CN115601357A - Stamping part surface defect detection method based on small sample - Google Patents

Abstract

The invention relates to the technical field of surface defect detection of stamping parts, and solves the problems that in the prior art, the missing detection rate of image fine defect detection is high and fine defects cannot be accurately detected, in particular to a stamping part surface defect detection method based on a small sample, which comprises the following steps: s1, acquiring image data I to be detected by using a CCD camera; s2, preprocessing the image data I to be detected by adopting a contrast enhancement network based on image inverse mapping to obtain an enhanced full-resolution image P; and S3, inputting the full-resolution image P into the improved SSD-Net detection network, and training the improved SSD-Net detection network by combining a loss function L. The method can accurately detect the tiny defects on the surface of the stamping part, further reduces the missing rate of defect detection, reduces the dependence of the traditional defect detection on a sample, and improves the precision and the efficiency of the tiny defect detection.

Description

Stamping part surface defect detection method based on small sample

Technical Field

The invention relates to the technical field of stamping part surface defect detection, in particular to a stamping part surface defect detection method based on a small sample.

Background

With the rapid development of science and technology and industry in the current society, the quality of life of people is greatly improved, and the requirements of the nation on various industrial products are more and more strict. In the traditional industrial production and manufacturing, because of the limitation of research technology, a defect detection method mainly adopting manual detection is still adopted to detect the defects on the surface of the product, and the method has the defects of low speed, low efficiency and high labor cost because of the manual limitation and the backward technology, and in the detection process, because of the randomness of the defects, the types and the sizes of the defects are various, fine defects on a plurality of workpieces cannot be manually detected, and the conditions of defect omission and error detection are easy to occur.

In the current society, with the development of scientific technology, the appearance and development of scientific technology such as artificial intelligence and the like and the deep research, an image surface defect detection technology based on machine vision appears in the field of image defect detection, in a highly automated production scene, the yield of products is particularly high, it is very time-consuming to collect defect samples, most of the current deep learning methods for defect detection are based on a large number of defect samples to establish a model, and the model is difficult to be on-line due to the lack of the defect samples.

The defect detection method based on the small sample greatly improves the efficiency of detection work, reduces the dependence of an algorithm on the sample, is widely applied to the fields of road tunnel engineering detection, workpiece surface quality detection and aerospace, and avoids the inaccuracy of subjective judgment existing in manual detection. However, the method has a poor effect in the aspect of image fine defect detection, and fine defects of products cannot be accurately detected.

Disclosure of Invention

Aiming at the defects of the prior art, the invention provides a stamping part surface defect detection method based on a small sample, which solves the problems that in the prior art, the missing detection rate of image fine defect detection is high and fine flaws cannot be accurately detected.

In order to solve the technical problems, the invention provides the following technical scheme: a stamping part surface defect detection method based on a small sample comprises the following steps:

s1, acquiring image data I to be detected by using a CCD camera;

s2, preprocessing the image data I to be detected by adopting a contrast enhancement network based on image inverse mapping to obtain an enhanced full-resolution image P;

s3, inputting the full-resolution image P into the improved SSD-Net detection network, and training the improved SSD-Net detection network by combining a loss function L;

and S4, carrying out defect detection on the surface image of the stamping part by using the trained improved SSD-Net detection network.

Further, in step S2, a contrast enhancement network based on image inverse mapping is used to pre-process the image data to be detected to obtain an enhanced full-resolution image P, and the specific process includes the following steps:

s21, sampling and coding image data I to be detected into a feature map, extracting local and global features, and connecting the local and global features in series to obtain a predicted low-resolution inverse mapping map group;

s22, upsampling the predicted low-resolution inverse mapping map set by using the bilateral grid to generate a full-resolution mapping map set S, and multiplying the full-resolution mapping map set S with each pixel point of the input image data I to be detected to obtain an enhanced full-resolution image P.

Furthermore, the improved SSD-Net detection network comprises a backbone network and a GGA module, wherein the backbone network comprises five coding blocks, an attention mechanism is embedded in the first two coding blocks, and each coding block comprises an attention unit and two convolution layers which are connected in series and are connected in parallel.

Furthermore, in the last three coding blocks of the backbone network, each coding block is formed by connecting four convolution layers with convolution kernel size of 3 × 3 in series.

Furthermore, the GGA module consists of an anchor point prediction area and a characteristic adaptation area, wherein the anchor point prediction area consists of two branches, namely a position prediction branch and a shape prediction branch.

