1. An automatic identification method for X-ray inspection defects of electronic components is characterized by comprising the following steps: the method comprises the following steps: carrying out denoising and enhancement pretreatment on the X-ray image; carrying out contrast enhancement on the acquired X-ray image of the electronic component by a histogram equalization and gray level conversion method;

step two: manually and semi-automatically or automatically labeling the sample, wherein the sample comprises a component type, a defect type and a defect position;

step three: classifying the defect types of the electronic components to be detected according to the encapsulation and defect forms of the electronic components; the defect types of the electronic components to be detected are classified into three types, namely cavity defects, consistency defects and angle defects;

step four: detecting the four types of hollow defects in the step three by using a semantic segmentation method based on a convolutional neural network;

step five: realizing consistency defect detection by using a registration method based on a convolutional neural network;

the consistency between the element to be detected and the standard qualified element is contrastively analyzed, so that the consistency defect is detected and distinguished; the method is realized by adopting an image registration method based on a convolutional neural network;

step six: detecting and identifying chip mounting inclination and undercut defects;

the detection process comprises the steps of firstly detecting a chip region according to the characteristics of an input image, further carrying out edge detection and extraction by adopting a Robinson algorithm, analyzing the installation angle and the undercut angle of a chip according to the extracted edge information, and finally judging and outputting according to a relevant judgment criterion; the judgment criterion of the defects is as follows: mounting and bonding semiconductor chips with an inclination of not more than 10 DEG with respect to a normal mounting surface;

in the process of judging the defects, the chip area is enhanced by a gray level conversion method, and corresponding detection and analysis are carried out on the basis of the chip area.

2. The method for automatically identifying the X-ray inspection defects of the electronic components as claimed in claim 1, wherein the method comprises the following steps: in the fourth step, the void defects comprise a single-layer chip bonding area void and a chip bonding area void of the multilayer hybrid integrated circuit; for the single-layer chip welding area cavity, the cavity area is divided by adopting a convolution-based neural network algorithm, the chip welding area is detected by a gray projection method, the areas of the cavity, the area and the welding area are respectively and automatically calculated by using a pixel statistics method, and finally the qualification is judged according to a judgment criterion through an area ratio, wherein the area of the welding area cavity exceeds 1/2 of the whole welding area; or when the single cavity transversely or longitudinally penetrates through the chip welding area and the area of the cavity exceeds 10 percent of the chip welding area, the single cavity is judged to be unqualified;

for the cavity of the chip welding area of the multilayer hybrid integrated circuit, the interference of the cavity of the welding interface of the substrate and the tube shell is eliminated through the combination of multi-angle X-ray images; training a network facing to the segmentation of the welding area cavity by accurately marking the image, and realizing the segmentation of the welding area cavity for images at different angles; according to the position relation of the cavities of different layers on the multi-angle image, eliminating the interference of the cavities of the substrate bonding layer to obtain a segmentation result; shooting X-ray images of the chip along the vertical direction and 45-degree inclination angles respectively, and detecting cavity areas of the images at different angles respectively by adopting a convolutional neural network-based method; correcting the oblique image through perspective transformation, performing matching analysis on the oblique image and the image in the vertical direction, removing cavities on the welding interface of the substrate and the tube shell according to the difference of the cavities on different layers in multi-angle shooting, detecting real cavities of the chip and the substrate welding area, and further judging the qualification of the chip according to related judgment rules; displacement of the chip and the cavity of the welding area of the substrate between imaging results at different angles, and fixing the positions of the cavity of the welding interface of the substrate and the tube shell on each figure;

for the sealing area cavity detection, the cavity area is divided by adopting an algorithm based on a convolutional neural network, meanwhile, the sealing area detection is carried out by adopting gray mapping, and the widths of the cavity area and the sealing area are respectively and automatically calculated; judging the qualification of the chip according to corresponding rules, namely judging the chip to be unqualified when the width of the cavity area exceeds 75% of the width of the sealing area;

for solder ball cavity detection, performing cavity segmentation based on a convolutional neural network, extracting the contour of a solder ball by adopting an algorithm of Robinson edge detection, and calculating the diameter of the solder ball; judging the qualification of each solder ball by comparing the diameter of the cavity with the diameter of the solder ball; when the diameter of the hollow exceeds 25% of the diameter of the solder ball, the solder ball is judged to be defective.

