Light detection system for object surface and control method thereof

Abstract

The invention relates to the technical field of optical detection of object surfaces, and discloses an optical detection system for object surfaces and a control method thereof.The optical detection system comprises a rack, a numerical controller, a transmission table, an objective table, a collector and a computer, the numerical controller, the transmission table, the objective table and the collector are installed on the rack, the computer is placed on the rack, the objective table is arranged on the transmission table, and the collector is arranged on the transmission table. The collector is arranged above the transmission table, the numerical controller is electrically connected to the collector, the transmission table and the computer, the numerical controller transmits signals collected by the collector to the computer, and the numerical controller is used for controlling operation of the transmission table; the computer is provided with an acquired signal processing system, and the acquired signal processing system is provided with a space-time convolution model and a visual identification model, and is used for carrying out image presentation and identification processing on the to-be-detected object. According to the system, through optical detection and image recognition, light irradiation uniformity of a light source, presentation of full light transmission and mask light transmission, image presentation and the like can be achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of light detection of object surface, and particularly relates to a light detection system for object surface and a control method thereof. BACKGROUND

[0002] With the increasing of electronic products, the electronic products will involve PCB boards with many electronic components. For the PCB boards with many electronic components, each electronic component needs to be soldered on the PCB board. In the production process, different degrees of defects will always occur in soldering, such as tin beads, cold joints, cracks, line scratches, pressure injuries, and lack of welding, which are caused by factors such as raw materials and temperature. The traditional detection is visual detection by detection personnel, and the personnel often have visual fatigue and missed detection, which affects the use of the PCB board. Then, the SMT (Surface Mounted Technology) detection equipment appears. The SMT detection equipment is a key tool for detecting the quality of pin soldering in electronic product manufacturing. It can perform 3D scanning of solder paste thickness and area to prevent printing defects, and can detect component mounting offset and poor welding through image recognition. These are all quality detection of PCB pin soldering by using optical principles. However, due to uneven light, it is difficult to clearly distinguish the soldering condition of the pin soldering when detecting the soldering condition of the pin soldering, and errors occur in the detection of soldering defects, which leads to that the good products are misdetected as unqualified or the unqualified products are misdetected as qualified, resulting in waste of PCB board resources or affecting the use of the back end. In the detection, the three-dimensional morphology of the PCB board with many electronic components and the corresponding soldering pins needs to be acquired and generated to better convert the physical image. However, the detection equipment in production often encounters errors or defects in three-dimensional morphology, making it difficult to detect by using three-dimensional morphology. In addition, the light projection brightness of the detection equipment in production is not uniform, the energy transfer efficiency is low, the brightness is dark, and other problems, which leads to deviation in the subsequent optical acquisition, abnormality in system processing and analysis, poor detection accuracy and stability, and the like. In addition, the light projection object is full projection, and unnecessary places are not shaded to avoid affecting the light irradiation collection of other areas.

[0003] Therefore, in view of the above problems, the existing pin soldering light detection technology needs to be further improved. SUMMARY

[0004] The purpose of the present application is to provide a light detection system for the surface of an object, which comprises a rack, a numerical controller, a transmission table, a carrier table, a collector and a computer, so as to solve the problems of poor brightness uniformity, low energy utilization, low brightness, regional light interference and the like of the existing light projection system; the computer is also provided with a collection signal processing system, and a space-time graph convolution network model and a visual recognition model are used to segment, classify and convolve the time and space of the object to be detected (such as a pin soldering object), so as to realize better detection of the object to be detected, improve the detection effect of the object to be detected, and better meet the detection needs in production.

[0005] The technical scheme of the present application is as follows: A light detection system for the surface of an object, comprising a rack, a numerical controller, a transmission table, a carrier table, a collector and a computer, wherein the numerical controller, the transmission table, the carrier table and the collector are installed on the rack, the computer is placed on the rack, the carrier table is arranged on the transmission table, the carrier table is used for placing an object to be detected, the transmission table is used for transmitting the carrier table, the collector is arranged above the transmission table, the numerical controller is electrically connected to the collector, the transmission table and the computer, the numerical controller transmits the signal collected by the collector to the computer, and the numerical controller is also used for controlling the operation of the transmission table. The collector comprises a light source, a reflector, a display assembly, a lens, an image lens and a housing, the light source, the reflector, the display assembly, the lens and the image lens are sequentially installed in the housing, the light source illuminates the object to be detected on the carrier table through the reflector, the display assembly, the lens and the image lens, the reflected light from the object to be detected on the carrier table is collected by the image lens, and the collected signal is transmitted to the computer through the numerical controller for image processing. The lens is a cemented lens. The display assembly comprises a heat shield and an LCD display device, the heat shield is arranged between the reflector and the LCD display device, and the LCD display device is electrically connected to the numerical controller. The computer is provided with a collection signal processing system, the collection signal processing system is provided with a space-time graph convolution network model, the space-time graph convolution network model can collect the graph structure of the object to be detected by the collector, and model the time distribution and the space distribution of the object to be detected based on the graph structure.

[0006] Further, the heat insulation sheet is selected from a glass sheet coated with an infrared absorption film layer or a plastic sheet coated with an infrared absorption film layer; since the light emitted by the light source carries some heat, the irradiation on the surface of the LCD display device can cause the temperature to rise, affecting the display effect of the LCD display device, etc., and if the temperature is too high, the middle part is prone to blackening or even damage, so the heat insulation sheet is used to insulate the LCD display device.

[0007] Further, the LCD display device is selected from an LCD display device compatible with full light transmission mode and mask light transmission mode.

[0008] Further, the transmission table includes a transmission motor, a driving wheel, an auxiliary driving wheel and a transmission belt, the transmission motor is installed on the rack, the driving wheel and the auxiliary driving wheel are rotatably installed on the rack, the driving end of the transmission motor is connected to the driven end of the driving wheel, the transmission belt is installed on the driving wheel and the auxiliary driving wheel, and the control end of the transmission motor is electrically connected to the control end of the numerical controller.

[0009] Further, the reflector is selected from a uniform light reflector; the reflector is used to converge and uniformize the light emitted by the light source, which can improve the utilization rate of light source light and thus improve the light energy utilization rate of the entire collector.

[0010] Further, the image lens includes a lens and an image collector, both of which are installed in the shell, and the image collector is electrically connected to the numerical controller; the lens projects the picture (with or without a mask) on the LCD display device onto the stage (projection surface); wherein the lens includes a plurality of different optical lenses, which can project the picture on the LCD display device onto the stage with a certain magnification, when the angle between the optical axis of the lens and the normal of the projection surface is β, the angle between the normal of the LCD display device and the optical axis of the lens is α, and α and β satisfy the Snell law of optical structure, a comprehensive and clear projection picture can be obtained.

[0011] Further, the collection signal processing system is also loaded with a visual recognition model, which collects the to-be-detected object through the collector and performs segmentation and classification on it.

