A liquid crystal display defect detection model, sorting production line and method

Abstract

The present invention discloses a liquid crystal display defect detection model for constructing a system for identifying liquid crystal display defects. It also provides a liquid crystal display defect detection and sorting production line and method for automatically sorting liquid crystal displays based on the results of liquid crystal display defect identification. This invention effectively addresses the current challenges faced by automated liquid crystal display production lines, where mura defects on liquid crystal displays are difficult to detect due to their small size, low contrast, irregular shape, and random location. The method utilizes deep learning for analysis, outputs identification and sorting results, and, in combination with an automatic sorting module, separates unqualified products, effectively improving both the efficiency and accuracy of mura defect detection on liquid crystal displays.

Description

Technical Field

[0001] The present invention relates to a liquid crystal display screen defect detection model, a sorting production line and a method, and belongs to the technical field of industrial product visual recognition, mechanism design and product sorting. Background Art

[0002] In recent years, with the rapid development of electronic technology, the application areas of liquid crystal displays (LCDs) have expanded year by year, and their use can be seen everywhere in daily life. LCDs are used as display devices in LCD televisions, digital cameras, game consoles, televisions, mobile phones, computers, and smart watches. Furthermore, LCDs are not only used in daily life, but are also widely used in industries such as industry, military, and aerospace. However, LCDs are prone to defects during the manufacturing process, which can affect the user's visual experience. This is especially true for TFT-LCD screens, which are prone to various defects during the manufacturing process, among which mura defects are particularly difficult to detect. Therefore, it is of great significance to develop a model, sorting production line, and method for fast and highly accurate detection of spot mura defects in LCDs. Summary of the Invention

[0003] The present invention provides a liquid crystal display defect detection model for constructing a system for identifying liquid crystal display defects, and provides a liquid crystal display defect detection and sorting production line and method for automatically sorting liquid crystal displays based on the liquid crystal display defect identification results.

[0004] The technical solution of the present invention is: a liquid crystal display defect detection and sorting production line, including a self-centering clamping tray 2, a conveying module 3, an automatic sorting module 4, and a defect detection module 5; the self-centering clamping tray 2 is placed on the conveying module 3, and the self-centering clamping tray 2 is driven to move by the conveying module 3, and the liquid crystal display on the self-centering clamping tray 2 is detected by the defect detection module 5, and the automatic sorting module 4 is used to sort the liquid crystal display according to the recognition results of the defect detection module 5.

[0005] It also includes a dust removal module 1, which includes a dust removal roller 6, an air-blowing dust collector 7, a frame I8, and a sprocket I9; wherein, the sprocket I9 is ​​installed at both ends of the roller shaft of the dust removal roller 6; the dust removal roller 6 and the air-blowing dust collector 7 are installed on the frame I8 in a front-to-back manner and the air-blowing dust collector 7 is close to the forward end of the conveying direction of the conveying module 3; the sprocket I9 is ​​powered by the servo motor I37 of the upper synchronous belt conveyor in the conveying module 3.

[0006] The self-centering clamping tray 2 includes a tray 10 and a self-centering clamping mechanism; wherein, the self-centering clamping mechanism is installed on the tray 10, and the self-centering clamping mechanism includes a transverse retractable rack 16, a longitudinal retractable rack 17, a buckle 24, and a trigger device 25; wherein, the buckle 24 is installed at one end of the transverse retractable rack 16 and the longitudinal retractable rack 17, and the other end of the transverse retractable rack 16 and the longitudinal retractable rack 17 cooperates with the trigger device 25 to realize the clamping of the LCD screen between the buckles 24.

