and perform post-processing on the output signal to obtain the defect identification result.
A cellophane defect identification device, characterized in that it comprises:

program;

a memory for storing the program;

A processor is used to load the program to execute the cellophane defect identification method according to any one of claims 1-7.
A storage medium, which stores a program, wherein when the program is executed by a processor, the cellophane defect identification method according to any one of claims 1-7 is implemented.