Further, in step S3, the full-resolution image P is input into the improved SSD-Net detection network, and the improved SSD-Net detection network is trained in combination with the loss function L, and the specific process includes the following steps:

s31, inputting the enhanced full-resolution image P into a first coding block 1 of an improved SSD-Net detection network, and solving a channel descriptor Xn of an input Un, wherein the Xn is subjected to two full connections to obtain a feature descriptor Xn';

s32, limiting the value of Xn' to the range of [0,1] by using a sigmoid function to obtain the weight S of each channel modulation weight set;

s33, convolving Un to obtain Un ', and applying the weight S to the characteristic mapping Un' to obtain an output characteristic U which is used as the input of the next coding block;

s34, inputting the output features U into the coding blocks 3, 4 and 5 to carry out convolution operation to obtain feature graphs with the same dimensionality;

s35, connecting the obtained feature graphs in series to obtain a feature graph group Fp;

s36, inputting the feature map group Fp into an anchor point prediction area of the GGA module to obtain the width and the height of a target detection frame;

s37, inputting the width and the height of a target detection frame into a feature adaptation area of the GGA module to obtain a new feature map Fp';

s38, determining a loss function L, wherein the loss function L comprises classification loss

Return loss of

Anchor point location loss

And anchor shape prediction loss

；

And S39, training the improved SSD-Net detection network according to the new feature diagram Fp' and the loss function L.

By means of the technical scheme, the invention provides a stamping part surface defect detection method based on a small sample, which at least has the following beneficial effects:

1. after the image of the target detection product is obtained, the invention can accurately detect the fine defects in the image, solve the problems of high missing detection rate of the fine defect detection of the image and incapability of accurately detecting the fine defects in the prior art, and further improve the accuracy of the defect detection.

2. The method can accurately detect the tiny defects on the surface of the stamping part, further reduces the omission factor of defect detection, solves the problems of high omission factor and low accuracy of manual tiny defect detection, reduces the dependence of the traditional defect detection on a sample, and improves the precision and the high efficiency of the tiny defect detection.

Drawings

The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiment(s) of the application and together with the description serve to explain the application and not to limit the application. In the drawings:

FIG. 1 is a flow chart of the method for detecting surface defects of a stamping part according to the present invention;

FIG. 2 is a network structure diagram of a contrast enhancement network based on image inverse mapping according to the present invention;

FIG. 3 is a network architecture diagram of the improved SSD-Net detection network of the present invention;

FIG. 4 is a schematic diagram of a CCD camera of the present invention acquiring surface image information of an object to be measured;

FIG. 5 is a schematic diagram showing the result of detecting the surface defect of the stamping part according to the method of the present invention in FIG. 4.

Detailed Description

In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below. Therefore, the realization process of how to apply technical means to solve the technical problems and achieve the technical effects can be fully understood and implemented.

It will be understood by those skilled in the art that all or part of the steps in the method for implementing the above embodiments may be implemented by relevant hardware instructed by a program, and therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

Referring to fig. 1 to fig. 5, a specific implementation manner of the present embodiment is shown, and after an image of a target inspection object is obtained, the present embodiment can accurately detect a fine defect in the image, so as to solve the problems in the prior art that the detection omission rate of the fine defect of the image is high and the fine defect cannot be accurately detected, and further improve the accuracy of defect detection.

Referring to fig. 1, the present embodiment provides a method for detecting surface defects of a stamping part based on a small sample, including the following steps:

s1, acquiring image data I to be detected by adopting a CCD camera;

s2, preprocessing the image data I to be detected by adopting a contrast enhancement network based on image inverse mapping to obtain an enhanced full-resolution image P, and improving the contrast of the image data to be detected;

referring to fig. 2, in step S2, the contrast enhancement network based on the image inverse mapping is used to pre-process the image data to be detected to obtain an enhanced full-resolution image P, and the specific process includes the following steps:

s21, sampling and coding image data I to be detected into a feature map, extracting local and global features, and connecting the local and global features in series to obtain a predicted low-resolution inverse mapping map group;

s22, upsampling the predicted low-resolution inverse mapping map set by using the bilateral grid to generate a full-resolution mapping map set S, and multiplying the full-resolution mapping map set S with each pixel point of the input image data I to be detected to obtain an enhanced full-resolution image P.