[0012] Further, the visual recognition model includes a Haar cascade model, which uses matrix features and the pixel value difference of the matrix region through the image of the to-be-detected object collected by the collector to display the local features of the segmented image of the to-be-detected object.

[0013] Further, the visual recognition model further comprises an AdaBoost model, which combines the rectangular features of each Haar cascade model with a threshold value into a weak classifier, and then through multiple rounds of iteration, focuses on the error samples, and finally weights the weak classifiers into a strong classifier.

[0014] Further, the machine frame is made of metal.

[0015] Further, the numerical controller is provided with a display screen.

[0016] Further, the numerical controller is provided with a touch display screen.

[0017] Further, the light source is an LED light source.

[0018] Further, the light source is a matrix distributed LED light source.

[0019] Further, the LED light source is a white light LED light source, etc., which is used to provide appropriate illumination brightness for the LCD display device; wherein the LED light source is closely attached to the incident port of the reflector, and the light emitting surface of the LED light source is smaller than the incident port of the reflector, so that the utilization rate of the LED light source light can be higher.

[0020] Further, the reflector comprises a table-shaped reflector and a conical reflector, the inside of the table-shaped reflector is provided with a conical reflector, the small size end of the table-shaped reflector is close to the light source, and the large size end of the table-shaped reflector is close to the display assembly; the light source light can be reflected multiple times in the reflector, and the reflector plays the role of a light homogenizer, so that the display assembly obtains uniform light intensity distribution.

[0021] Further, the conical reflector is selected to be at least two.

[0022] Further, the conical reflector is selected to be a circular conical reflector or a prismatic conical reflector.

[0023] Further, the table-shaped reflector is selected to be an isosceles circular table-shaped reflector or an isosceles prismatic table-shaped reflector.

[0024] Further, the LCD display device can make the light from the light source to the LCD display device again obtain uniform light intensity distribution; wherein the heat insulation sheet is used for partitioning and reducing the heat of the LCD display device, so as to ensure that the LCD display device can work normally.

[0025] Further, the heat insulation sheet is selected to be a glass sheet coated with antimony tin oxide on both sides or a glass-plastic sheet coated with antimony tin oxide on both sides.

[0026] Further, the heat insulation sheet is selected from a glass sheet coated with 10 μm thick antimony tin oxide on both sides or a glass-plastic sheet coated with 10 μm thick antimony tin oxide on both sides.

[0027] Further, the antimony tin oxide is deposited on the glass sheet or the plastic sheet by a magnetron sputtering technique.

[0028] Further, the cemented lens can change the light field distribution and improve the brightness uniformity of the light, so that more light is projected into the stage far from the optical axis, thereby improving the brightness uniformity of the stage; in addition, the cemented lens can also deflect the off-axis light, so that more off-axis light enters the lens, thereby improving the brightness and uniformity of the projected light.

[0029] Further, the cemented lens comprises a convex lens and a concave lens, the convex lens is a symmetrical convex lens, and the convex surface of the convex lens and the concave surface of the concave lens are matched and bonded by glue.

[0030] Further, the concave lens is selected from a concave lens with one concave surface and the other flat surface.

[0031] Further, the image collector is selected from a CCD sensor type image collector or a CMOS sensor type image collector.

[0032] Further, the housing comprises a main housing and a lens housing, the lens housing is installed on the main housing, the light source, the reflector, the display assembly and the lens are sequentially installed in the main housing, and the image lens is installed in the lens housing.

[0033] Further, the Haar cascade model is combined with an AdaBoost model, first, the system is initialized, then the acquisition data of the collector is read, and image processing is performed, the image processing here is mainly to convert the RGB image of the object to be detected into a gray scale image, which will present different gray scale regions, then the Haar cascade classifier and the AdaBoost model are processed, the Haar cascade classifier divides the image into several squares, each square includes several pixels, then each square is processed and a threshold result is generated, the threshold can reflect the difference value of the light and dark area, and the sample comparison model is combined to process the defects; if there is no defect, the qualified product result is output, and is stored in the memory as a standard detection piece; if there is no defect in the subsequent detection, the qualified product result is output, and is continuously filled into the standard detection piece memory, so that the standard detection piece memory library is more accurate; if the subsequent detection has defects and differs from the standard detection piece in the standard detection piece memory library, the unqualified product result is output.

[0034] Further, the AdaBoost model selects a single rectangular feature that can separate positive examples from negative examples, and for each single matrix feature, the weak classifier determines a threshold classification function to achieve a reduced number of misclassified samples.

[0035] Further, Hj(x) in the AdaBoost model is composed of a feature Fj, a threshold θj and a polarity Pj, where Hj(x)={1, if PjFj(x) is less than Pjθj} or Hj(x)={0, if PjFj(x) is not less than Pjθj}, where x is an AxB pixel sub-window, A and B are the length and width of the pixel sub-window, respectively, and A and B are selected to be equal or not equal.

[0036] Further, the visual recognition model further comprises a convolutional neural network model, which comprises a convolutional layer, an activation layer, a pooling layer and a fully connected layer, the convolutional layer performs local feature extraction of the object to be detected, and is given to the activation layer for activation, and then to the pooling layer for dimension reduction, and then to the fully connected layer for regression processing.

[0037] Further, the convolutional layer comprises a plurality of learnable filters, each filter is small in spatial dimension (such as width and height) but extends along the full depth range of the input data, each filter is convolved along the spatial dimension of the input data, and the dot product of the filter elements and the input data at any position is calculated, with the sliding of the filter in the spatial dimension of the input data, a two-dimensional activation map is generated, which presents the response of the filter at each spatial position, and similarly, each filter will generate an activation map, and the activation maps are stacked along the depth dimension to generate an output volume.

[0038] Further, the activation layer selects a ReLU activation function, which is a nonlinear operation function, and is an additional operation after each convolution operation. Since convolution is a linear operation, the ReLU activation function is introduced to realize nonlinear operation, which does not change the size of the volume.

[0039] Further, the pooling layer reduces the network computing requirement step by step after the use of the convolutional layer and the activation layer, and maximally reduces the possibility of overfitting, and adopts maximum pooling (such as maximum value) and average pooling (such as average value) operation, first defines a spatial neighborhood window, and then takes the maximum element and the average element of the corrected feature map in the window for output; it can realize obtaining an image representation that is almost scale invariant, and can also detect the object to be detected at any position in the image.

[0040] Further, the fully connected layer performs full integration calculation and output of the convolutional layer, the activation layer and the pooling layer.

