Chlor-alkali plant crystallization detection method based on image refinement segmentation

Abstract

The invention relates to a chlor-alkali plant crystal detection method based on image refinement segmentation. The method comprises the following steps: S1, carrying out regional segmentation processing with a buffer area on an acquired original crystal detection image; s2, inputting a detection area image obtained after segmentation into the trained YOLO neural network model for detection, and obtaining an output result of a detection area image target; s3, mapping the output result of the detection area image target to the original detection image to obtain the output result of the original detection image target; and S4, performing duplicate removal processing on an output result of the original detection image target to obtain a final crystallization detection result. Regional segmentation processing with a buffer area is carried out on an acquired crystal original image, so that a network can learn wider context information more quickly, the relative receptive field of a target in the image is indirectly increased, an identification result is mapped into the original image, and duplicate removal processing is carried out, so that an accurate identification result can be obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of visual recognition, and in particular to a method for crystallization detection in a chlor-alkali plant based on image refinement and segmentation. Background Art

[0002] As a computer vision technology, target detection can automatically identify the location of target objects of a specific category through processes such as candidate region generation, feature extraction, target classification, and bounding box regression. With the continuous development of deep learning technology, target detection technology has been widely used in fields including intelligent security, autonomous driving, and intelligent manufacturing. In addition, the accuracy and efficiency of this technology have gradually matured with the development of technologies such as image processing and machine learning.

[0003] Target detection technology is an image-based recognition technology. The quality of the image, the size of the target and the visibility of the target features directly affect the effect of target detection.

[0004] The chlor-alkali industry is an important chemical process that is mainly used to produce chlorine, hydrogen and caustic soda (sodium hydroxide), which are widely used in many industrial fields. In the production process of chlor-alkali plants, crystallization on equipment is an issue that requires special attention because it is directly related to production efficiency and product quality. For example, crystallized substances may cause corrosion to equipment, affecting the service life and safety of the equipment.

[0005] However, the size of chemical crystals in chlor-alkali plants is very small and their features are not obvious, especially their color information is not easy to distinguish from background information, which makes it difficult for neural networks to extract the features of crystal targets in images with huge amounts of information, thus increasing the challenge of target detection. As a result, there is currently no method that can solve the problem of chemical crystal detection in chlor-alkali plants. Summary of the invention

[0006] In order to solve the problems existing in the above-mentioned technology, the present invention provides a method for crystallization detection in a chlor-alkali plant based on image refinement and segmentation, comprising the following steps:

[0007] S1 performs regional segmentation processing with a buffer zone on the collected original detection image of the crystal;

[0008] S2 inputs the detection area image obtained after segmentation into the trained YOLO neural network model for detection, and obtains the output result of the detection area image target;

[0009] S3 maps the output result of the detection area image target to the original detection image to obtain the output result of the original detection image target;

[0010] S4 performs deduplication processing on the output result of the original detection image target to obtain the final crystallization detection result.

[0011] Furthermore, the training method of the YOLO neural network model in step S2 is:

[0012] Collecting the original training image of the crystallization, performing regional segmentation processing with a buffer zone on the original training image, annotating the segmented training area image, and forming a sample file together with the corresponding label file;

[0013] Divide the sample files into training set, validation set and test set;

[0014] The training set is input into the YOLO neural network model for training, and the model effect is tested through the validation set, and finally the model is tested through the test set.

[0015] Furthermore, the regional segmentation process with a buffer is to divide the original image into n x *n y The image processing is performed to obtain a plurality of area images of the same size, wherein the buffer zone is an overlapping portion of adjacent area images;

[0016] The length and width of the region image are calculated according to the following formula:

[0017]

[0018] Among them, l seg and w seg Represent the length and width of the region image respectively, Overlapping is the overlapping length or width of two adjacent region images, x and y represent the length and width of the original image respectively, and nx and ny represent the number of divisions of the length and width of the original image respectively;

[0019] Wherein, when the original image is an original detection image, the region image is a detection region image;

[0020] When the original image is an original training image, the region image is a training region image.

[0021] Furthermore, the output result of the detection area image target is expressed as:

[0022]

[0023] Among them, x i ,y i , w i ,h i Respectively represent the horizontal coordinate of the upper left corner point of the i-th target box, the vertical coordinate of the upper left corner point, the width of the target box, and the height of the target box;

[0024] id is the type of the i-th target;

[0025] C i is the confidence of the i-th target.