3. The method for automatically identifying the X-ray inspection defects of the electronic components as claimed in claim 1, wherein the method comprises the following steps: the image registration method based on the convolutional neural network in the fifth step is realized by the following process: firstly, selecting a standard qualified product image for each packaged device, and then accurately registering an image to be detected and the standard image, wherein when an element to be detected has the defects of redundancy, broken bonding wire and wrong element installation, the image to be detected and the registered qualified product image still have difference; effective detection of defect positions and types thereof is realized by comparing and analyzing differences; and training the convolutional neural network through the sample to realize accurate detection of various consistency defects.

4. The automatic identification and detection system for the X-ray inspection defects of the electronic components, which is built by the method of claim 1, is characterized in that: the detection system comprises: the device comprises a cavity defect detection unit, a consistency defect detection unit and an angle defect detection unit;

the hole defect detection unit adopts a semantic segmentation algorithm based on a convolutional neural network and is designed based on a DeeplabV3+ basic architecture; the method comprises the steps of utilizing a semantic segmentation algorithm based on a convolutional neural network to achieve segmentation of various cavity regions, further performing segmentation of relevant background regions according to specific packaging forms and defect types, training the convolutional neural network through samples, achieving accurate segmentation of various cavity defects, and improving automatic defect identification efficiency.

5. The automatic defect identification and detection system for X-ray inspection of electronic components as claimed in claim 4, wherein: for the single layer of the cavity in the chip welding area, firstly, the algorithm based on the convolutional neural network is adopted to segment the cavity area, meanwhile, the chip welding area is detected by a gray projection method, the areas of the cavity area and the welding area are respectively and automatically calculated by a pixel statistics method, and finally, the qualification is judged according to the corresponding judgment criterion through the area ratio.

6. The automatic defect identification and detection system for X-ray inspection of electronic components as claimed in claim 4, wherein: for the multilayer of the chip welding area cavities of the hybrid integrated circuit, in order to eliminate the interference problem of the cavities of the welding interface of the substrate and the tube shell, the detection is realized by combining a multi-angle X-ray image; firstly, training a network facing to the segmentation of the welding area cavity by accurately marking images, and realizing the segmentation of the welding area cavity for images at different angles; and according to the position relation of the cavities on different layers on the multi-angle image, eliminating the interference of the cavities on the bonding layer of the substrate and obtaining an accurate segmentation result.

7. The automatic defect identification and detection system for X-ray inspection of electronic components as claimed in claim 4, wherein: for the detection of the cavity in the sealing area, firstly, the algorithm based on the convolutional neural network is adopted to carry out the segmentation of the cavity area, meanwhile, the gray mapping method is adopted to carry out the detection of the sealing area, and the widths of the cavity area and the sealing area are respectively and automatically calculated; and finally, judging the qualification of the chip according to a corresponding rule.

8. The automatic defect identification and detection system for X-ray inspection of electronic components as claimed in claim 4, wherein: for the detection of the solder ball cavity, firstly, the cavity segmentation based on the convolutional neural network is carried out, meanwhile, the outline of the solder ball is extracted by adopting the Robinson edge detection algorithm, and the diameter of the solder ball is calculated; finally, the qualification of each solder ball is judged by comparing the diameter of the cavity with the diameter of the solder ball.

9. The automatic defect identification and detection system for X-ray inspection of electronic components as claimed in claim 4, wherein: the consistency defect detection unit is realized by adopting an image registration method based on a convolutional neural network, and consistency defect detection and judgment can be realized by comparing and analyzing consistency between the element to be detected and the standard qualified element; and selecting a standard qualified product image for each packaged device, and then accurately registering the image to be detected and the standard image, wherein when the element to be detected has the defects of redundancy, broken bonding wire and wrong element installation, the image and the registered qualified product image still have large difference.

10. The automatic defect identification and detection system for X-ray inspection of electronic components as claimed in claim 4, wherein: the angle defects comprise detection and identification of chip installation inclination and undercut defects; detecting a chip region according to the characteristics of an input image, carrying out edge detection and extraction by adopting a Robinson algorithm, analyzing the installation angle and the undercut angle of the chip according to the extracted edge information, and finally judging and outputting according to a relevant judgment criterion.