S3, inputting the full-resolution image P into an improved SSD-Net detection network, and training the improved SSD-Net detection network by combining a loss function L;

the improved SSD-Net detection network comprises a backbone network and a GGA module, wherein the backbone network comprises five coding blocks, an attention mechanism is embedded in the first two coding blocks, and each coding block comprises an attention unit and two serially connected convolution layers which are connected in parallel;

in the last three coding blocks, each coding block is formed by connecting four convolution layers with convolution kernel size of 3 multiplied by 3 in series;

the GGA module consists of two parts, namely an anchor point prediction area and a characteristic adaptation area, wherein the anchor point prediction area consists of two branches, namely a position prediction branch and a shape prediction branch;

referring to fig. 3, in step S3, the full-resolution image P is input into the improved SSD-Net detection network, and the improved SSD-Net detection network is trained in combination with the loss function L, which includes the following steps:

s31, inputting the enhanced full-resolution image P into a first coding block 1 of an improved SSD-Net detection network, and solving a channel descriptor Xn of an input Un, wherein the Xn is subjected to two full connections to obtain a feature descriptor Xn';

s32, limiting the value of Xn' to the range of [0,1] by using a sigmoid function to obtain the weight S of the modulation weight set of each channel;

s33, convolving Un to obtain Un ', applying the weight S to the feature mapping Un' to obtain an output feature U, and using the output feature U as the input of a next coding block to enhance important features and weaken unimportant features, so that the extracted feature directivity is stronger;

the calculation formula of the channel descriptor Xn is:

the formula for calculating the weight S is:

the calculation formula of the output characteristic U is as follows:

in the above formula, H and W are the spatial dimensions of Un,

is ReLU function, g is sigmoid function.

S34, inputting the output features U into the coding blocks 3, 4 and 5 to carry out convolution operation to obtain feature graphs with the same dimensionality;

and respectively performing up-sampling on feature maps obtained by the output features U through 3 convolution kernel groups of the coding blocks 3, 4 and 5 to obtain feature maps with the same dimension.

S35, connecting the obtained feature graphs in series to obtain a feature graph group Fp, wherein the operation can collect rich position information in a lower layer and collect strong semantic information in a higher layer;

s36, inputting the feature map group Fp into an anchor point prediction area of the GGA module to obtain the width and height of a target detection frame;

inputting the feature map group Fp into a GGA module, wherein the GGA module consists of an anchor point prediction area and a feature adaptation area. The position and shape of the detection box is characterized by a 4-tuple of the form (x, y, w, h), (x, y) being the spatial coordinates of the center, w being the width, h being the height.

The anchor point prediction area consists of two branches, the position prediction branch performs 1 multiplied by 1 convolution on the characteristic diagram Fp to obtain a target degree score diagram, and then the target degree score diagram is converted into a probability value through a sigmoid function of elements to output a probability prediction diagram with the same size as the input characteristic diagram Fp

Anchor points with a probability value above a predefined threshold L are considered foreground points, for

The anchor point position (i, j) corresponds to a position on the input image of

. The shape prediction branch performs 1 × 1 convolution on the feature map Fp to generate dw and dh, and then maps the dw and dh to (w, h), thereby obtaining the width w and the height h of the target detection box, and the mapping formula is as follows:

where k is the step size and where,

is an empirical scale factor

。

S37, inputting the width and the height of a target detection frame into a feature adaptation area of the GGA module to obtain a new feature map Fp';

specifically, the width and height of the wide and high input feature adaptive region obtained from the anchor point prediction region are represented by a 2-channel feature map, and a 3 × 3 variable convolution layer is applied to the 2-channel feature map to obtain the ith position-adaptive feature

Namely, the original feature map Fp is subjected to variable convolution operation to obtain a new feature map Fp'.

Location-adapted features of ith

Expressed as:

wherein Lt is a 3 × 3 deformable convolution, wherein

Is a characteristic of the ith position and,

and

is the corresponding anchor shape.

S38, determining a loss function L, wherein the loss function L comprises classification loss

Regression loss

Anchor point location loss

And anchor shape prediction loss

；

Specifically, the calculation formula of the loss function L is as follows:

wherein:

wherein:

represents aThe number of positive samples in each batch,

representing the number of negative samples in a batch.

The width and height of the bounding box predicted for the predefined width and height of the bounding box, w, h,

is the smooth L1 loss, and the smooth L1 loss,

，

the regression parameters of the ith bounding box are represented.

For regularizing the parameters, take

。

And S39, training the improved SSD-Net detection network according to the new feature diagram Fp' and the loss function L.