[0026] Furthermore, in step S3, the method of mapping the target frame of the detection area image to the original detection image is expressed as:

[0027]

[0028] Among them, X i , Y i , W i , H i They represent the horizontal coordinate of the upper left corner of the target box on the original detection image, the vertical coordinate of the upper left corner, the width of the target box, and the height of the target box respectively; jx and jy represent the order of the detection area image from left to right and from top to bottom respectively.

[0029] Furthermore, the deduplication process in step S4 is as follows: the IOU value is calculated for the target boxes with overlapping areas, and two targets with an IOU value greater than 0.8 are regarded as the same target, and the target with the higher ID is retained.

[0030] Furthermore, the YOLO neural network model is a YOLOv3 neural network model;

[0031] Multi-head attention modules are added to the Backbone and Head parts of the YOLOv3 neural network model respectively.

[0032] The beneficial effects of the present invention are as follows: the present invention proposes a method for crystallization detection in a chlor-alkali plant based on image refinement and segmentation, using YOLOv3 with a multi-head attention module as a detection framework, and performing nx*ny regional segmentation processing with a buffer on the collected original crystallization image, so that the network can learn a wider range of contextual information more quickly, indirectly increasing the relative receptive field of the target in the image, and sending the processed image to the YOLOv3 detection framework to identify the target. After that, the recognition result is mapped to the original image and deduplication processing is performed, so that accurate recognition results can be obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0034] Figure 1 It is a schematic diagram of an image segmentation method with a buffer zone of the present invention;

[0035] Figure 2 It is a schematic diagram of a method for deduplication processing of the output result of the original detected image target of the present invention. DETAILED DESCRIPTION

[0036] The technical scheme in the embodiment of the present invention is described clearly and completely below in conjunction with the accompanying drawings in the embodiment of the present invention. In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present invention, so the present invention is not limited by the specific embodiments disclosed below.

[0037] A method for crystallization detection in a chlor-alkali plant based on image refinement and segmentation comprises the following steps:

[0038] 1. Train the neural network model YOLOv3, and add multi-head attention modules (Coordinate attention) to the Backbone and Head parts of the YOLOv3 neural network model.

[0039] Collecting the original training image of the crystallization, performing regional segmentation processing with a buffer zone on the original training image, annotating the segmented training area image, and forming a sample file together with the corresponding label file;

[0040] The sample files are divided into training set, validation set and test set in the ratio of 80%, 10% and 10%;

[0041] The training set is input into the YOLO neural network model for training, and the model effect is tested through the validation set. Finally, the model is tested through the test set to obtain the trained neural network model.

[0042] Among them, Figure 1 As shown, the regional segmentation process with a buffer zone is to segment the original image into nx*ny parts to obtain multiple regional images of the same size, and the buffer zone is the overlapping part of adjacent regional images;

[0043] The length and width of the region image are calculated according to the following formula:

[0044]

[0045] Among them, l seg and w seg Respectively represent the length and width of the regional image, Overlapping is the overlapping length or width of two adjacent regional images, x and y represent the length and width of the original image, nx and ny represent the number of divisions of the length and width of the original image, respectively. In this embodiment, nx and ny are both 2;

[0046] Among them, the original image is the original training image, and the region image is the training region image.

[0047] 2. Perform regional segmentation processing with a buffer zone on the collected original detection image of the crystal according to the method in the previous step. During the segmentation processing, the original image is the original detection image, and the regional image is the detection regional image;

[0048] 3. Input the detection area image obtained after segmentation into the trained YOLO neural network model for detection, and obtain the output result of the detection area image target;

[0049] The output result of the detection area image target is expressed as:

[0050]

[0051] Among them, x i ,y i , w i ,h i Respectively represent the horizontal coordinate of the upper left corner point of the i-th target box, the vertical coordinate of the upper left corner point, the width of the target box, and the height of the target box;

[0052] id is the type of the i-th target, and all types are 0;

[0053] C i is the confidence of the i-th target, which refers to the probability that the target box is a class 0 target, and its value is in the range of [0, 1].