[0041] Further, a method for operating a light detection system for an object surface, the method comprising the following steps: Step 1: Setting up the detection system Select a rack, a numerical controller, a transmission table, a carrier table, a collector and a computer, the numerical controller, the transmission table, the carrier table and the collector are installed on the rack, the computer is placed on the rack, the carrier table is arranged on the transmission table, the carrier table is used to place the object to be detected, the transmission table is used to transmit the carrier table, the collector is arranged above the transmission table, the numerical controller is electrically connected to the collector, the transmission table and the computer, the numerical controller transmits the signal collected by the collector to the computer, and the numerical controller is also used to control the operation of the transmission table; The transmission table includes a transmission motor, a driving wheel, an auxiliary driving wheel and a transmission belt, the transmission motor is installed on the rack, the driving wheel and the auxiliary driving wheel are rotatably installed on the rack, the driving end of the transmission motor is connected to the driven end of the driving wheel, the transmission belt is installed on the driving wheel and the auxiliary driving wheel, and the control end of the transmission motor is electrically connected to the control end of the numerical controller; the transmission motor can realize forward and reverse rotation; the transmission belt is selected as a chain type or a belt type; The carrier table is a planar carrier table, a plurality of limiting columns are arranged on the carrier table, and the limiting columns are arranged in a rectangular shape, and the object to be detected can be stably placed in the limiting columns; Step 2: Charging the collector Based on step 1, select a light source, a reflector, a display assembly, a lens, an image lens and a collector shell, the light source, the reflector, the display assembly, the lens and the image lens are sequentially installed in the collector shell, the light source illuminates the object to be detected on the carrier table through the reflector, the display assembly, the lens and the image lens, the reflected light from the object to be detected on the carrier table is collected by the image lens, and the collected signal is transmitted to the computer through the numerical controller for image processing; the reflector is selected as a uniform light reflector; the display assembly includes a heat shield and an LCD display device, the heat shield is arranged between the reflector and the LCD display device, and the LCD display device is electrically connected to the numerical controller; the LCD display device is selected as an LCD display device compatible with full light transmission mode and mask light transmission mode; the lens is selected as a cemented lens; Step 3: Setting up the acquisition signal processing system Based on step 2, select a computer, and the computer is loaded with an acquisition signal processing system, the acquisition signal processing system is provided with a spatio-temporal graph convolution network model, the spatio-temporal graph convolution network model can collect the graph structure of the object to be detected through the collector, and model the time distribution and spatial distribution of the object to be detected based on the graph structure; the acquisition signal processing system is also provided with a visual recognition model, which collects the object to be detected through the collector and classifies and classifies the object to be detected; The visual recognition model comprises a Haar cascade model, which uses matrix features and pixel value differences of matrix regions to display the local features of the segmentation image of the object to be detected through the image of the object to be detected collected by the collector. The visual recognition model further comprises an AdaBoost model, which combines the rectangular features of each Haar cascade model with a threshold value to form a weak classifier, and then focuses on error samples through multiple iterations to finally integrate the weak classifiers into a strong classifier. The visual recognition model further comprises a convolutional neural network model, which comprises a convolutional layer, an activation layer, a pooling layer, and a fully connected layer. The convolutional layer extracts the local features of the object to be detected, and the activation layer is activated, and then the pooling layer is dimensionally reduced. The process is repeated, and finally the fully connected layer is used for regression processing. The depth learning model is built on the signal processing system, which can cooperate with the Haar cascade model and the AdaBoost model to realize image processing. Step four: detection of the object to be detected Based on step three, the object to be detected is placed in the limiting column of the object table, and the numerical controller controls the transmission motor to operate. The transmission table will send the object to be detected to the detection area below the collector. The numerical controller starts the collector to operate, and the light source emits white light through the reflector to the heat insulation sheet. The heat insulation sheet filters out the infrared light, and the remaining light is projected onto the LCD display device, which can be fully transmitted or mask projected, and then through the lens to the object to be detected. The object to be detected is illuminated. The object to be detected reflects light to the image collector, which converts the collected signal into an electrical signal and transmits it to the computer through the numerical controller. The computer converts the electrical signal into image data signal through the signal processing system and presents it. The Haar cascade model combined with the AdaBoost model is first initialized, then the collection data of the collector is read, and the image is processed. The RGB image of the object to be detected is converted into a grayscale image, which will present different gray value regions. Then the Haar cascade classifier and the AdaBoost model are processed. The Haar cascade classifier divides the image into several matrix windows, each of which includes several pixels. Then each matrix window is processed and a threshold result is generated. The threshold value can reflect the difference value of light and dark areas. This process also integrates the spatio-temporal graph convolution network model, the convolutional neural network model, and the depth learning model. The feature vector processed by the Haar cascade classifier and the AdaBoost model and the flattened feature vector of the convolution layer, the activation layer and the pooling layer of the convolution neural network model are fused, the feature vector refers to a string of numbers with information; then the output of the image data is completed in the full connection layer; The space-time graph convolution model is used, the graph structure of the object to be detected collected by the collector is used to model the time distribution and spatial distribution of the object to be detected based on the graph structure, the time distribution can select the state of the object to be detected changing with time, and the time distribution can be constructed for the change over time; the spatial distribution can select the spatial state of the object to be detected at a certain time, and the spatial distribution of all objects in the collector field of view at a certain time point can be constructed by the spatial distribution; Step five: recycling of the object to be detected Based on step four, after the object to be detected is detected, the numerical controller controls the transmission motor to operate, the transmission table carries the object table to the recycling area, and the external six-axis manipulator takes and places the object to be detected, if it is a qualified product, it is placed in the qualified box, if it is an unqualified product, it is placed in the unqualified box, then the transmission motor is reversed, the object table is carried to the feeding area, the next object to be detected is loaded, and the detection of all objects to be detected is completed. Beneficial effects

[0042] The light detection technology of the pin solder is reasonably designed, a light detection system for the surface of an object and a detection method thereof are provided, a rack, a numerical controller, a transmission table, an object table, a collector and a computer are used to solve the problems of poor brightness uniformity, low energy utilization, low brightness, regional light interference and the like of the existing light projection system; the computer can be equipped with a signal acquisition and processing system, and a space-time graph convolution network model and a visual recognition model are used to segment, classify and convolve the time and space of the object to be detected (such as pin solder), so as to better detect the object to be detected, improve the detection effect of the object to be detected, and better adapt to the detection needs in production. The automatic operation of the numerical controller saves time and labor, and can well control the cooperative operation of the transmission table, the collector and the computer, so that the detection is more convenient; the signal acquisition and processing system on the computer can better present the detection image and detection data, and better detect the qualified products and unqualified products; the mutual cooperation of the reflector, the display assembly, the lens and the lens can make the light of the light source more evenly irradiate on the object to be detected, and better collect the image of the object to be detected. BRIEF DESCRIPTION OF DRAWINGS

[0043] Fig. 1 It is a structure diagram of the light detection system for the surface of an object.

[0044] Fig. 2 It is a structure diagram of a collector of a light detection system for object surface of the application.

[0045] Fig. 3 It is a local explosion structure diagram of a collector of a light detection system for object surface of the application.

