CN109658396A - A kind of casting x-ray image defect identification method based on selective search and CNN - Google Patents

Abstract

The invention belongs to casting to manufacture process matched therewith correlative technology field, and disclose a kind of casting x-ray image defect identification method based on selective search and CNN comprising: radioscopic image is shot for each casting of pending quality testing and carries out image preprocessing；All suspected defects whole frames in image are elected using selective search algorithm；Model image database is established, constructs defect classification model based on convolutional neural networks, and training is until meet precision of prediction requirement；Defect classification model based on above-mentioned building is predicted each defect classification and is marked, while calculating defect relative size and position according to image pixel information, obtains the report of casting radioscopic image defects detection.Through the invention, all kinds of defects in casting radioscopic image can more accurate, be easily detected automatically, and determine defect classification automatically, reduce labor workload, and partial detection can directly be connected with production, to provide data feedback for optimization casting technique.

Description

A kind of casting x-ray image defect identification method based on selective search and CNN

Technical field

The invention belongs to casting to manufacture process matched therewith correlative technology field, be searched more particularly, to one kind based on selectivity The casting x-ray image defect identification method of rope and CNN (convolutional neural networks).

Background technique

Casting is one of fundamental technology of Modern Manufacturing Industry.The safety that all kinds of defects of casting seriously affect cast article makes With, therefore realize that casting defect detection is to control the important link of casting quality.With the high speed development of non-destructive testing technology, respectively Class detection technique is applied widely in field of industry detection, such as ultrasound detection, infrared detection and ray detection, these inspections Result is surveyed to promote the quality safety of product and providing data support；Wherein x-ray imaging detection, can by transmiting inside test specimen The information such as internal defects quantity, each defect form and position distribution are efficiently obtained, therefore obtain this field more and more Attention and concern.

However, further investigations have shown that, existing casting radioscopic image detection scheme still has defect below or not Foot: firstly, way conventional at present is, when casting is after X-ray real time scan and imaging, corresponding egative film physical printed is gone out Come, professional carries out artificial detection to imaging film again, therefore, it is determined that the type of defect, quantization size and record position information Deng；This kind of manual inspection method there are personnel's subjectivity dependences strong, defect the problems such as there are erroneous detection missing inspections, and process steps are more Cumbersome, defects detection is inefficient, and information preservation and inquiry are extremely inconvenient, is unable to the quality condition that Real-time Feedback is detected casting. Secondly, the link and data information being related to are more for the whole production line of moulding casting, quality testing link is made For one of the key link for examining casting process achievement, keep the timely relevance of testing result and whole production procedure information outstanding To be important, so as to provide accurate data feedback in time for process optimization.Correspondingly, this field needs to find and targetedly solve Certainly scheme, preferably to meet the above technical need faced in actual production practice.

Summary of the invention

For the above insufficient or Improvement requirement of the prior art, the present invention provides one kind to be based on selective search and CNN Casting x-ray image defect identification method, wherein the related data feature and technique that pass through moulding by casting production line of combining closely Demand analysis, selection introduces specific picture search and defect class prediction model from many algorithms, accordingly not only can be more For accurate, feedback casting quality testing result information, tracking retrospect obtain the root and crucial mistake of casting quality defect in time Journey, and then the regularity that entire pouring technology procedure quality problem generates can be quickly grasped, therefore be particularly suitable for modernizing Need to execute casting quality the application of high efficiency, high quality on-line monitoring in manufacturing process.

To achieve the above object, it is proposed, according to the invention, provide a kind of casting x-ray image based on selective search and CNN Defect identification method, which is characterized in that this method includes the following steps:

(S1) radioscopic image is shot for each casting of pending quality testing, and image filtering and increasing is executed to it Strength reason；

(S2) all suspected defects whole frames in image are elected candidate and given using selective search algorithm and compile Number, wherein the suspected defects had not only included real defect region but also had included gray scale sudden change region caused by casting structure design；

(S3) by casting quality defect be divided into high density is mingled with, low-density is mingled with, stomata, shrinkage cavity, incomplete fusion, crackle, Shrinkage defect and casting structure background totally 8 classifications, then handle defect area in image, under giving defective labels and shearing Come, establish image data set, data set is further divided into training set according to special ratios and verifying collects；Then, it is based on convolution Neural network constructs defect classification prediction model, then training set of images is sent into model and is trained and optimizes, and then tests Card collection verifying model accuracy is until meet the requirements；

(S4) pre- after the candidate region that above-mentioned steps (S2) frame is elected to be input to (S3) optimization through the above steps Model is surveyed, the generic of casting quality defect is accordingly obtained and gives location position, thus complete whole casting x-ray image Defect recognition process.

As it is further preferred that in step (S1), the enhancing processing preferably include brightness enhancing, coloration enhancing, The operation such as contrast enhancing and acutance enhancing.

