CN110110758A - A kind of surface roughness classification method based on convolutional neural networks - Google Patents

Abstract

The invention discloses a method for classifying surface roughness based on a convolutional neural network, which belongs to the field of quality inspection, and obtains light scattering distribution images of samples with different roughnesses; compresses the light scattering distribution images, and converts the compressed images into pixel matrices , construct a label matrix according to the level of the roughness sample block corresponding to the scattering image; normalize the pixel matrix, use the pixel matrix and label matrix as the input and output of the convolutional neural network, and set the basic parameters and training parameters; initialize The convolutional neural network is trained to obtain the optimal convolutional neural network classification model, and the classification accuracy of the model is analyzed with test samples; in actual measurement, the scattering image is obtained, and through corresponding processing, the optimal classification model is obtained. Measure the surface roughness level of the workpiece. The invention avoids the situation that the prediction error of the surface roughness is too large due to improper selection of characteristic parameters, and has high measurement accuracy and fast measurement speed.

Description

A Surface Roughness Classification Method Based on Convolutional Neural Network

technical field

The invention belongs to the technical field of quality inspection, in particular to a surface roughness classification method based on a convolutional neural network.

Background technique

Surface roughness is an important indicator for evaluating the surface quality of workpieces, and has an important impact on product quality, performance and life. At present, the methods for measuring the surface roughness of workpieces can be divided into two types: contact and non-contact. Contact measurement methods generally include touch method, comparison method, impression method and stylus method, etc. These methods have slow measurement speed, low measurement accuracy, and are easy to cause damage to the surface of the workpiece to be measured; Ultrasonic method, laser triangulation method, speckle method, scattering method and machine vision method, etc., these methods will not cause damage to the surface of the workpiece to be measured, and are currently the main measurement methods used. Among them, the measurement method based on the principle of light scattering has the characteristics of high sensitivity, simple structure, and low environmental requirements, and is suitable for on-line measurement in industrial environments. With the rapid development of machine vision technology, machine vision technology can be combined with scattering technology to measure the surface roughness of workpieces.

Min Li et al. used machine vision technology to obtain the surface image of the turning workpiece, and extracted 14 surface texture feature parameters based on the gray level co-occurrence matrix as the input of the BP neural network, and constructed a surface roughness prediction model. Wang Cuiting designed a DSP-based online surface roughness detection system, and extracted the characteristic parameters in the scattering image as the input of the support vector machine. The system can judge the surface roughness level of different workpieces. However, the above-mentioned surface roughness prediction models all need to extract characteristic parameters from images that can evaluate surface roughness, and the correspondence between characteristic parameters and surface roughness directly affects the prediction accuracy of the model. In other words, if the correspondence between the characteristic parameters and the surface roughness is not accurate enough, the accuracy of the prediction model trained with the characteristic parameters will not be too high.

The convolutional neural network is essentially a multi-layer perceptron. Unlike the traditional neural network, its neurons are connected to each other in a non-fully connected manner, and the connection weights between some neurons are shared. . Based on the above characteristics, convolutional neural networks are suitable for more complex application scenarios. At the same time, the close connection and spatial information between the layers in the convolutional neural network make it especially suitable for image processing, and can automatically extract color, texture and shape and other related characteristics from the image. The convolutional neural network can directly use the image as the input parameter of the network, thereby avoiding the process of feature extraction and data reconstruction in the traditional algorithm, and shortening the time. Currently, there are no reports on predicting surface roughness with convolutional neural networks.

Contents of the invention

Purpose of the invention: In view of the problems in the prior art, provide a surface roughness classification method based on convolutional neural network, which is suitable for various measurement occasions and measurement objects, and avoids surface roughness prediction caused by improper selection of characteristic parameters When the error is too large, the measurement accuracy is high and the measurement speed is fast.

Technical solution: In order to solve the above technical problems, the present invention provides a surface roughness classification method based on a convolutional neural network, comprising the following steps:

(1) sample acquisition;

(2) Compress the image and convert it into a pixel matrix to construct a label matrix;

(3) normalized pixel matrix;

(4) Set the basic parameters and training parameters of CNN;

(5) Initialize CNN and train;

(6) Test the convolutional neural network prediction model after training, that is, the CNN classification model, and analyze the accuracy of the model for surface roughness classification.

Further, the specific steps of sample acquisition in the step (1) are as follows:

(1.1) Obliquely shoot the collimated laser beam onto the surface of the workpiece, use the camera to capture the spatial distribution images of reflected light and scattered light carrying surface roughness information, and obtain the scattered images of standard samples with different roughness;

(1.2) The scattering image obtained from step (1.1) is used as a sample set.