[0046] The figure mark: 01, computer; 02, numerical controller; 03, rack; 04, transmission table; 05, object table; 06, collector; 21, main shell; 22, lens shell; 16, light source device; 15, reflector; 14, heat insulation sheet; 13, LCD display device; 12, cemented lens; 11, image lens. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the application will be apparently and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work belong to the protection scope of the application.

[0048] Reference Figs. 1-3 As shown in the figure, the application provides a light detection system for object surface, which comprises a rack 03, a numerical controller 02, a transmission table 04, an object table 05, a collector 06 and a computer 01. The rack 03 is made of metal, and the numerical controller 02, the transmission table 04, the object table 05 and the collector 06 are installed on the rack 03. The numerical controller 02 is selected to have a touch display screen, and can control the start and stop of the transmission table 04 and the speed operation, and can also control the operation of the collector 06. The computer 01 is placed on the rack 03, preferably on one side of the rack 03, for convenient operation of the operator. The object table 05 is arranged on the transmission table 04, and is used for placing the object to be detected. The transmission table 04 is used for transmitting the object table 05, so as to transmit the object to be detected into a detection area. The collector 06 is arranged above the transmission table 04 and in the detection area, and can collect information of the object to be detected in the detection area. The numerical controller 02 is electrically connected to the collector 06, the transmission table 04 and the computer 01, and transmits the signal collected by the collector 06 to the computer 01. The numerical controller 02 is also used for controlling the operation of the transmission table 04. The numerical controller 02 is preferably an embedded industrial computer. The transmission station 04 comprises a transmission motor, a driving wheel, an auxiliary driving wheel and a transmission belt. The transmission motor is installed on the rack 03. The driving wheel and the auxiliary driving wheel are rotatably installed on the rack 03. The driving end of the transmission motor is connected to the driven end of the driving wheel. The transmission belt is installed on the driving wheel and the auxiliary driving wheel. The control end of the transmission motor is electrically connected to the control end of the numerical controller 02. The transmission motor is a servo motor. The transmission motor can realize forward and reverse rotation. The transmission belt is a chain type or a belt type. For the chain type transmission belt, the driving wheel and the auxiliary driving wheel are provided with chain gears matched with the chain type transmission belt, thereby driving the operation of the object table. For the belt type transmission belt, the driving wheel and the auxiliary driving wheel are provided with rough surfaces matched with the belt type transmission belt, thereby driving the operation of the transmission belt. The object table 05 is a planar object table. A plurality of limiting columns are arranged on the object table 05 in a rectangular distribution. The limiting columns can stably place the object to be detected (such as a PCB). The collector 06 comprises a light source 16, a reflector 15, a display assembly, a lens, an image lens 11 and a shell. The light source 16, the reflector 15, the display assembly, the lens and the image lens 11 are sequentially installed in the shell. The shell comprises a main shell 21 and a lens shell 22. The lens shell 22 is installed on the main shell 21. The light source 16, the reflector 15, the display assembly and the lens are sequentially installed in the main shell 21. The image lens 11 is installed in the lens shell 22. The light source 16 is a matrix distributed white light LED light source. A plurality of white light LED lamp beads are connected in series or parallel on a lamp panel. The LED lamp beads are arranged in a matrix. The light source 16 emits light which is reflected by the reflector 15, the display assembly, the lens and the image lens 11 to illuminate the object to be detected on the object table 05. The reflected light of the object to be detected on the object table 05 is collected by the image lens 11. The collected signal is transmitted to the computer 01 through the numerical controller 02 for image processing. The reflector 15 is a uniform light reflector. The reflector 15 comprises a table-shaped reflector and a conical reflector. The conical reflector is arranged in the table-shaped reflector. The small size end of the table-shaped reflector is close to the light source 16. The large size end of the table-shaped reflector is close to the display assembly. The light source light can be reflected multiple times in the reflector 15. The reflector 15 functions as a light homogenizer to obtain uniform light intensity distribution of the display assembly. The conical reflector is selected as six circular conical reflectors. The table-shaped reflector is selected as an isosceles circular table-shaped reflector. The bottom of the circular conical reflector faces the small size end of the table-shaped reflector. The top of the circular conical reflector faces the large size end of the table-shaped reflector. The circular conical reflector is gradually increased in size with the isosceles circular table-shaped reflector. The light of the light source is uniformly reflected by the reflector and the illumination area is expanded. The display assembly comprises the heat insulation sheet 14 and the LCD display device 13, the heat insulation sheet 14 is arranged between the reflector 15 and the LCD display device 13, and the LCD display device 13 is electrically connected to the numerical controller 02; the LCD display device 13 is selected from an LCD display device compatible with full light transmission mode and mask light transmission mode, that is, if the LCD display device in full light transmission mode is required, the numerical controller controls the LCD display device to arrange the liquid crystal molecules to allow the light from the reflector to pass through without obstruction; if the LCD display device in mask light transmission mode is required, the numerical controller controls the LCD display device to arrange the liquid crystal molecules to form selective light shielding and selective light transmission, thereby forming a mask-like pattern, and the light from the reflector partially transmits through the LCD display device, thereby forming a projection on the object table; the LCD display device in the two modes is very helpful for the object to be detected (such as a PCB); the LCD display device 13 can make the light from the light source on the LCD display device 13 have uniform illumination distribution or projection distribution; the heat insulation sheet 14 is used to reduce the heat of the LCD display device 13, so as to ensure that the LCD display device 13 can work normally; the heat insulation sheet 14 is selected from a glass sheet coated with an infrared absorption film or a plastic sheet coated with an infrared absorption film, which can also be called a heat insulation filter or an infrared filter; since the white light of the light source contains heat-emitting infrared light, if this heat-emitting light irradiates the LCD display device for a long time, the temperature of the LCD display device will increase, which will affect the normal use effect of the LCD display device; in particular, the surface temperature of the LCD display device 13 is too high, which is easy to cause blackening or even damage; therefore, the heat insulation sheet 14 is used to insulate the LCD display device 13; the heat insulation sheet 14 is selected from a glass sheet coated with 10 μm thick antimony tin oxide on both sides; the antimony tin oxide is deposited on the glass sheet by magnetron sputtering technology; The lens is selected from a cemented lens 12; the cemented lens 12 can change the light field distribution, improve the brightness uniformity of light, make more light be projected into the object table 05 far from the optical axis, thereby improving the brightness uniformity of the object table 05; in addition, the cemented lens 12 can also deflect off-axis light, so that more off-axis light enters the lens, thereby improving the brightness and uniformity of the projected light; the cemented lens 12 comprises a convex lens and a concave lens, the convex lens is a symmetrical convex lens, and the convex surface of the convex lens and the concave surface of the concave lens are matched and bonded by glue; the concave lens is selected from a concave lens with a concave surface on one side and a flat surface on the other side; The image lens 11 includes a lens and an image collector, both of which are installed in the shell, and the image collector is electrically connected to the digital controller 02; the lens projects the picture on the LCD display device 13 (including a picture and no picture) onto the stage 05 (projection surface); the lens includes a plurality of different optical lenses, which can project the picture on the LCD display device 13 on the stage 05 with a certain magnification; when the angle between the optical axis of the lens and the normal of the projection surface is β, and the angle between the normal of the LCD display device 13 and the optical axis of the lens is α, α and β satisfy the Snell law of the optical path structure, and a comprehensive and clear projection picture can be obtained; the image collector selects a CCD sensor type image collector or a CMOS sensor type image collector.