[0007] The trigger device 25 includes a trigger cover 18, a gear 15, a one-way ratchet mechanism, a disc-shaped coil spring 21, a fixed shaft 22, a return spring 92, and a fixed pin 95; the trigger cover 18 is mounted on one end of the fixed shaft 22 via the fixed pin 95, and a strip-shaped through hole is provided on the cylindrical body of the trigger cover 18 for cooperating with the fixed pin 95, so that the trigger cover 18 can achieve circumferential rotation and axial movement. A return spring 92 is installed between the fixed shaft 22 and the trigger cover 18, and the entire one-way ratchet mechanism is fixed on the fixed shaft 22. The gear 15 is engaged with the transverse contraction rack 16 and the longitudinal contraction rack 17. At the same time, the one-way ratchet 19 of the one-way ratchet mechanism is fixedly connected to the gear 15. When the gear 15 rotates, it drives the one-way ratchet 19 of the one-way ratchet mechanism to rotate together. One end of the rotating coil spring 21 is installed in the fixed groove of the gear 15, and the other end of the rotating coil spring 21 is fixed on the self-centering clamping mechanism; the one-way ratchet mechanism includes a one-way ratchet 19, a pawl 20, a pawl mounting frame 94, and a spring II 96; wherein, the pawl mounting frame 94 is installed in the one-way ratchet 19, one end of the pawl mounting frame 94 is splined with the fixed shaft 22 and is fastened by a fixed fixing nut 93, the pawl 20 is installed on the outer ring of the pawl mounting frame 94, and a spring II 96 is arranged between the outer ring of the pawl mounting frame 94 and the pawl 20, and the insert arranged at one end of the trigger cover 18 cooperates with the pawl 20.

[0008] The conveying module 3 includes an upper synchronous belt conveyor and a lower double-speed chain conveyor; the upper synchronous belt conveyor includes a first conveying module and a first adjusting module, the first conveying module is used to provide power-driven movement through a first power system, and the movement of the first conveying module drives the self-centering clamping pallet 2 to follow the movement; the first adjusting module is used to drive the nut block Ⅰ41 installed on the adjusting screw rod Ⅰ40 to move up and down through the adjusting screw rod Ⅰ40, and the up and down movement of the nut block Ⅰ41 drives the angle of the two first support arms 39 connected to the nut block Ⅰ41 to change, and the change in the angle of the first support arm 39 drives the first conveying module at the other end of the two first support arms 39 to move in the opposite direction with the first support arm 39, and the movement direction is perpendicular to the conveying direction of the first conveying module. ; The lower-level double-speed chain conveyor is installed at the lower part of the upper-level synchronous belt conveyor, and includes a second conveying module and a second adjusting module. The second conveying module is used to provide power-driven movement through the second power system, and the movement of the second conveying module drives the self-centering clamping tray 2 sorted out by the automatic sorting module 4 to follow the movement; the second adjusting module is used to drive the nut block Ⅱ52 installed on the adjusting screw rod Ⅱ53 to move up and down through the adjusting screw rod Ⅱ53, and the up and down movement of the nut block Ⅱ52 drives the angle of the two second support arms 51 connected to the nut block Ⅱ52 to change, and the change of the angle of the second support arms 51 drives the second conveying module at the other end of the two second support arms 51 to move in the opposite direction with the first support arm 39, and the movement direction is perpendicular to the transportation direction of the second conveying module.

[0009] The defect detection module 5 includes a photoelectric sensing stop switch, one or more groups of defect detection components, and each group of defect detection components has the same structure, including an image acquisition module and a controller. The photoelectric counter in each group of image acquisition modules is used to count the passing pallets and transmit the count number to the respective controllers. The controller drives the camera 75 to capture images of the LCD screen based on the trigger signal of the photoelectric sensing stop switch and the count number of the photoelectric counter, and performs defect detection on the LCD screen based on the captured image.

[0010] According to another aspect of the present invention, a liquid crystal display defect detection model is provided. Based on the YOLOV4-Tiny model, the output of the first CSP module in the backbone feature extraction network is added as a shallow feature map, and the original deep feature map is subjected to a convolution operation to adjust the number of channels and divided into two outputs: one output is upsampled and then feature fused with the middle feature map, and the fused result is upsampled and then fused with the newly added shallow feature map to obtain a feature map. After being processed by a Ghost module, the obtained feature map is input into the YOLO head for classification prediction and regression prediction; the other output is processed by the Ghost module and then input into the YOLO head for classification prediction and regression prediction.