As it is further preferred that it is also preferable to include obtain casting quality defect relative size size in step (S4) Operation, the Process Design is as follows: after determining the defect classification of the candidate region, preferably being marked lacking in image Side is fallen into, while printing the coordinate value of two reference points on each defect circumscribed area box lower-left and upper right diagonal line, thus Position of the defect in candidate region is given and is positioned；Then the coordinate value of the two reference points is made the difference to calculate external box Width it is high, thus the size to defect in candidate region, which is given, characterizes.

As it is further preferred that it is also preferable to include be single with a complete casting radioscopic image in step (S4) Position, the corresponding operation for generating casting defect examining report, and this report storage is spare in the database.

In general, through the invention it is contemplated above technical scheme is compared with the prior art, mainly have below Technological merit:

1, this method scans for and classifies automatically according to target object feature in image, departing from artificial subjectivity, subtracts Missing inspection and erroneous detection situation are lacked；Furthermore entire detection process is not necessarily to printing egative film into independent testing and management, saves manpower Material resources cost, while manual labor amount is reduced, shorten detection time, improves detection efficiency；

2, the present invention has also further been carried out by the particular model respective algorithms to casting defect class prediction process excellent Change design, more actual test shows that its phase should be able to significantly improve the prediction accuracy and defective bit of casting defect type Set the data precision with size；

3, the detection method real-time detection casting goes out electronic defects detection report, report storage and filing inquiry and casting Part information is directly corresponding, and partial detection can be extracted individually to be associated with castings production manufacture, timely for optimization casting technique Accurate data feedback is provided.

Detailed description of the invention

Fig. 1 is the integrated artistic according to casting x-ray image defect identification method constructed by the preferred embodiment for the present invention Flow diagram；

Fig. 2 is more specifically to show the treatment process for constructing disaggregated model and prediction using CNN according to the present invention Schematic diagram.

Specific embodiment

In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and It is not used in the restriction present invention.As long as in addition, technical characteristic involved in the various embodiments of the present invention described below Not constituting a conflict with each other can be combined with each other.

Fig. 1 is the integrated artistic according to casting x-ray image defect identification method constructed by the preferred embodiment for the present invention Flow diagram.As shown in fig. 1, which mainly includes following scheme step:

Firstly, each casting for pending quality testing shoots radioscopic image, and image filtering and increasing are executed to it The pretreatments such as strength reason.

More specifically, it can choose suitable image filtering according to casting radioscopic image target signature and be combined with enhancing, To improve the picture quality of subsequent processing.In view of there are the lesser defects of size for casting when filtering, so in selection filter Or pay attention to retaining image detail when window size.It is gray level image for casting radioscopic image, it is bright can suitably enhances image Degree, coloration, contrast and acutance improve subsequent detection effect to improve the comparison of image object and background.

Then, selective search algorithm preferably can be used, all suspected defects whole frames in image is elected into candidate simultaneously It gives and numbers, wherein the suspected defects had not only included real defect region but also dashed forward comprising gray scale caused by casting structure design Become region；

More specifically, casting radioscopic image after treatment, can detecte the similarity of target in image, including line Similarity, scale similarity and overlapping similarity etc. are managed, is selected suspicious region whole frame in image based on selective search algorithm It out, and for example can be according to image from left to right serial number from top to bottom.Target herein both includes real defect, it is also possible to Including gray scale sudden change region caused by casting structure design, all candidate regions remain subsequent exhaustive division.

Then, the step of being building and training the defect class prediction model based on CNN.

If Fig. 2 is more particularly shown in, in the present invention by casting quality defect be divided into high density is mingled with, low-density is mingled with, Then stomata, shrinkage cavity, incomplete fusion, crackle, shrinkage defect and casting structure background totally 8 classifications handle defect area in image, Given defective labels are simultaneously shear off, and establish image data set, further according to special ratios by data set be divided into training set and Test set；Then, constructed based on convolutional neural networks defect classification prediction model, then by training set of images be sent into model in into Row training and optimization, then verifying collection verifying model accuracy is until meet the requirements；Model structure mainly includes input layer, 5 layers of volume Product pond layer, 3 layers of full articulamentum and output layer, this building principle is known in the art, therefore details are not described herein.

More specifically, before carrying out model training, first prepare image data set.The candidate that image inside casing is elected Region gives label, specific category be high density is mingled with, low-density is mingled with, stomata, shrinkage cavity, incomplete fusion, crackle, shrinkage defect and 8 classifications such as casting structure background, classification foundation are the gray scale difference of target of all categories, boundary rule degree, edge shape and interior Portion's textural characteristics etc., above-mentioned data preferably in proportion 4:1 points are training set and verifying collection.