[0049] The computer 01 is provided with a signal acquisition and processing system, and a spatiotemporal graph convolution network model is built on the signal acquisition and processing system; the spatiotemporal graph convolution network model can collect the graph structure of the object to be detected by the collector 06, and model the time distribution and spatial distribution of the object to be detected based on the graph structure; the time distribution can select the change state of the object to be detected over time, such as the pin soldering on the components of a printed circuit board (PCB); the soldering that meets the standard is first a bright metal surface on the pin, and with the passage of time, the surface will be oxidized, lose luster, and become dull; the soldering that does not meet the standard will be presented by a bright metal surface at the beginning, and with the passage of time, such soldering will absorb moisture or dust in the air, possibly produce a chemical reaction, and cause the color of the soldering to become darker (such as yellowing, brown, etc.); for such time-dependent changes, the time distribution can be constructed; the spatial distribution can select the spatial state of the object to be detected at a certain time, such as the pin soldering on the components of a printed circuit board (PCB) and the printed circuit board and electronic components, etc.; through the spatial distribution, the spatial distribution of all objects (including printed circuit boards, electronic components, soldering, etc.) at a certain time point and within the field of view of the collector can be constructed.

[0050] The signal acquisition and processing system is also provided with a visual recognition model, which collects the object to be detected (such as a PCB) by the collector and segments and classifies it; for example, classifying by electronic component category, segmenting by electronic component position, segmenting by electronic component pin, classifying and segmenting the soldering of the electronic component pin, etc. The visual recognition model includes a Haar cascade model, which uses rectangular features and the pixel value difference of the rectangular region to display the local features (such as target objects, positions, etc.) of the segmented image of the object to be detected by using the image of the object to be detected collected by the collector. In order to improve the construction accuracy of the Haar cascade model, the visual recognition model further comprises an AdaBoost model, which combines the rectangular features of each Haar cascade model with a threshold value to form a weak classifier, and then through multiple rounds of iteration, focuses on the error samples, and finally weights the weak classifiers to integrate a strong classifier; that is, the AdaBoost model integrates the weak classifier of the local features of the image segmented by the Haar cascade model, and further constructs a strong classifier of the object to be detected, so as to ensure that the Haar cascade model and the spatiotemporal graph convolution network model are more accurate.

[0051] In the AdaBoost model, a single matrix feature capable of separating positive examples and negative examples is selected, and for each single matrix feature, a weak classifier determines a threshold classification function to reduce the number of misclassified samples.

[0052] In the AdaBoost model, Hj(x) is composed of a feature Fj, a threshold value θj and a polarity Pj, wherein, Hj(x)={1, if PjFj(x) is less than Pjθj} or Hj(x)={0, if PjFj(x) is not less than Pjθj}, wherein x is an A×B pixel sub-window, and A and B are the length and width of the pixel sub-window, respectively, and A and B are selected to be equal or unequal.

[0053] The Haar cascade model is combined with an AdaBoost model, first initializing the system, then reading the acquisition data of the collector, and performing image processing. The image processing mainly converts the RGB image of the object to be detected (such as a PCB board) into a gray-scale image, which will present different gray-scale value areas, and then performs Haar cascade classifier and AdaBoost model processing. The Haar cascade classifier divides the image into a plurality of matrix windows (sub-regions), each matrix window including a plurality of pixels, and then processes each matrix window and generates a threshold result. The threshold value can reflect the difference value of light and dark areas. This is combined with a sample comparison model to process defects. If there is no defect, the qualified product detection result is output and stored in the memory as a standard detection piece. In future detection, if there is no defect, the qualified product detection result is output and continues to be filled into the standard detection piece memory, so that the standard detection piece memory library is more accurate. If the subsequent detection has defects and differs from the standard detection piece in the standard detection piece memory library, the unqualified product result is output. The Haar cascade model here is completed in cooperation with the spatiotemporal graph convolution network model. The target value of the defect is, for example, whether the electronic component pin soldering on the PCB board exists and whether the size of the existing area deviates from the qualified standard. If the target value does not exist, the defect result is directly recorded. If the target value exists and deviates from the qualified standard, the object to be detected is identified, then the defect classification of the object to be detected (mainly involving single-sided defects, double-sided defects, high and low degree defects, etc.) is performed, and finally the defect result is recorded.

[0054] The visual recognition model further includes a convolutional neural network model, which includes a convolutional layer, an activation layer, a pooling layer, and a fully connected layer. The convolutional layer extracts local features of the object to be detected and activates the activation layer. Then, the pooling layer is reduced in dimension, and the cycle is repeated. Finally, the fully connected layer is processed by regression.

[0055] The convolutional layer includes a plurality of learnable filters. Each filter is small in size in the spatial dimension (such as width and height), but extends along the full depth range of the input data. Each filter is convolved along the spatial dimension of the input data, and the dot product of the filter elements and the input data at any position is calculated. As the filter slides in the spatial dimension of the input data, an activation map is generated, showing the response of the filter at each spatial position. Each filter generates an activation map, and the activation maps are stacked along the depth dimension to generate an output volume.

[0056] Wherein, the activation layer selects the ReLU activation function, the ReLU activation function is a nonlinear operation function, is the additional operation after each convolution operation, since the convolution is a linear operation, the nonlinear operation is realized by introducing the ReLU activation function, which also does not change the size of the volume.

[0057] Wherein, the pooling layer is used after the convolution layer and the activation layer, gradually reduces the network computing requirement, and maximally reduces the possibility of overfitting, adopts maximum pooling (such as maximum value) and average pooling (such as average value) operation, first defines a spatial neighborhood window, and then takes the maximum element and average element of the corrected feature map in the window for output; can realize obtaining nearly scale invariant image representation, and can detect the object to be detected at any position in the image.

[0058] Wherein, the full connection layer integrates and outputs the convolution layer, the activation layer and the pooling layer.