[0011] According to another aspect of the present invention, a method for detecting and sorting defects in liquid crystal displays is provided, comprising: starting a visual real-time detection and sorting system for liquid crystal displays; driving a liquid crystal display placed on a self-centering clamping tray 2 toward a defect detection module 5 via a conveying module 3; performing defect detection on the liquid crystal display on the self-centering clamping tray 2 via the defect detection module 5; and sorting the liquid crystal displays according to the recognition results of the defect detection module 5 via an automatic sorting module 4.

[0012] The defect detection of the LCD screen on the self-centering clamping tray 2 is performed by the defect detection module 5, including: the photoelectric counter in the image acquisition module in the defect detection module 5 is used to count the passing trays and transmit the count number to the respective controllers, the controller drives the camera 75 to capture the image of the LCD screen according to the trigger signal of the photoelectric sensing stop switch and the count number of the photoelectric counter, and performs defect detection on the LCD screen based on the captured image.

[0013] The controller drives the camera 75 to capture the image of the LCD screen according to the trigger signal of the photoelectric sensor stop switch and the count number of the photoelectric counter, and performs defect detection on the LCD screen according to the captured image. Specifically, if the defect detection components in the defect detection module 5 are a group, for the current detection, the photoelectric counter counts to 1 and the photoelectric sensor stop switch is triggered, then the controller drives the upper synchronous belt conveyor in the transmission module 3 to stop moving, drives the camera 75 to capture the image of the LCD screen, and calls the frozen model obtained according to the LCD screen defect detection model to perform defect detection on the captured LCD screen image. When the detection is completed, the photoelectric counter Reset to zero, and perform the next detection until it ends; if there are n groups of defect detection components in the defect detection module 5, for the current detection, if the maximum increment value of the n photoelectric counters is m and the photoelectric sensing stop switch is triggered, the mth controller drives the upper synchronous belt conveyor in the transmission module 3 to stop moving, and the first m controllers respectively drive the camera 75 to collect images of the LCD screen, and call the frozen model obtained according to the LCD screen defect detection model to perform defect detection on the collected LCD screen images. When the detection is completed, all photoelectric counters are reset to zero, and the next detection is performed until it ends; wherein, n photoelectric counters correspond to n controllers one by one.

[0014] The beneficial effects of the present invention are as follows: the sorting production line of the present invention reduces the false detection rate during detection by performing dust removal processing through the dust removal module; the self-centering clamping pallet uses the pallet's own gravity to trigger the automatic clamping mechanism to fix the position of the display on the pallet and thus complete the signal power connection; the conveying module is divided into two parts, one part is the upper detection and transportation mechanism and the other part is the lower defective product transportation mechanism. The conveying module has the function of adjustable centering width to adapt to pallets of different sizes. The upper detection conveyor adopts a synchronous belt conveyor with higher motion accuracy, which ensures the reliability of the signal connection between the detection equipment and the pallet; the automatic sorting module separates the defective products by rotating 90°, and uses the size difference between the length and width of the pallet to achieve a simpler separation of the defective products; the image acquisition module can be used not only for single-station acquisition, but also for multi-station acquisition The image acquisition module uses a counter number to determine whether to capture images of the display screen passing directly under the image acquisition module, effectively avoiding the problem of repeated detection and improving detection efficiency. Furthermore, the model and sorting method disclosed in the present invention effectively solve the problems of low efficiency and low detection accuracy of manual visual detection in the current TFT-LCD liquid crystal display mura defect detection operation, and the traditional detection method is not suitable for point mura (spot mura) defect detection. By using the improved YOLOV4-Tiny deep learning method for analysis, the recognition and sorting results are automatically output. Combined with the automatic sorting module, unqualified products are automatically rejected, effectively reducing the labor intensity of workers and improving the sorting accuracy. Therefore, the product quality of the TFT-LCD liquid crystal display is guaranteed, and users are provided with a better user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a flow chart of the present invention;

[0016] Figure 2 This is a Mura defect image of a TFT-LCD display;

[0017] Figure 3 Flowchart for constructing a dataset of mura defects on TFT-LCD displays;