Further, in the step (2), the image is compressed and converted into a pixel matrix, and the specific steps for constructing a label matrix are as follows:

Compress each scattering image in the sample set to 28*28, convert the compressed image into a 1*784 matrix, and form a pixel matrix according to the number of images in the sample set, and then according to the roughness sample corresponding to the scattering image in the sample set The level of the block constructs the label matrix, and they are used as the input and output of the convolutional neural network respectively.

Further, the basic parameters and training parameters of CNN are set in the step (4) and the specific steps are as follows:

(4.1) The basic parameters of the convolutional neural network are set; the basic parameters include the number of convolutional layers and downsampling layers, the size of the convolution kernel and the downsampling rate of decline;

(4.2) The training parameters of the convolutional neural network are set; the training parameters include the learning rate, the number of total samples for batch training and the number of iterations.

Further, the specific steps of initializing CNN and training in the step (5) are as follows:

(5.1) Initialize the convolutional neural network; the initialization content includes a convolution kernel and a bias;

(5.2) Training the convolutional neural network; the training process includes forward propagation, back propagation and weight modification of the network.

Further, in the step (1.1), the collimated laser beam is obliquely incident on the surface of the workpiece at the incident angle of the set angle, and a ground glass screen with a height consistent with the height of the laser is placed in the reflection direction of the corresponding angle, and a camera is used to shoot the ground glass screen. Scattered image.

Further, there are six levels of surface roughness classification in the step (6): 0.025 μm, 0.05 μm, 0.1 μm, 0.2 μm, 0.4 μm and 0.8 μm.

The invention uses Matlab software to realize the compression of scattering images, the normalization processing of sample data, the training of convolutional neural network model and the test of model classification effect, wherein the training and test of the model use the DeepLearnToolbox-master toolbox.

Compared with the prior art, the present invention has the advantages of:

(1) The present invention builds a relatively simple measurement system based on the principle of laser scattering, which is applicable to various measurement environments and measurement objects, and can realize online detection;

(2) The present invention uses a convolutional neural network as a classification model of surface roughness, and uses the scattering image collected by the measurement system as the input of the network, which avoids the phenomenon that the surface roughness prediction error is too large due to improper selection of characteristic parameters;

(3) The present invention uses a convolutional neural network as a classification model of surface roughness, which avoids processes such as extraction of characteristic parameters and data reconstruction of traditional algorithms, and shortens the prediction time.

Description of drawings

Fig. 1 is a flowchart of the present invention;

Fig. 2 is the light scattering distribution diagram corresponding to different roughnesses in the specific embodiment; Wherein 2 (a) is the light scattering distribution diagram corresponding to the surface roughness of 0.025 μm; 2 (b) is the light corresponding to the surface roughness of 0.05 μm Scattering distribution diagram; 2(c) is the light scattering distribution diagram corresponding to the surface roughness of 0.1 μm; 2(d) is the light scattering distribution diagram corresponding to the surface roughness of 0.2 μm; 2(e) is the surface roughness of 0.4 The light scattering distribution map corresponding to μm; 2(f) is the light scattering distribution map corresponding to the surface roughness of 0.8 μm.

Detailed ways

The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. The embodiments described in the present invention are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, other embodiments obtained by persons of ordinary skill in the art without making creative efforts all fall within the protection scope of the present invention.

As shown in Figure 1, a surface roughness classification method based on a convolutional neural network disclosed in an embodiment of the present invention mainly includes the following steps:

(1) The collimated laser beam is slanted onto the surface of the workpiece, and the camera is used to capture the spatial distribution images of reflected and scattered light carrying surface roughness information to obtain the scattered images of standard samples with different roughness. In this step, the spatial light scattering distribution images corresponding to different roughness values can be obtained through the built measurement system. The measurement system in this step includes a laser, a ground glass screen and a camera. A semiconductor collimated laser with a wavelength of 632.8nm is selected as a light source, and a collimated beam is emitted, which obliquely hits the surface of the workpiece to be measured at an incident angle of a certain angle, such as 30°; A frosted glass screen is placed in the reflection direction of 30°, and its height is consistent with the height of the laser to collect the spatial distribution of reflected and scattered light carrying surface roughness information; a 5-megapixel industrial camera is used to capture scattered images.

The spatial light scattering distribution image captured by the camera is shown in Figure 2, which is a band-shaped scattering image formed in space centered on the reflected light point. Among them, the surface roughness values corresponding to the six images (a)-(f) are 0.025 μm, 0.05 μm, 0.1 μm, 0.2 μm, 0.4 μm and 0.8 μm, respectively.