[0059] Wherein, a control method for a light detection system for an object surface, the control method steps are as follows: Step one: building of the detection system Select the rack, numerical controller, transmission table, object table, collector and computer, the rack is installed with the numerical controller, transmission table, object table and collector, the computer is placed on the rack, the object table is arranged on the transmission table, the object table is used for placing the object to be detected, the transmission table is used for transmitting the object table, the collector is arranged above the transmission table, the numerical controller is electrically connected to the collector, the transmission table and the computer, the numerical controller transmits the signal collected by the collector to the computer, and the numerical controller is also used for controlling the operation of the transmission table; Wherein, the transmission table comprises a transmission motor, a driving wheel, an auxiliary driving wheel and a transmission belt, the transmission motor is installed on the rack, the driving wheel and the auxiliary driving wheel are rotatably installed on the rack, the driving end of the transmission motor is connected to the driven end of the driving wheel, the transmission belt is installed on the driving wheel and the auxiliary driving wheel, and the control end of the transmission motor is electrically connected to the control end of the numerical controller; the transmission motor can realize forward and reverse rotation; the transmission belt selects a chain type or a belt type transmission belt; Wherein, the object table is a planar object table, a plurality of limiting columns are arranged on the object table, and the limiting columns are arranged in a rectangular shape, and the object to be detected can be stably placed in the limiting columns; Step two: supplement of the collector Based on step one, the light source device, reflector, display assembly, lens, image lens and collector of the shell are selected, the light source device, reflector, display assembly, lens and image lens are installed in the shell in turn, the light source device emits light which reaches the stage through the reflector, display assembly, lens and image lens to illuminate the object to be detected, the reflected light of the object to be detected on the stage is collected by the image lens, and the collected signal is transmitted to the computer through the numerical controller for image processing; the reflector is selected as a uniform light reflector; the display assembly includes a heat shield and an LCD display device, the heat shield is arranged between the reflector and the LCD display device, and the LCD display device is electrically connected to the numerical controller; the LCD display device selects a LCD display device compatible with full light transmission mode and mask light transmission mode; the lens selects a cemented lens; Step three: building of the signal acquisition and processing system Based on step two, a computer is selected, and the computer is equipped with a signal acquisition and processing system, the signal acquisition and processing system is built with a spatio-temporal graph convolution network model, the spatio-temporal graph convolution network model can collect the graph structure of the object to be detected through the collector, and model the time distribution and spatial distribution of the object to be detected based on the graph structure; the signal acquisition and processing system is also equipped with a visual recognition model, which collects the object to be detected through the collector and classifies and classifies it; The visual recognition model includes a Haar cascade model, which uses matrix features and pixel value differences in matrix regions through a cascade classifier to display the local features of the segmented image of the object to be detected based on the image of the object to be detected collected by the collector; The visual recognition model also includes an AdaBoost model, which combines the rectangular features of each Haar cascade model with a threshold to form a weak classifier, and then focuses on the samples with errors through multiple iterations, and finally weights the weak classifiers into a strong classifier; The visual recognition model also includes a convolutional neural network model, which includes a convolutional layer, an activation layer, a pooling layer and a fully connected layer, the convolutional layer extracts local features of the object to be detected and gives the activation layer for activation, then to the pooling layer for dimensionality reduction, and then to the fully connected layer for regression processing; The signal acquisition and processing system is built with a deep learning model, which can cooperate with the Haar cascade model and the AdaBoost model to realize image processing; Step four: detection of the object to be detected Based on step three, the object to be detected is placed in the limiting column of the object table, the numerical controller controls the transmission motor to operate, the transmission table will send the object to be detected to the detection area below the collector, the numerical controller starts the operation of the collector, the light source will emit white light through the reflector to the heat insulation sheet, the heat insulation sheet filters out the infrared light, and the remaining light is irradiated to the LCD display device, which can be fully transmitted or mask projected, and then passed through the lens to the object to be detected, illuminating the object to be detected; The object to be detected will reflect light to the image collector, which will convert the collected signal into an electrical signal and transmit it to the computer through the numerical controller; Wherein, the computer will convert the electrical signal into image data signal through the acquisition signal processing system and present it; Wherein, through the Haar cascade model combined with the AdaBoost model, first initialize the system, then read the acquisition data of the collector, and process the image, convert the RGB image of the object to be detected into a gray image, which will present different gray value areas, then process the Haar cascade classifier and AdaBoost model, the Haar cascade classifier divides the image into several matrix windows (subareas), each matrix window includes several pixels, then each matrix window is processed and a threshold result is generated, the threshold can reflect the difference value of light and dark area, and then the sample comparison model is combined to process defects; In this process, the space-time graph convolution network model, the convolution neural network model and the deep learning model are also integrated; Wherein, the feature vectors of Haar cascade classifier and AdaBoost model processing and the feature vectors of convolution layer, activation layer and pooling layer of convolution neural network model are fused, the feature vector refers to a string of numbers with information; Then go to the full connection layer to complete the output of image data; Wherein, the spatio-temporal graph convolution model is used to collect the graph structure of the object to be detected by the collector, and the time distribution and the space distribution of the object to be detected are modeled based on the graph structure. The time distribution can be selected as the change state of the object to be detected over time, such as the pin soldering on the components of the PCB board. The soldering that meets the standard is first a bright metal surface on the pin, and with the passage of time, its surface will be oxidized, lose luster, and become dull. The soldering that does not meet the standard will be presented by a bright metal surface at first due to pollution, impurities invasion or flux residue, and with the passage of time, such soldering will absorb moisture or dust in the air, possibly produce a chemical reaction, causing the color of the soldering to become darker (such as yellowing, brown, etc.). For such time-dependent changes, the time distribution can be constructed. The space distribution can be selected as the spatial state of the object to be detected at a certain time point, such as the pin soldering on the components of the PCB board and the circuit board and electronic components, etc. Through the space distribution, the spatial distribution of all objects (including circuit boards, electronic components, soldering, etc.) at a certain time point and within the collector's field of view can be constructed. Step five: recycling of the object to be detected Based on step four, after the detection of the object to be detected, the numerical controller controls the transmission motor to operate, and the transmission table will transport the object table to the recycling area, and the external six-axis manipulator will take and place the object to be detected. If it is a qualified product, it will be placed in the qualified box, and if it is an unqualified product, it will be placed in the unqualified box. Then the transmission motor is reversed to transport the object table to the feeding area to load the next object to be detected, and the cycle is repeated to complete the detection of all objects to be detected.