[0054] 4. Map the output result of the detection area image target to the original detection image to obtain the output result of the original detection image target;

[0055] The method of mapping the target box of the detection area image to the original detection image is expressed as:

[0056]

[0057] Among them, X i , Y i , W i , H i They represent the horizontal coordinate of the upper left corner of the target box on the original detection image, the vertical coordinate of the upper left corner, the width of the target box, and the height of the target box respectively; jx and jy represent the order of the detection area image from left to right and from top to bottom respectively.

[0058] 5. If Figure 2 As shown, the output result of the original detection image target is deduplicated to obtain the final crystallization detection result.

[0059] The deduplication process is as follows: the IOU value is calculated for the target boxes with overlapping areas. Two targets with an IOU value greater than 0.8 are considered to be the same target, and the target with the earlier ID is retained.

[0060] The calculation method of IOU value is as follows:

[0061]

[0062] Among them, A and B are two overlapping detection frames.

[0063] 6. The retained target frame is used as the final crystal target detection result.

[0064] Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

Claims

1. A method for crystallization detection in a chlor-alkali plant based on image refinement and segmentation, characterized in that: The steps include: S1 performs regional segmentation processing with a buffer zone on the collected original detection image of the crystal; S2 inputs the detection area image obtained after segmentation into the trained YOLO neural network model for detection, and obtains the output result of the detection area image target; S3 maps the output result of the detection area image target to the original detection image to obtain the output result of the original detection image target; S4 performs deduplication processing on the output result of the original detection image target to obtain the final crystallization detection result.

2. The method for crystallization detection in a chlor-alkali plant based on image refinement and segmentation according to claim 1, characterized in that: The training method of the YOLO neural network model in step S2 is: Collecting the original training image of the crystallization, performing regional segmentation processing with a buffer zone on the original training image, annotating the segmented training area image, and forming a sample file together with the corresponding label file; Divide the sample files into training set, validation set and test set; The training set is input into the YOLO neural network model for training, and the model effect is tested through the validation set, and finally the model is tested through the test set.

3. The method for crystallization detection in a chlor-alkali plant based on image refinement and segmentation according to claim 2, characterized in that: The regional segmentation process with a buffer is to divide the original image into n x *n y The image processing is performed to obtain a plurality of area images of the same size, wherein the buffer zone is an overlapping portion of adjacent area images; The length and width of the region image are calculated according to the following formula: Among them, l seg and w seg Represent the length and width of the region image respectively, Overlapping is the overlapping length or width of two adjacent region images, x and y represent the length and width of the original image respectively, and nx and ny represent the number of divisions of the length and width of the original image respectively; Wherein, when the original image is an original detection image, the region image is a detection region image; When the original image is an original training image, the region image is a training region image.

4. The method for crystallization detection in a chlor-alkali plant based on image refinement and segmentation according to claim 3 is characterized in that: The output result of the detection area image target is expressed as: Among them, x i ,y i , w i ,h i Respectively represent the horizontal coordinate of the upper left corner point of the i-th target box, the vertical coordinate of the upper left corner point, the width of the target box, and the height of the target box; id is the type of the i-th target; C i is the confidence of the i-th target.

5. The method for crystallization detection in a chlor-alkali plant based on image refinement and segmentation according to claim 4, characterized in that: In step S3, the method of mapping the target frame of the detection area image to the original detection image is expressed as: Among them, X i , Y i , W i , H i They represent the horizontal coordinate of the upper left corner of the target box on the original detection image, the vertical coordinate of the upper left corner, the width of the target box, and the height of the target box respectively; jx and jy represent the order of the detection area image from left to right and from top to bottom respectively.

6. The method for crystallization detection in a chlor-alkali plant based on image refinement and segmentation according to claim 5, characterized in that: The deduplication process in step S4 is as follows: the IOU value is calculated for the target boxes with overlapping areas, and two targets with an IOU value greater than 0.8 are regarded as the same target, and the target with the higher ID is retained.

7. A method for crystallization detection in a chlor-alkali plant based on image refinement and segmentation according to any one of claims 1 to 6, characterized in that: The YOLO neural network model is a YOLOv3 neural network model; The Backbone and Head parts in the YOLOv3 neural network model are respectively added with multi-head attention modules.

Images