(2) Compress each scattering image in the sample set into a 28*28 size image, and convert the compressed image into a 1*784 matrix, and form a pixel matrix according to the number of images in the sample set as a convolutional neural network input of. In this step, the scattering image is first cropped into a 1360*1360 size image, and the noise is filtered out by Gaussian filtering, and then compressed into a 28*28 size image, and converted into a 1*784 matrix, and then according to the number of images in the sample set A pixel matrix is formed. If there are 120 scattered images in the sample set, a 120*784 pixel matrix is formed. At the same time, corresponding to the scattering images of different roughness samples in the sample set, a label matrix is constructed using roughness levels. For example, there are 120 scattering images in the sample set, corresponding to 6 roughness levels, and a 6*120 label matrix is formed.

(3) Normalize the pixel matrix, and use the normalized pixel matrix as the input of the convolutional neural network, and use the label matrix as the output of the convolutional neural network.

(4) Set the basic parameters such as the number of convolutional layers and downsampling layers, the size of the convolution kernel, and the downsampling rate in the convolutional neural network. Set training parameters such as the learning rate, the number of total samples for batch training, and the number of iterations in the convolutional neural network.

(5) Initialize the convolution kernel and bias in the convolutional neural network, and then start training the convolutional neural network classification model. Among them, the forward propagation, back propagation and weight modification of the network are all automatically completed by the computer.

(6) Test the convolutional neural network prediction model after training, and analyze the accuracy of the model for surface roughness classification. Similarly, the scattering images in the test samples were cropped, Gaussian filtered, compressed, matrix transformed and normalized, and used as a test set to analyze the accuracy of the model classification. The classification accuracy of this model can reach 100%, and the prediction results are shown in Table 1.

Table 1. Convolutional neural network prediction results

(7) During the actual measurement, use the measurement system to collect the scattering image of the workpiece surface, and through cropping, Gaussian filtering, compression, matrix conversion and normalization, according to the trained optimal convolutional neural network classification model, the obtained Measure the surface roughness level of the workpiece.

Claims (7)

1. a kind of surface roughness classification method based on convolutional neural networks, which comprises the steps of:

(1) sample acquisition；

(2) it compresses image and is converted to picture element matrix, construct label matrix；

(3) picture element matrix is normalized；

(4) basic parameter and training parameter of CNN are set；

(5) CNN and training are initialized；

(6) convolutional neural networks prediction model, that is, CNN disaggregated model after the completion of test training, analyzes the model to rough surface Spend the accuracy of classification.

2. a kind of surface roughness classification method based on convolutional neural networks according to claim 1, which is characterized in that Specific step is as follows for sample acquisition in the step (1):

(1.1) it is slanted by collimated laser beam to workpiece surface, the reflection of surface roughness information is carried using camera shooting Light and scattering light spatial distribution image, obtain the dispersion image of different roughness standard sample；

(1.2) dispersion image obtained from step (1.1) is as sample set.

3. a kind of surface roughness classification method based on convolutional neural networks according to claim 1, which is characterized in that Compressed picture and picture element matrix is converted in the step (2), constructing label matrix, specific step is as follows:

By each width dispersion image boil down to 28*28 in sample set, compressed image is changed into the matrix of 1*784, and root According in sample set amount of images constitute picture element matrix, then further according to dispersion image in sample set correspond to roughness sample block etc. Grade building label matrix, using their outputting and inputting as convolutional neural networks.

4. a kind of surface roughness classification method based on convolutional neural networks according to claim 1, which is characterized in that Specific step is as follows for the basic parameter of setting CNN and training parameter in the step (4):

(4.1) basic parameter of convolutional neural networks is set；The basic parameter includes the quantity of convolutional layer and down-sampled layer, volume Product core size and the down-sampled range of decrease；

(4.2) training parameter of convolutional neural networks is set；The training parameter includes the quantity of learning rate, batch total sample of training And the number of iterations.

5. a kind of surface roughness classification method based on convolutional neural networks according to claim 1, which is characterized in that Initialization CNN and specific step is as follows for training in the step (5):

(5.1) convolutional neural networks are initialized；The initialization content includes convolution kernel and biasing；

(5.2) training convolutional neural networks；The training process includes forward-propagating, backpropagation and the weight modification of network.

6. a kind of surface roughness classification method based on convolutional neural networks according to claim 2, which is characterized in that Collimated laser beam is slanted with the incidence angle of set angle to workpiece surface, in the reflection side of respective angles in the step (1.1) Upward placing height and the consistent ground glass of laser height use the dispersion image on camera shooting ground glass.

7. a kind of surface roughness classification method based on convolutional neural networks according to claim 1, which is characterized in that The grade that surface roughness is classified in the step (6) has 0.025 μm, 0.05 μm, 0.1 μm, 0.2 μm, 0.4 μm and 0.8 μm six A grade.