[0060] Specifically, if a soldered PCB with multiple electronic components, the solder on the pins of the electronic components on it meets the soldering standard, it is placed on the stage and sent to the collector below through the transmission table. The LCD display device of the collector can be in full transmission mode, and the light source device can evenly illuminate the PCB. Then the image collector collects the image of the PCB and transmits it to the computer for processing and presentation. The processing system detects and locates the image through the Haar cascade model and the AdaBoost model combined with the deep learning model. A rectangular window is used for processing to locate the target object position with classification (such as resistance, inductance, chip, capacitor, diode, etc. and corresponding pins). A boundary box coordinate list containing the target object position is generated, and corresponding feature extraction and quantitative analysis are performed. Spatial relationship analysis and single-point quality analysis are performed through the spatio-temporal graph convolution model and the convolutional neural network model to form a standard soldering pattern and data, which are stored in the standard library as a detection reference. If a soldered PCB with multiple electronic components, the solder on the pins of the electronic components on it meets the soldering standard, it is placed on the stage and sent to the collector below through the transmission table. The LCD display device of the collector can be in mask mode, and the light source device can evenly illuminate the PCB in mask mode. Then the image collector collects the image of the PCB and transmits it to the computer for processing and presentation. With the existence of the mask mode, the influence of some irrelevant objects can be avoided. The processing system detects and locates the image through the Haar cascade model and the AdaBoost model combined with the deep learning model. A rectangular window is used for processing to locate the target object position with classification (such as resistance, inductance, chip, capacitor, diode, etc. and corresponding pins). A boundary box coordinate list containing the target object position is generated, and corresponding feature extraction and quantitative analysis are performed. Spatial relationship analysis and single-point quality analysis are performed through the spatio-temporal graph convolution model and the convolutional neural network model to form a standard soldering pattern and data, which are stored in the standard library as a detection reference.If the soldering of the electronic component pins on a soldered PCB with multiple electronic components is unclear whether it meets the soldering standard, it is placed on the object table and sent to the collector below through the transmission table. The LCD display device of the collector can be in full transmission mode or mask mode, and the light of the light source device is evenly irradiated on the PCB. Then the image collector collects the image of the PCB and transmits it to the computer for processing and presentation. The processing system detects and locates the image through the Haar cascade model and AdaBoost model combined with the deep learning model. Rectangular windows are used for processing to locate the positions of the classified targets (such as resistors, inductors, chips, capacitors, diodes, etc. and corresponding pins), generate a list of boundary box coordinates containing target positions, and perform corresponding feature extraction and quantitative analysis. Spatial relationship analysis and single-point quality analysis are performed through the spatio-temporal graph convolution model and the convolutional neural network model. If it meets the standard soldering pattern and data, it is stored in the standard library as a detection reference. If it does not meet the standard soldering pattern and data, it is stored in the defect library as a defective product for subsequent defect state statistics and process improvement. If a soldered PCB with multiple electronic components is placed for a period of time, the soldering of the electronic component pins on the PCB is unclear whether it meets the soldering standard. It is placed on the object table and sent to the collector below through the transmission table. The LCD display device of the collector can be in full transmission mode or mask mode, and the light of the light source device is evenly irradiated on the PCB. Then the image collector collects the image of the PCB and transmits it to the computer for processing and presentation. The processing system detects and locates the image through the Haar cascade model and AdaBoost model combined with the deep learning model. Rectangular windows are used for processing to locate the positions of the classified targets (such as resistors, inductors, chips, capacitors, diodes, etc. and corresponding pins), generate a list of boundary box coordinates containing target positions, and perform corresponding feature extraction and quantitative analysis. Spatial relationship analysis and single-point quality analysis are performed through the spatio-temporal graph convolution model and the convolutional neural network model. If it meets the standard soldering pattern and data, it is stored in the standard library as a detection reference. If it does not meet the standard soldering pattern and data, it is stored in the defect library as a defective product for subsequent defect state statistics and process improvement. Here, there may be a situation where the solder is placed for a period of time. If the soldering has defects, impurities will be introduced, and the color of the solder will change. By processing the time distribution of the spatio-temporal graph convolution network model, the defective product (i.e. defective product) is detected and stored in the defect library for subsequent defect state statistics and process improvement.

[0061] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the claims and their equivalents.

Claims

1. A light detection system for a surface of an object, characterized in that The light detection system comprises a rack (03), a numerical controller (02), a conveying table (04), a carrier table (05), a collector (06) and a computer (01), the numerical controller (02), the conveying table (04), the carrier table (05) and the collector (06) are installed on the rack (03), the computer (01) is placed on the rack (03), the carrier table (05) is arranged on the conveying table (04), the carrier table (05) is used for placing an object to be detected, the conveying table (04) is used for conveying the carrier table (05), the collector (06) is arranged above the conveying table (04), the numerical controller (02) is electrically connected to the collector (06), the conveying table (04) and the computer (01), the numerical controller (02) transmits a signal collected by the collector (06) to the computer (01), and the numerical controller (02) is also used for controlling operation of the conveying table (04). The collector (06) comprises a light source device (16), a reflector (15), a display assembly, a lens, an image lens (11) and a device shell, the light source device (16), the reflector (15), the display assembly, the lens and the image lens (11) are sequentially installed in the device shell, the light source device (16) illuminates the object to be detected on the carrier table (05) through the reflector (15), the display assembly, the lens and the image lens (11), the light reflected by the object to be detected on the carrier table (05) is collected by the image lens (11), and the collected signal is transmitted to the computer (01) through the numerical controller (02) for image processing. The lens is a cemented lens (12). The display assembly comprises a heat insulation sheet (14) and an LCD display device (13), the heat insulation sheet (14) is arranged between the reflector (15) and the LCD display device (13), and the LCD display device (13) is electrically connected to the numerical controller (02). The computer (01) is provided with a collection signal processing system, the collection signal processing system is provided with a space-time graph convolution network model, the space-time graph convolution network model can collect a graph structure of the object to be detected by the collector (06), and the time distribution and the space distribution of the object to be detected are modeled based on the graph structure.

2. A light detection system for a surface of an object according to claim 1, characterized in that The heat insulation sheet (14) is a glass sheet coated with an infrared absorption film layer or a plastic sheet coated with an infrared absorption film layer.

3. A light detection system for a surface of an object according to claim 1, wherein, The LCD display device (13) is an LCD display device compatible with full light transmission mode and mask light transmission mode.

4. A light detection system for a surface of an object according to claim 1, wherein, The conveying table (04) comprises a conveying motor, a driving wheel, a driven wheel and a conveying belt, the conveying motor is installed on the rack (03), the driving wheel and the driven wheel are rotatably installed on the rack (03), a driven end of the driving wheel is connected to a driving end of the conveying motor, the conveying belt is installed on the driving wheel and the driven wheel, and a control end of the conveying motor is electrically connected to a control end of the numerical controller (02).

5. A light detection system for a surface of an object according to claim 1, wherein, The reflector (15) is a uniform light reflector.

6. A light detection system for a surface of an object according to claim 1, wherein, The image lens (11) comprises a lens and an image collector, both of which are installed in the casing, and the image collector is electrically connected to the numerical controller (02).

7. A light detection system for a surface of an object according to claim 1, wherein, The signal acquisition and processing system is also provided with a visual recognition model, which acquires the object to be detected by the collector (06) and performs segmentation and classification on the object.