Background

Military electronic components are basic elements of national defense and military industry electronic systems, and mainly comprise microelectronic devices such as semiconductor discrete devices and integrated circuits, and special devices such as relays and sensors. Military electronic components are the material guarantee of information systems such as nuclear, aviation, aerospace, ships, weapons, electronics and the like, and the stability and reliability of the military electronic components directly influence the function and performance indexes of the information systems.

Packaging defects of military electronic components are one of the main factors affecting the quality of military electronic components. X-ray inspection is a means for nondestructive detection of internal defects of components, can quickly detect defects generated in the processes of bonding, welding, sealing and the like, and is required to be carried out in component screening, identification and inspection, DPA (differential Power analysis) and failure analysis. The X-ray detection carries out X-ray imaging on the electronic component, the image of the X-ray imaging can reflect the internal defects of the electronic component, the current X-ray detection mainly depends on manual identification, and the X-ray image is browsed by detection personnel and is manually interpreted and analyzed according to the detection standard. However, with the continuous development of the processing and manufacturing level of electronic components, the packaging forms of the electronic components are more and more, the packaging density is higher and more, the detection number and tasks of the electronic components are increased, the speed and accuracy of manual identification cannot be met, the manual interpretation analysis mode is influenced by subjective factors such as experience of testers and physical conditions, the detection efficiency and reliability are low, the detection quality of the components is difficult to effectively guarantee, risks and hidden dangers are brought to the use of the components, and the increasing requirements of military electronic systems cannot be met. Therefore, it is necessary to form an automatic defect identification method for electronic component X-ray inspection by means of computer vision inspection technology.

With the development of computer vision and artificial intelligence technology, the computer vision detection technology based on deep learning has the characteristic of automatically learning sample characteristics due to a processing mechanism of simulating human brain vision, has greater performance advantages in the field of target detection and identification than other traditional methods, and is widely applied to various fields such as aviation, aerospace, electronics, medicine, industrial manufacturing and the like.

Aiming at the requirement of automatic identification of X-ray inspection defects of electronic components, the method for automatically identifying X-ray inspection defects of electronic components based on the deep learning visual inspection technology is deeply researched, and is an urgent requirement for quality detection of military electronic components.

At present, in China, the defect detection of electronic components mainly focuses on the related researches of detecting welding spot defects of Printed Circuit Boards (PCBs), detecting mounted components, packaging SMT surfaces and the like, and few methods for detecting chip images of the components are available.

Disclosure of Invention

In view of the above problems, the present invention is proposed to provide an automatic X-ray defect detection method for electronic components, which overcomes or at least partially solves the above problems, and mainly utilizes a deep learning visual inspection technique to obtain automatic X-ray defect detection methods for electronic components in different defect forms, thereby realizing automatic detection functions for electronic components in different defect forms.

The invention provides an automatic identification method for X-ray inspection defects of electronic components,

the method comprises the following steps: carrying out denoising and enhancement pretreatment on the X-ray image;

and carrying out contrast enhancement on the acquired X-ray image of the electronic component by using a histogram equalization and gray level conversion method.

Step two: manually and semi-automatically or automatically labeling the sample, wherein the sample comprises a component type, a defect type and a defect position;

step three: and classifying the defect types of the electronic components to be detected according to the packaging and defect forms of the electronic components. The defect types of the electronic components to be detected are classified into three types, namely a cavity defect, a consistency defect and an angle defect.

Step four: and (3) detecting the four types of hole defects in the step three by using a semantic segmentation method based on a convolutional neural network.

The hollow defects comprise a single-layer chip welding area hollow and a chip welding area hollow of the multilayer hybrid integrated circuit; for the single-layer chip welding area cavity, the cavity area is divided by adopting a convolution-based neural network algorithm, the chip welding area is detected by a gray projection method, the areas of the cavity, the area and the welding area are respectively and automatically calculated by using a pixel statistics method, and finally the qualification is judged according to a judgment criterion through an area ratio, wherein the area of the welding area cavity exceeds 1/2 of the whole welding area; or the single hollow transversely or longitudinally penetrates through the chip bonding area, and the area of the hollow is more than 10% of the chip bonding area, the chip bonding area is judged to be unqualified.