And S4, carrying out defect detection on the surface image of the stamping part by using the trained improved SSD-Net detection network.

Referring to fig. 4, a schematic diagram of a CCD camera acquiring information of a surface image of an object to be detected is shown, for example, the surface image of the object to be detected is detected by the method for detecting surface defects of a stamping part provided in this embodiment to obtain a detection result as shown in fig. 5, so that fine defects on the surface of the stamping part can be accurately detected, the missing rate of defect detection is further reduced, the problems of high missing rate and low accuracy of artificial fine defect detection are solved, the dependence of conventional defect detection on a sample is reduced, and the accuracy and the efficiency of fine defect detection are improved.

The present invention has been described in detail with reference to the foregoing embodiments, and the principles and embodiments of the present invention have been described herein with reference to specific examples, which are provided only to assist understanding of the methods and core concepts of the present invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed, and in summary, the content of the present specification should not be construed as a limitation to the present invention.

Claims (6)

1. A stamping part surface defect detection method based on a small sample is characterized by comprising the following steps:

s1, acquiring image data I to be detected by using a CCD camera;

s2, preprocessing the image data I to be detected by adopting a contrast enhancement network based on image inverse mapping to obtain an enhanced full-resolution image P;

s3, inputting the full-resolution image P into the improved SSD-Net detection network, and training the improved SSD-Net detection network by combining a loss function L;

and S4, carrying out defect detection on the surface image of the stamping part by using the trained improved SSD-Net detection network.

2. The method for detecting the surface defects of the stamping part according to claim 1, characterized in that: in step S2, a contrast enhancement network based on image inverse mapping is used to pre-process the image data to be detected to obtain an enhanced full-resolution image P, and the specific process includes the following steps:

s21, down-sampling and coding image data I to be detected into a feature map, extracting local and global features and connecting the local and global features in series to obtain a predicted low-resolution inverse mapping map group;

s22, upsampling the predicted low-resolution inverse mapping map set by using the bilateral grid to generate a full-resolution mapping map set S, and multiplying the full-resolution mapping map set S with each pixel point of the input image data I to be detected to obtain an enhanced full-resolution image P.

3. The method for detecting the surface defects of the stamping part according to claim 1, characterized in that: the improved SSD-Net detection network comprises a backbone network and a GGA module, wherein the backbone network comprises five coding blocks, an attention mechanism is embedded in the first two coding blocks, and each coding block comprises an attention unit and two convolution layers connected in series in parallel.

4. The method for detecting the surface defects of the stamping part according to claim 1, characterized in that: in the last three coding blocks of the backbone network, each coding block is formed by connecting four convolution layers with convolution kernel size of 3 multiplied by 3 in series.

5. A stamping part surface defect detection method according to claim 3, characterized in that: the GGA module consists of two parts, namely an anchor point prediction area and a characteristic adaptation area, wherein the anchor point prediction area consists of two branches, namely a position prediction branch and a shape prediction branch.

6. The method for detecting the surface defects of the stamping part according to claim 1, characterized in that: in step S3, the full-resolution image P is input into the improved SSD-Net detection network, and the improved SSD-Net detection network is trained in combination with the loss function L, which specifically includes the following steps:

s31, inputting the enhanced full-resolution image P into a first coding block 1 of an improved SSD-Net detection network, and solving a channel descriptor Xn of an input Un, wherein the Xn is subjected to two full connections to obtain a feature descriptor Xn';

s32, limiting the value of Xn' to the range of [0,1] by using a sigmoid function to obtain the weight S of each channel modulation weight set;

s33, convolving Un to obtain Un ', and applying the weight S to the feature mapping Un' to obtain an output feature U which is used as the input of the next coding block;

s34, inputting the output characteristic U into the coding blocks 3, 4 and 5 to carry out convolution operation to obtain a characteristic diagram with the same dimension;

s35, connecting the obtained feature graphs in series to obtain a feature graph group Fp;

s36, inputting the feature map group Fp into an anchor point prediction area of the GGA module to obtain the width and the height of a target detection frame;

s37, inputting the width and the height of the special target detection box into a feature adaptation area of the GGA module to obtain a new feature map Fp';

s38, determining a loss function L, wherein the loss function L comprises classification loss

Return loss of

Anchor point location loss

And anchor shape prediction penalties

；

And S39, training the improved SSD-Net detection network according to the new feature diagram Fp' and the loss function L.

Images