CN110335244A - A kind of tire X-ray defect detection method based on more Iterative classification devices - Google Patents

Abstract

The tire X-ray defect detection method based on more Iterative classification devices that the invention discloses a kind of, model realization of this method based on more Iterative classification devices, model includes three parts: for obtain pretreatment after image characteristic pattern convolutional layer, for extracting the region candidate network of candidate region in characteristic pattern and the sorter network of iteration；The sorter network of iteration uses three layers of Fast R-CNN；Every layer of Fast R-CNN includes area-of-interest pond layer and sorter network, the candidate frame that the sorter network of upper one layer of Fast R-CNN obtains is sent into the pond layer interested of next layer of Fast R-CNN in three layers of Fast R-CNN, class probability accesses the flattening layer of next layer of Fast R-CNN classifier, and iteration obtains accurate defect type and position.The method of the present invention based on classifier successive ignition, model performance can be made more preferable, the precision ratio of defect is higher.

Description

A kind of tire X-ray defect detection method based on more Iterative classification devices

Technical field

The invention belongs to computer visions and industrial detection technical field, are related to a kind of tire based on more Iterative classification devices X-ray defect detection method is to be allowed to through computer by the algorithm of target detection of classifier successive ignition come detection wheel Defect in tire x-ray image.

Background technique

Tire shooting X-ray belongs to the last one quality monitoring stage of tire production, and present major manufacturer is general It is all the quality for taking class Three's rotation system to control this stage.Specific process is that tire is sent to X-ray machine channel, and X-ray machine is given Tire shoots x-ray image, and quality-monitoring personnel just see x-ray image in front of the display, and then quality inspection personnel will be sentenced according to X-ray Whether disconnected have defect, and if there is then this current tires can be sent to special channel and handled defect, otherwise tire enters Next process.The quality inspection mode of country tire production quotient is primarily present following problems at present:

Firstly, the x-ray image of a tire is generally all bigger under efficiency is relatively low, skilled quality inspection personnel is complete Differentiate whether a tire has defect to require to consume tens of seconds time；Secondly, differentiate accuracy it is not high, judge by accident, fail to judge it is more, See that display screen screen people is easy fatigue for a long time, in that case it is possible to occur having the image discriminating of defect for no defect, It causes to fail to judge or defect classification judges incorrectly, cause to judge by accident.It judges by accident and fails to judge in both of these case it is crucial that solve It fails to judge problem, this is also the most concerned problem of industry, and defect classification misdeems, due to not being sold to businessman, so It not will cause safety problem, but fail to judge, once security risk may be had by having arrived in businessman's hand；Finally, carrying out people for a long time Work differentiates, damages to quality inspection personnel health larger.It is long-term eyes to be damaged against display screen, cause quality inspection personnel to occur The case where having a dizzy spell.

According to nearly 3 years statistics, the misdetection rate that personnel detect X-ray picture is about 2.3%, and so-called misdetection rate will There is the picture of defect to be determined as no defect.

In recent years, target detection (Object Detection) algorithm was developed rapidly due to the emergence of deep learning, Propose algorithm of target detection much based on convolutional neural networks.But in terms of tire X-ray check, wanted there are also very long road It walks.Position such as last detection defect is still inaccurate.

Summary of the invention

For the above problem existing for existing quality inspection mode, the present invention provides a kind of tires based on more Iterative classification devices X-ray defect detection method solves defect position with more Iterative classification devices and judges that still inaccurate, misdetection rate is not still low and asks Topic.It realizes computer generation and differentiates whether current x-ray image has defect for human eye, and be which kind of defect, at the same time can also Mark the position of defect in the picture.

A kind of tire X-ray defect detection method based on more Iterative classification devices, target are input tire X-ray picture, output Whether current x-ray image has defect, has the position of which kind of defect and defect in the picture, includes the following steps:

(1) image preprocessing, more Iterative classification devices can be inputted by converting the tire X-ray picture for being used for training and detecting to The format and size of model；

(2) model and initialization model are built.The model includes three parts:

Convolutional layer (Conv layers): original image passes through the available a series of characteristic pattern of convolutional layer

Region candidate network (Region proposal network): characteristic pattern is input to region candidate network, and region is waited Network selection network, which extracts in characteristic pattern, to obtain candidate region containing the region of defect.