For the cavity of the chip welding area of the multilayer hybrid integrated circuit, the interference of the cavity of the welding interface of the substrate and the tube shell is eliminated through the combination of multi-angle X-ray images. And training a network facing the segmentation of the welding area holes by accurately marking the images, and realizing the segmentation of the welding area holes for the images at different angles. And eliminating the interference of the cavities of the substrate bonding layer according to the position relation of the cavities of different layers on the multi-angle image to obtain a segmentation result. And shooting X-ray images of the chip along the vertical direction, the left 45-degree inclination angle and the right 45-degree inclination angle respectively, and detecting the cavity regions of the images at different angles respectively by adopting a convolution neural network-based method. Correcting the oblique image through perspective transformation, performing matching analysis with the image in the vertical direction, eliminating the cavities on the welding interface of the substrate and the tube shell according to the difference of the cavities on different layers in multi-angle shooting, detecting real cavities in the welding area of the chip and the substrate, and judging the qualification of the chip according to related judgment rules. The displacement of the chip and the substrate welding area hollow hole between different angle imaging results is more obvious, and the positions of the substrate and the tube shell welding interface hollow hole on each figure are relatively fixed.

For the sealing area cavity detection, the cavity area is divided by adopting an algorithm based on a convolutional neural network, meanwhile, the sealing area detection is carried out by adopting gray mapping, and the widths of the cavity area and the sealing area are respectively and automatically calculated; and judging the qualification of the chip according to corresponding rules, wherein when the width of the cavity area exceeds 75% of the width of the sealing area, the chip is judged to be unqualified.

And for the solder ball cavity detection, performing cavity segmentation based on a convolutional neural network, extracting the contour of the solder ball by adopting an algorithm of Robinson edge detection, and calculating the diameter of the solder ball. And judging the qualification of each solder ball by comparing the diameter of the cavity with the diameter of the solder ball. When the diameter of the hollow exceeds 25% of the diameter of the solder ball, the solder ball is judged to be defective.

Step five: and realizing consistency defect detection by using a registration method based on a convolutional neural network.

The consistency between the element to be detected and the standard qualified element is contrastively analyzed, so that the consistency defect is detected and distinguished; the method is realized by adopting an image registration method based on a convolutional neural network. Firstly, selecting a standard qualified product image for each packaged device, and then accurately registering an image to be detected and the standard image, wherein when an element to be detected has the defects of redundancy, broken bonding wire and wrong element installation, the image to be detected and the registered qualified product image still have large difference. And the effective detection of the defect position and the defect type is realized by comparing and analyzing the difference. The convolutional neural network is trained through a large number of samples, and accurate detection of various consistency defects is achieved.

Step six: and detecting and identifying chip mounting inclination and undercut defects.

In the detection process, a chip region is detected according to the characteristics of an input image, edge detection and extraction are carried out by adopting a Robinson algorithm, the installation angle and the undercut angle of the chip are analyzed according to the extracted edge information, and finally judgment and output are carried out according to a relevant judgment criterion. The judgment criterion of the defects is as follows: the semiconductor chip mounted and bonded is inclined by not more than 10 ° with respect to the normal mounting surface.

In the process of judging the defects, the detection of the chip is the first link. The difference between the chip and the surrounding area is small, so that the detection difficulty is high. In the research, the chip area is firstly enhanced by a gray scale conversion method, and then corresponding detection and analysis are carried out on the basis of the chip area.

The detection system for realizing the method comprises the following steps: the device comprises a cavity defect detection unit, a consistency defect detection unit and an angle defect detection unit.

The hole defect detection unit adopts a semantic segmentation algorithm based on a convolutional neural network and is designed based on a DeeplabV3+ basic architecture, and compared with the traditional segmentation algorithm, the hole defect detection unit has stronger universality. The traditional method usually needs to design a corresponding segmentation algorithm aiming at each type of defect problem and set relevant parameters according to experience. The method has the advantages that various cavity regions can be segmented by utilizing a semantic segmentation algorithm based on the convolutional neural network, further, relevant background regions are segmented according to specific packaging forms and defect types, the convolutional neural network is trained through a large number of samples, various cavity defects can be accurately segmented, and the automatic defect identification efficiency is greatly improved.