[0018] Figure 4 Label the TFT-LCD display mura defect image for the LabelImg software;

[0019] Figure 5 Flowchart for model training and generation of frozen models;

[0020] Figure 6 This is the improved YOLOV4-Tiny deep learning network model diagram of the present invention;

[0021] Figure 7 This is the submodule structure diagram of the YOLOV4-Tiny deep learning network model;

[0022] Figure 8 This is a comparison chart of the detection effect;

[0023] Figure 9 This is a comparison chart of the AP results of the improved YOLOV4-Tiny target detection algorithm and the YOLOV4-Tiny original network;

[0024] Figure 10 This is an image acquisition implementation diagram of the image acquisition module;

[0025] Figure 11 It is the three-dimensional axonometric assembly drawing of the system;

[0026] Figure 12 This is the main view of the inspection system (the current situation is that camera 2 detects a defective product);

[0027] Figure 13 This is the structure diagram of the dust removal module;

[0028] Figure 14 This is the structural diagram of the self-centering clamping pallet;

[0029] Figure 15 Structural diagram of the automatic clamping mechanism with and without the upper cover removed;

[0030] Figure 16 This is the axonometric drawing of the synchronous belt conveyor;

[0031] Figure 17 This is a top view of the synchronous belt conveyor;

[0032] Figure 18 This is the axonometric drawing of the double-speed chain conveyor;

[0033] Figure 19 This is the axonometric diagram of the image acquisition module;

[0034] Figure 20 This is the axonometric drawing of the automatic sorting module;

[0035] Figure 21 This is the structural diagram of the signal connector;

[0036] Figure 22 It is the structural diagram of the conductive wheel;

[0037] Figure 23 Remove the automatic clamping mechanism diagram for the self-centering clamping pallet;

[0038] Figure 24 This is the internal coordination diagram of the automatic clamping mechanism;

[0039] Figure 25Top view of the mounting shell for the automatic clamping mechanism;

[0040] Figure 26 To trigger the capping from different angles;

[0041] Figure 27 This is a diagram showing the positional relationship between the pawl, spring, and one-way ratchet when the one-way ratchet is working;

[0042] Figure 28 This is a diagram showing the relationship between the trigger cover and the pawl positions when the trigger device is not triggered;

[0043] Figure 29 This is a top view of the cross-section when the trigger mechanism is not triggered (position relationship between the trigger cover and the pawl);

[0044] Figure 30 This is a diagram showing the relationship between the trigger cover and the pawl positions when the trigger device is triggered;

[0045] Figure 31 This is a top view cutaway diagram of the trigger mechanism when it is triggered (the position relationship between the trigger cover and the pawl);

[0046] Figure 32 This is an exploded view of the trigger device;

[0047] Figure 33 Install a roadmap for the triggering device.

[0048] The numbers in the figure are: 1-dust removal module, 2-self-centering clamping tray, 3-transmission module, 4-automatic sorting module, 5-defect detection module, 6-dust removal roller, 7-air dust collector, 8-frame I, 9-sprocket I, 10-tray, 11-conductive copper sheet, 12-signal interface, 13-guide, 14-guide wheel, 15-gear, 16-transverse retraction rack, 17-longitudinal retraction rack, 18-trigger cover, 19-one-way ratchet, 20-pawl, 21-disc coil spring, 22-fixed shaft, 23-mounting shell, 24-clip, 25-trigger device , 26-frame II, 27-synchronous belt, 28-synchronous pulley, 29-auxiliary wheel, 30-pulley mounting support, 31-auxiliary wheel mounting support, 32-beam, 33-pulley mounting plate, 34-tensioning pulley, 35-drive shaft, 36-gearbox I, 37-servo motor I, 38-mounting plate I, 39-second support arm, 40-adjusting screw I, 41-nut block I, 42-slider link support, 43-slider I, 44-slide rail I, 45-frame III, 46-slide rail II, 47-slider II, 48-slider mounting support, 49-chain guide, 5 0-sprocket Ⅱ, 51-second support arm, 52-nut block Ⅱ, 53-adjusting screw, 54-sprocket mounting block, 55-driven sprocket shaft, 56-driving sprocket shaft, 57-support arm mounting seat, 58-adjusting screw fixing seat, 59-gear box Ⅱ, 60-servo motor Ⅱ, 61-speed chain, 62-anti-slip platform, 63-gear shaft, 64-rack, 65-90° reversing cylinder, 66-mounting plate Ⅱ, 67-vertical lifting cylinder, 68-mounting base plate, 69-beam, 70-column, 71-slider Ⅲ, 72-slide rail Ⅲ, 73-screw, 74 -nut, 75-camera, 76-light shield, 77-servo motor III, 78-interface, 79-connecting pin, 80-magnet, 81-signal line, 82-spring I, 83-wire screw plug, 84-connecting block, 85-electromagnet, 86-mounting bracket, 87-rocker arm, 88-clamping device, 89-conductive roller, 90-power interface, 91-signal interface, 92-reset spring, 93-fixing nut, 94-pawl mounting bracket, 95-fixing pin, 96-spring II, 97-screw mounting block, 98-top cover, 99-one-way ratchet fixing screw. DETAILED DESCRIPTION