8. A light detection system for a surface of an object according to claim 7, wherein, The visual recognition model comprises a Haar cascade model, which uses matrix features and pixel value differences of matrix regions to display the local features of the segmented image of the object to be detected by using the image of the object to be detected acquired by the collector (06).

9. A light detection system for a surface of an object according to claim 8, wherein, The visual recognition model also comprises an AdaBoost model, which combines the rectangular features of each Haar cascade model with a threshold to form a weak classifier, and then focuses on the error samples through multiple iterations, and finally weights the weak classifiers to form a strong classifier.

10. A method of handling a light detection system for a surface of an object, characterized in that The control method comprises the following steps: Step one: construction of the detection system Select the rack (03), numerical controller (02), transmission table (04), object table (05), collector (06) and computer (01), the rack (03) is provided with the numerical controller (02), transmission table (04), object table (05) and collector (06), the computer (01) is placed on the rack (03), the object table (05) is arranged on the transmission table (04), the object table (05) is used for placing the object to be detected, the transmission table (04) is used for transmitting the object table (05), the collector (06) is arranged above the transmission table (04), the numerical controller (02) is electrically connected to the collector (06), transmission table (04) and computer (01), the numerical controller (02) transmits the signal collected by the collector (06) to the computer (01), and the numerical controller (02) is also used for controlling the operation of the transmission table (04); The transmission table (04) comprises a transmission motor, a driving wheel, an auxiliary driving wheel and a transmission belt, the transmission motor is installed on the rack (03), the driving wheel and the auxiliary driving wheel are rotatably installed on the rack (03), the driving end of the transmission motor is connected to the driven end of the driving wheel, the transmission belt is installed on the driving wheel and the auxiliary driving wheel, and the control end of the transmission motor is electrically connected to the control end of the numerical controller (02); the transmission motor can realize forward and reverse rotation; the transmission belt is selected as a chain type or a belt type transmission belt; The object table (05) is a planar object table, which is provided with a plurality of limiting columns, and the limiting columns are arranged in a rectangular shape, and the object to be detected can be stably placed in the limiting columns; Step two: supplement of the collector Based on step one, the light source device (16), reflector (15), display component, lens, image lens (11) and collector (06) of the shell are selected, the light source device (16), reflector (15), display component, lens and image lens (11) are installed in the shell in turn, the light source device (16) emits light to the reflector (15), display component, lens and image lens (11) to illuminate the object to be detected on the object table (05), and the reflected light of the object to be detected on the object table (05) is collected by the image lens (11), and the collected signal is transmitted to the computer (01) through the numerical controller (02) for image processing; the transmitter selects a uniform reflector (15); the display component includes a heat shield (14) and an LCD display device (13), the heat shield (14) is arranged between the reflector (15) and the LCD display device (13), and the LCD display device (13) is electrically connected with the numerical controller (02); the LCD display device (13) selects an LCD display device compatible with full light transmission mode and mask light transmission mode; the lens selects a cemented lens (12); Step three: building of signal acquisition and processing system Based on step two, a computer (01) is selected, and a signal acquisition and processing system is mounted on the computer (01), and a spatiotemporal graph convolution network model is built on the signal acquisition and processing system, the spatiotemporal graph convolution network model can collect the graph structure of the object to be detected through the collector (06), and model the time distribution and spatial distribution of the object to be detected based on the graph structure; a visual recognition model is also mounted on the signal acquisition and processing system, and the visual recognition model collects the object to be detected through the collector (06) and segments and classifies it; The visual recognition model includes a Haar cascade model, which uses matrix features and pixel value differences in matrix regions through a cascade classifier to display the local features of the segmented image of the object to be detected based on the image of the object to be detected collected by the collector (06); The visual recognition model also includes an AdaBoost model, which combines the rectangular features of each Haar cascade model with a threshold to form a weak classifier, and then focuses on error samples through multiple iterations to finally weight the weak classifiers into a strong classifier; The visual recognition model also includes a convolutional neural network model, which includes a convolutional layer, an activation layer, a pooling layer and a fully connected layer, the convolutional layer extracts local features of the object to be detected, and the activation layer is activated, and then the pooling layer is pooled and dimensionally reduced, and then the fully connected layer is processed by regression; The signal acquisition and processing system is built with a deep learning model, which can cooperate with the Haar cascade model and the AdaBoost model to realize image processing; Step four: detection of object to be detected Based on step three, the object to be detected is placed in the limiting column of the object table (05), the numerical controller (02) controls the transmission motor to operate, and the transmission table (04) will send the object to be detected to the detection area below the collector (06), the numerical controller (02) starts the collector (06) to operate, the light source (16) will emit white light through the reflector (15) uniform light transmission to the heat insulation sheet (14), the heat insulation sheet (14) filters out the infrared light, and the remaining light is irradiated on the LCD display device (13), which can be fully transmitted or mask projection, and then through the lens to the object to be detected, illuminating the object to be detected; The object to be detected will reflect light to the image collector (06), which will convert the collected signal into an electrical signal and transmit it to the computer (01) through the numerical controller (02); Wherein, the computer (01) will convert the electrical signal into image data signal through the acquisition signal processing system to present; Wherein, through the Haar cascade model combined with the AdaBoost model, first initialize the system, then read the acquisition data of the collector (06), and process the image, convert the RGB image of the object to be detected into a gray image, which will present different gray value areas, then process the Haar cascade classifier and the AdaBoost model, the Haar cascade classifier divides the image into several matrix windows, each matrix window includes several pixels, then each matrix window is processed and a threshold result is generated, the threshold can reflect the difference value of light and dark area, and this is combined with sample comparison model to process defects; In this process, the space-time graph convolution network model, the convolution neural network model and the deep learning model are also integrated; Wherein, the feature vectors of Haar cascade classifier and AdaBoost model processing and the feature vectors of convolution layer, activation layer and pooling layer of convolution neural network model are fused, the feature vector refers to a string of numbers with information; Then go to the full connection layer to complete the output of image data; Wherein, the space-time graph convolution model is used, the graph structure of the object to be detected collected by the collector (06) is used to model the time distribution and spatial distribution of the object to be detected based on the graph structure, the time distribution can select the state of the object to be detected changing with time, and for this time change, the time distribution can be constructed; The spatial distribution can select the spatial state of the object to be detected at a certain time, and through the spatial distribution, the spatial distribution of all objects at a certain time point and within the field of view of the collector (06) can be constructed; Step five: recycling of the object to be detected Based on step four, after the detection of the object to be detected, the numerical controller (02) controls the transmission motor to operate, and the transmission table (04) will transport the object table (05) to the recycling area, and the external six-axis manipulator will take and place the object to be detected, if it is a qualified product, it will be placed in the qualified box, if it is a unqualified product, it will be placed in the unqualified box, then the transmission motor is reversed, the object table (05) is transported to the feeding area, the next object to be detected is loaded, and the cycle is repeated to complete the detection of all objects to be detected.

Images