Further, for the cavity (single layer) of the chip welding area, firstly, the cavity area is divided by adopting an algorithm based on a convolutional neural network, meanwhile, the chip welding area is detected by a gray projection method, the areas of the cavity area and the welding area are respectively and automatically calculated by a pixel counting method, and finally, the qualification is judged according to the corresponding judgment criterion through the area ratio. The problem that automatic calculation cannot be carried out and the judgment can be carried out only manually in the prior art is solved.

Furthermore, for the chip welding area cavity (multilayer) of the hybrid integrated circuit, in order to eliminate the interference problem of the substrate and the shell welding interface cavity, the detection is realized by combining a multi-angle X-ray image. Firstly, training a network facing to the segmentation of the welding area holes by accurately marking images, and realizing the segmentation of the welding area holes for images at different angles. And further eliminating the interference of the cavities of the substrate bonding layer according to the position relation of the cavities of different layers on the multi-angle image to obtain an accurate segmentation result. The problem that the traditional method is only manually judged is solved.

Further, for the detection of the cavity in the sealing area, firstly, the algorithm based on the convolutional neural network is adopted to segment the cavity area, and meanwhile, the gray mapping method is adopted to detect the sealing area, so that the widths of the cavity area and the sealing area are respectively and automatically calculated. And finally, judging the qualification of the chip according to a corresponding rule. The problem that automatic calculation cannot be carried out and the judgment can be carried out only manually in the prior art is solved.

Further, for solder ball void detection, firstly, the void segmentation based on the convolutional neural network is carried out, meanwhile, the outline of the solder ball is extracted by adopting the Robinson edge detection algorithm, and the diameter of the solder ball is calculated. Finally, the qualification of each solder ball is judged by comparing the diameter of the cavity with the diameter of the solder ball.

The consistency defect detection unit is realized by adopting an image registration method based on a convolutional neural network, and consistency defect detection and judgment can be realized by comparing and analyzing consistency between the element to be detected and the standard qualified element.

Furthermore, a standard qualified product image is selected for each packaged device, the image to be detected and the standard image are accurately registered, and when the element to be detected has the defects of redundancy, broken bonding wires, wrong installation of the element and the like, the image and the registered qualified product image still have large difference. By comparing and analyzing the difference, the defect position and the type thereof can be effectively detected.

The angle defects comprise detection and identification of chip installation inclination and undercut defects. The method comprises the steps of firstly detecting a chip region according to the characteristics of an input image, further carrying out edge detection and extraction by adopting a Robinson algorithm, analyzing the installation angle and the undercut angle of a chip according to the extracted edge information, and finally judging and outputting according to a relevant judgment criterion.

Further, in the process of distinguishing the defects, the detection of the chip is the first link. The difference between the chip and the surrounding area is small, so that the detection difficulty is high. Firstly, the chip area is enhanced by a gray level conversion method, and then corresponding detection and analysis are carried out on the basis of the chip area.

Compared with the prior art, the method for automatically identifying the X-ray inspection defects of the electronic components based on the deep learning computer vision inspection technology solves the problems that the detection efficiency and reliability are low, the detection quality of the components is difficult to effectively guarantee and the like caused by only depending on manual identification in the prior art.

The semantic segmentation algorithm based on the convolutional neural network is adopted, the problem that a corresponding segmentation algorithm needs to be designed for each defect problem in the traditional method is solved, the method is high in universality, segmentation of various cavity regions can be achieved, relevant background regions can be segmented for specific packaging forms and defect types, the convolutional neural network is trained through a large number of samples, accurate segmentation of various cavity defects is achieved, and automatic defect identification efficiency is greatly improved. Meanwhile, a chip welding area, a sealing area and the like are detected by a gray projection method, a cavity area, a width, a welding area, a sealing area width and the like are respectively and automatically calculated by a pixel counting method, and finally, the qualification is judged according to a corresponding judgment criterion. The problem that automatic calculation cannot be carried out and the judgment can be carried out only manually in the prior art is solved.