The sorter network of iteration:

Including three layers of Fast R-CNN；Every layer Fast R-CNN points are two parts:

First part is area-of-interest pond layer；Area-of-interest pond is passed through in the candidate region of region candidate network output Change layer specification to unified size；

Second part is sorter network；This layer calculates the classification for whether having defect and defect using candidate region characteristic pattern, Remove the candidate region for being judged as no defect, while to being judged as that the frame for having defect carries out frame and return that obtain defect more smart True position after obtaining more accurate candidate frame, backs into the area-of-interest pond layer of next layer of Fast R-CNN, together When obtained class probability is accessed to next layer of Fast R-CNN classifier flattening layer in, carry out next iteration；

That is the candidate frame for obtaining the classifier of upper one layer of Fast R-CNN in three layers of Fast R-CNN is sent into next layer The pond layer interested of Fast R-CNN, class probability access the flattening layer of next layer of Fast R-CNN classifier, and iteration obtains Accurate defect type and position；

(3) training pattern utilizes the above-mentioned model of tire X-ray picture training for training；

(4) it by tire X-ray picture input model to be detected, exports whether current x-ray image has defect, there is which kind of disease The position of defect and defect in the picture.

The beneficial effects of the present invention are:

Tire X-ray defect detection based on target detection can avoid because of inefficiency caused by human factor, in master It solves the problems, such as to solve erroneous judgement to a certain extent in the case where failing to judge, to greatly improve the safety of tire, and can mitigate The pressure of quality inspection personnel.

Secondly, the method for the present invention based on classifier successive ignition, model performance can be made more preferable, the precision ratio of defect is more It is high.

Detailed description of the invention

Fig. 1 is the tire X-ray defect detection method flow chart based on more Iterative classification devices；

Fig. 2 is more Iterative classification device model schematics；

Fig. 3 is model training procedure chart；

Fig. 4 is ROI Pooling schematic diagram.

Specific embodiment

In order to more specifically describe the present invention, with reference to the accompanying drawing and specific embodiment to the present invention carry out specifically It is bright.

This example is intended to realize by the present invention and detect to the defect of tire x-ray image.This method process includes that image is located in advance It manages, build model, training pattern, input picture to be detected and obtain result, as shown in Figure 1, the specific implementation process is as follows:

(1) image preprocessing.The original picture size of this example is the size of 20000*1900, is when marking defect Directly marked in original graph.Since algorithm calculates the limitation of power, be classified as the small figure of 11 1900*1900, small figure it Between have certain overlapping (next with about Chong Die 82 pixels of a upper picture), it is therefore desirable to by corresponding coordinate position It carries out rewriting the xml document of record defect type and coordinate from transformation.Since the defect of picture is very sparse, figure is cut almost The omission of defect is not will cause.

Voc file format is converted by image, voc file mainly includes three files.First file be Annotations, what this file the inside was mainly stored is the xml document for describing defect type and position in picture；Second A file is JPGImages, and what this document folder was mainly stored is all picture files；Third file is ImageSets, this document folder is lower, and store is four txt files, be respectively train.txt, val.txt, trainval.txt, Test.txt, this four text files record training picture respectively, verify picture, verify training picture and test picture.

(2) target detection model and initialization model are built.Model is as shown in Fig. 2, include three parts:

Convolutional layer.Multilayer convolutional neural networks of the picture Jing Guo Conv-ReLU-Pooling structure are inputted, feature is extracted Figure is used for subsequent region candidate network and sorter network.Convolutional layer uses VGG16 model, first by the original image of P × Q The size of M × N is zoomed to, 13 Conv-ReLU layers and four maximum value pond layers are then passed through.The core of all convolutional layers is all It is 3 × 3, Filling power (padding) is 1, and step-length (stride) is also 1, this is to guarantee the output after each convolution The size of characteristic pattern is consistent with input size.The core of pond layer is 2 × 2, and Filling power 0, step-length 2, this is for each pond Characteristic pattern size halves after change.The picture of M × N becomes the characteristic pattern of (M/16) × (N/16) after convolutional layer.

Region candidate network.Region candidate network is for generating candidate region.Region candidate network is practical be divided into two it is flat Row network, one obtains foreground and background i.e. by normalization exponential function classifier (softmax) classification anchor (anchors) and has Without defect, another frame recurrence offset for being used to calculate for anchor, to obtain accurate candidate region.Finally synthesis contains The anchor and frame of defect return offset and obtain candidate region, while rejecting too small and beyond boundary candidate region.