In addition, constructing defect classification model based on convolutional neural networks (CNN), model mainly includes input layer, convolution pond Change layer, full articulamentum and output layer.It with target image is X when training, label is that Y carries out model training.After model training is good Again by verifying collection image input, precision of prediction is verified, improves model structure or regularized learning algorithm rate if precision is unsatisfactory for requiring Etc. parameters optimize, until reach demand precision.

Finally, being input to the candidate region that above-mentioned frame is elected by the prediction model after optimization, casting is accordingly obtained The generic of part mass defect simultaneously gives location position, thus completes whole casting x-ray image defect recognition process.

More specifically, the image that will have been subjected to pretreatment and region detection is sent into trained above-mentioned model, exports and is Above-mentioned 8 classifications, category label print the external box lower-left of each defect and upper right diagonal line two in image by defect side Point coordinate value, to the positioning of defect on the image.And this two o'clock coordinate value work is poor, the width for calculating external box is high, with this The size of value characterization defect.In summary defect information obtained, as unit of a complete casting radioscopic image, Corresponding to generate a casting defect examining report, report name and present image title are of the same name, and report content is remembered by defect classification Number, relative position and the size of each defect are recorded, and carries out summarizing rear statistical magnitude.

Another preferred embodiment according to the invention can also increase after the prediction for completing casting quality defect The trace back step of mass defect.The step is more particularly to as follows: by information such as the mass defects comprising defective classification and position It is mapped in the database, and makes this mass defect information associated with the data in production system and process system, Design parameter of the casting in system is specified, and then casting quality is traced come online based on the existing recommended parameter range of casting The concrete reason and relevant information of defect.

More specifically, by quality information map to casting current life cycle historical record by the problem casting with Similar casting such as casting process card information in process of production and monitoring of tools information compare, and explore multiple casting The regularity that defect problem generates, the root that analysis quality problems generate.

Such as, it is desirable to it checks the relevant information of specific casting number casting, corresponding number can be inputted in precise search column, It can check that all information of the casting of typing in resource management system, such as technological design scheme and history production are being made Monitoring information checks numerical difference away from exceeded parameter, to trace back to by comparing process design parameter and actual production information The possibility root of casting generation defect.

To sum up, recognition methods according to the invention, by the particular model to casting defect class prediction process and accordingly The optimization design of algorithm accordingly can be improved the prediction accuracy of casting defect type and the data essence of defective locations and size Degree, while being collectively formed quality analysis system with other records of production process, so efficiently, explore convenient for manipulation and wherein advise Rule property problem.In addition, whole operation process can carry out automatically, mitigate manual labor amount, promotes detection accuracy and detection efficiency, and Qualitative data feedback is provided for castings production process modification to support.

As it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention, not to The limitation present invention, any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should all include Within protection scope of the present invention.

Claims (4)

1. a casting X-ray image defect identification method based on selective search and CNN, is characterized in that, the method comprises the following steps: (S1) photographing an X-ray image for each casting to be inspected for quality, and performing image filtering and enhancement processing on it; (S2) using a selective search algorithm to select all the suspicious defects in the image as candidates and give them a number, wherein the suspicious defects include both the real defect area and the gray-scale mutation area caused by the structural design of the casting; (S3) Divide the casting quality defects into 8 categories: high-density inclusions, low-density inclusions, pores, shrinkage cavities, unfused, cracks, shrinkage defects and casting structural background, and then process the defect areas in the image, and give the defect labels And cut it down, establish an image data set, and further divide the data set into a training set and a validation set according to a specific ratio; then, build a defect classification prediction model based on a convolutional neural network, and then send the image training set into the model for training and evaluation. Optimization, and then the validation set verifies the model accuracy until it meets the requirements; (S4) Input the candidate area selected by the box in the above step (S2) into the prediction model optimized by the above step (S3), correspondingly obtain the category of the casting quality defect, and give the position calibration, thereby completing the overall casting X-ray image defect identification process. 2 . The method for identifying defects in X-ray images of castings as claimed in claim 1 , wherein, in step ( S1 ), the enhancement processing preferably includes operations such as brightness enhancement, chroma enhancement, contrast enhancement, and sharpness enhancement. 3 . 3. casting X-ray image defect identification method as claimed in claim 1 or 2, is characterized in that, in step (S4), preferably also comprises the operation that draws casting quality defect relative size, and this process is designed as follows: when After determining the defect category of the candidate area, it is preferable to mark it on the defect side in the image, and print the coordinate values of the two reference points on the lower left and upper right diagonal lines of the bounding area of each defect, so as to identify the defect. The position in the candidate area is positioned; then the coordinate values of the two reference points are differenced to calculate the width and height of the circumscribed box, thereby characterizing the size of the defect in the candidate area. 4. The casting X-ray image defect identification method according to any one of claims 1-3, characterized in that, in step (S4), preferably also comprising taking a complete casting X-ray image as a unit, correspondingly generating a casting The operation of defect detection report, and this report is stored in the database for future use.