The consistency between the element to be detected and the standard qualified element is contrastively analyzed by adopting an image registration method based on a convolutional neural network, so that the consistency defect is detected and distinguished; and the effective detection of the defect position and the defect type is realized by comparing and analyzing the difference. The convolutional neural network is trained through a large number of samples, and accurate detection of various consistency defects is achieved. The problem that the judgment is simply carried out manually in the prior art is solved.

For the chip welding area cavity of the multilayer hybrid integrated circuit, the interference generated by the welding interface cavity of the substrate and the tube shell is eliminated through the combination of multi-angle X-ray images, the network for cutting the welding area cavity is trained through accurately marking the images according to the position relation of different layers of cavities on the multi-angle images, and the cutting of the welding area cavity is realized for the images with different angles. The problem of interference of a cavity of a welding interface between a substrate and a tube shell in a hybrid integrated circuit is solved.

Drawings

FIG. 1 is a schematic diagram of a partition structure of a void area network according to an embodiment of the present invention;

fig. 2 is a schematic structural diagram of an image registration network according to an embodiment of the present invention.

FIG. 3 is a schematic diagram of an overall architecture of a defect detection algorithm according to an embodiment of the present invention;

FIG. 4 is a schematic diagram illustrating a process of detecting a void in a bonding pad of a single-layer chip according to an embodiment of the invention.

FIG. 5 is a schematic diagram illustrating a process of detecting voids in a bonding pad of a multi-layered hybrid integrated circuit according to an embodiment of the present invention.

FIG. 6 is a schematic view of a process of detecting voids in a sealing area of a cermet package according to an embodiment of the present invention.

Fig. 7 is a schematic diagram illustrating a solder ball void defect detection process according to an embodiment of the invention.

Fig. 8 is a schematic view of a consistent defect detection process based on image registration according to an embodiment of the present invention.

FIG. 9 is a schematic diagram of a chip mounting tilt and undercut defect detection process according to an embodiment of the invention.

Detailed Description

To further explain the technical means and effects of the present invention adopted to achieve the intended purpose, the present invention will be described in detail with reference to the accompanying drawings and preferred embodiments.

The embodiment of the invention discloses an automatic identification method for X-ray inspection defects of electronic components, which is shown in figure 1. The flow of the automatic X-ray defect detection algorithm for the electronic components is as follows:

the method comprises the following steps: preprocessing the X-ray image such as denoising and enhancing;

preprocessing is the primary link of various image analysis and identification. Due to the particularity of the imaging principle, the main problem in the analysis and identification of the X-ray image is low contrast, so that the contrast enhancement is mainly used in the early preprocessing step. Aiming at the typical characteristics of an X-ray image of an electronic component and the specific requirements of subsequent defect identification, the contrast is enhanced mainly by a histogram equalization and gray level conversion method.

Step two: manually and semi-automatically or automatically labeling the sample, wherein the sample comprises a component type, a defect type and a defect position;

step three: and classifying the possible defect types of the device to be detected according to the packaging and defect form of the device.

According to different discrimination methods, the defect types of the electronic components can be divided into three categories, namely hole defects, consistency defects, angle defects and the like. The void defects refer to defect problems caused by the fact that voids at positions such as a chip welding contact area, a sealing area or a solder ball exceed relevant criteria. The consistency defect refers to a situation that the actual condition of the chip to be detected does not accord with the designed ideal condition, such as redundancy, the number of bonding wires does not accord with or break, and components or bonding wires in the hybrid integrated circuit do not accord with. The angle defect refers to the condition that the angle of a specific position of the chip is not satisfactory, such as chip mounting inclination or undercut.

Step four: and (3) utilizing a semantic segmentation method based on a convolutional neural network to realize the detection of the hole defects.

Compared with the traditional segmentation algorithm, the semantic segmentation algorithm based on the convolutional neural network has stronger universality. The traditional method usually needs to design a corresponding segmentation algorithm aiming at each type of defect problem and set relevant parameters according to experience. The method has the advantages that various cavity regions are segmented by utilizing a semantic segmentation algorithm based on the convolutional neural network, then relevant background regions are segmented according to specific packaging forms and defect types, the convolutional neural network is trained through a large number of samples, various cavity defects can be accurately segmented, and the automatic defect identification efficiency is greatly improved.