The classification layer of iteration.This example has used three layers of Fast R-CNN.Fast R-CNN, can be to candidate while classification Frame is returned, this makes the Classification Loss of candidate frame and recurrence loss combine the loss total as one and train together, Not only it improves speed and also promotes precision.

Fast R-CNN is divided into two parts again:

First part is area-of-interest pond layer.The size of characteristic pattern is different, but since sorter network makes The reason of with full articulamentum, it is desirable that input size be it is fixed, therefore region candidate network output candidate region need by Area-of-interest pond layer specification is to unified size.The concrete operations of area-of-interest pond layer are as shown in figure 3, on a left side In four black regions of figure, one maximum numerical value of each regional choice, result in formation of the final results of right figure.Specific behaviour As: according to the candidate region of input, by interested area maps to characteristic pattern corresponding position, then by the region after mapping It is divided into the section of same size, the operation in maximum value pond is finally carried out to each section.

Second part is sorter network.This layer calculates the classification for whether having defect and defect using candidate region characteristic pattern, Remove the candidate region for being judged as no defect, at the same again to be judged as have defect frame carry out frame return obtain defect more Add accurate position.After obtaining more accurate candidate frame, the area-of-interest pond of next layer of Fast R-CNN is backed into Layer, while obtained class probability being accessed in the flattening layer of the classifier of next layer of Fast R-CNN, carry out next iteration. After three stackings generation, final accurate position and more accurate defect classification can be obtained.

(3) training pattern, training pattern process is as shown in figure 4, be divided into the following steps:

Training region candidate network.The good model (VGG) of the pre-training provided is provided, repetitive exercise is started；

Using trained region candidate network, candidate region is collected.Using region candidate network obtained in the previous step, obtain Candidate region and foreground classification probability are taken, the training of the sorter network for following iteration；

The Fast R-CNN network of training iteration for the first time.Obtain candidate region and prospect probability.Net is inputted from data Layer Network.The last identification normalization exponential function classifier of training and final frame return；

Second of training region candidate network.With Fast RCNN network parameter initialization area candidate network, region is waited The learning rate of those of the shared convolutional layer of network selection network, Fast R-CNN is set as 0, that is, does not update, and only update area is waited Those of the peculiar network layer of network selection network；

The trained region candidate network of previous step is utilized again, collects candidate region；

The Fast R-CNN network of second of training iteration, specific steps are identical as first time.

This is the training process recycled twice, and multiple training can allow model performance more outstanding, and misdetection rate is more It is low.But circulation does not just improve more times.

(4) tire X-ray picture to be detected is inputted into trained model, exports whether current x-ray image has defect, has The position of which kind of defect and defect in the picture.

Finally it should be noted that above example is only used to illustrate and not limit the technical solutions of the present invention, although referring to upper It states example to describe the invention in detail, it will be understood by those of skill in the art that the present invention can still be repaired Change the perhaps any modification or part replacement of equivalent replacement without departing from the spirit and scope of the present invention, should all cover In scope of the presently claimed invention.

Claims (1)

1. a kind of tire X-ray defect detection method based on more Iterative classification devices, which comprises the steps of:

(1) image preprocessing, converting the tire X-ray picture for being used for training and detecting to can input based on more Iterative classification devices The format and size of model；

(2) model based on more Iterative classification devices and initialization are constructed

The model based on more Iterative classification devices includes three parts:

Convolutional layer: for obtain pretreatment after image characteristic pattern；

Region candidate network: candidate region may be generated containing the region of defect for extracting in characteristic pattern；

The sorter network of iteration: including three layers of Fast R-CNN；Every layer Fast R-CNN points are two parts:

First part is area-of-interest pond layer；Area-of-interest pond layer is passed through in the candidate region of region candidate network output Standardize unified size；

Second part is sorter network；This layer calculates the classification for whether having defect and defect using candidate region characteristic pattern, removes It is judged as the candidate region of no defect, while to being judged as that the frame for having defect carries out frame and return that obtain defect more accurate Position after obtaining more accurate candidate frame, backs into the area-of-interest pond layer of next layer of Fast R-CNN, simultaneously will Obtained class probability accesses in the flattening layer of the classifier of next layer of Fast R-CNN, carries out next iteration；

(3) training pattern utilizes the above-mentioned model of tire X-ray picture training for training；

(4) it by tire X-ray picture input model to be detected, exports whether current x-ray image has defect, there is which kind of defect, and The position of defect in the picture.