Referring to fig. 2, based on the Deeplab V3+ basic architecture, and aiming at the characteristics of an X-ray image and the specific requirements of a cavity region segmentation task, each link of the network is optimally designed, so that the extraction and discrimination capability of the network on effective characteristics is improved from multiple angles, and a better segmentation effect is achieved.

For the chip bonding area hole (single layer), firstly, the algorithm based on the convolutional neural network is adopted to divide the hole area, meanwhile, the chip bonding area is detected by the gray projection method, the areas of the hole, the area and the welding area are respectively and automatically calculated by the pixel statistics method, and finally, the qualification is judged according to the corresponding judgment criterion by the area ratio, which is shown in figure 3.

For the chip welding area cavity (multilayer) of the hybrid integrated circuit, in order to eliminate the interference problem of the substrate and the shell welding interface cavity, the detection is realized by the combination of multi-angle X-ray images. Firstly, training a network facing to the segmentation of the welding area holes by accurately marking images, and realizing the segmentation of the welding area holes for images at different angles. And further eliminating the interference of the cavities of the substrate bonding layer according to the position relation of the cavities of different layers on the multi-angle image to obtain an accurate segmentation result. And shooting X-ray images of the chip along the conventional vertical direction and 45-degree inclination angles of the left side and the right side respectively, and detecting the cavity regions of the images at different angles respectively by adopting a convolution neural network-based method. Correcting the oblique image through perspective transformation, performing matching analysis with the image in the vertical direction, eliminating the cavity of the welding interface between the substrate and the tube shell according to the difference of the cavities of different layers in multi-angle shooting, detecting the real cavity of the welding area between the chip and the substrate, and further judging the qualification of the chip according to related judgment rules, which is shown in figure 4. The displacement of the chip and the substrate welding area hollow hole between different angle imaging results is more obvious, and the positions of the substrate and the tube shell welding interface hollow hole on each figure are relatively fixed.

For the detection of the cavity in the sealing area, firstly, the algorithm based on the convolutional neural network is adopted to carry out the segmentation of the cavity area, and meanwhile, the gray mapping method is adopted to carry out the detection of the sealing area, so that the widths of the cavity area and the sealing area are respectively and automatically calculated. Finally, the qualification of the chip is judged according to the corresponding rule, as shown in fig. 5.

For the detection of the solder ball holes, firstly, the hole segmentation based on the convolutional neural network is carried out, meanwhile, the outline of the solder ball is extracted by adopting the Robinson edge detection algorithm, and the diameter of the solder ball is calculated. Finally, the qualification of each solder ball is judged by comparing the diameter of the hollow and the diameter of the solder ball, as shown in fig. 6.

Step five: and realizing consistency defect detection by using a registration method based on a convolutional neural network.

The consistency between the element to be detected and the standard qualified element is contrastively analyzed, so that the consistency defect can be detected and distinguished. The method is realized by adopting an image registration method based on a convolutional neural network, and is shown in figure 7.

Firstly, selecting standard qualified product images for each packaged device, and then accurately registering the image to be detected and the standard image, wherein when the element to be detected has the defects of redundancy, broken bonding wire, wrong element installation and the like, the image and the registered qualified product image still have large difference. By comparing and analyzing the difference, the defect position and the type thereof can be effectively detected. The convolutional neural network is trained through a large number of samples, and accurate detection of various consistency defects is achieved. See fig. 8.

Step six: and detecting and identifying chip mounting inclination and undercut defects.

In the detection process, firstly, the chip region is detected according to the characteristics of an input image, then edge detection and extraction are carried out by adopting a Robinson algorithm, the installation angle and the undercut angle of the chip are analyzed according to the extracted edge information, and finally, judgment and output are carried out according to a relevant judgment criterion, as shown in figure 9. The judgment criterion of the defects is as follows: the inclination of the mounted and bonded semiconductor chip with respect to the normal mounting surface should not exceed 10 °.

In the process of judging the defects, the detection of the chip is the first link. The difference between the chip and the surrounding area is small, so that the detection difficulty is high. In the research, the chip area is firstly enhanced by a gray scale conversion method, and then corresponding detection and analysis are carried out on the basis of the chip area.

While the invention has been described in connection with specific embodiments thereof, it is to be understood that it is intended by the appended drawings and description that the invention may be embodied in other specific forms without departing from the spirit or scope of the invention.