[0049] The invention will be further described below with reference to the accompanying drawings and embodiments, but the content of the present invention is not limited to the scope of the drawings.

[0050] Example 1: Figure 11-33As shown, according to one aspect of an embodiment of the present invention, a liquid crystal display defect detection and sorting production line is provided, comprising a self-centering clamping tray 2, a conveying module 3, an automatic sorting module 4, and a defect detection module 5; the self-centering clamping tray 2 is placed on the conveying module 3, and the self-centering clamping tray 2 is driven to move by the conveying module 3, and the liquid crystal display on the self-centering clamping tray 2 is detected for defects by the defect detection module 5, and the automatic sorting module 4 is used to sort the liquid crystal display according to the recognition results of the defect detection module 5.

[0051] Optionally, a dust removal module 1 is also included, such as Figure 13 As shown, the dust removal module 1 includes a dust removal roller 6, an air-blowing dust collector 7, a frame I8, and a sprocket I9; wherein, the sprocket I9 is ​​installed at both ends of the roller shaft of the dust removal roller 6 using a flat key, and is axially fixed using a shaft shoulder and a shaft end retaining ring; the dust removal roller 6 and the air-blowing dust collector 7 are installed on the frame I8 in a front-to-back manner, and the air-blowing dust collector 7 is close to the forward end of the conveying direction of the conveying module 3; the sprocket I9 is ​​powered by the servo motor I37 of the upper synchronous belt conveyor in the conveying module 3. Optionally, the dust removal roller 6 fixed to the frame Ⅰ 8 by a bearing support is composed of a roller shaft, a sponge core, and a soft cotton cloth provided from the inside to the outside. The sponge core and the roller shaft are fixedly connected by bonding. The soft cotton cloth can disturb the dust on the LCD screen and further facilitate the dust removal in cooperation with the air-blowing dust collector 7; the lengths of the sponge core and the soft cotton cloth are adapted to the display panel; the air-blowing dust collector is composed of a steel pipe with small holes and a valve, which is installed at both ends of the steel pipe through pipe threads. The valve is used to adjust the force of the air blowing, and the air-blowing dust collector 7 is arranged parallel to the dust removal roller 6 and the air-blowing dust collector 7 does not affect the normal rotation of the dust removal roller 6. The air-blowing dust collector 7 can adjust the air blowing angle and the distance to the display screen, and the distance between the dust removal roller and the display panel can also be adjusted; the entire dust removal module 1 is installed on the upper synchronous belt conveyor by screws, and the power source required for the rotation of the dust removal roller 6 is provided by the upper synchronous belt conveyor through chain transmission; Figure 10 Taking the transport direction shown as an example, the drum rotates clockwise. Air dust collector 7 is used to further remove fine dust after the dust removal drum 6 has cleaned it. Air blows downward and to the left, with one velocity component directed against the tray's forward motion. The dust removal module removes dust and particles from the surface of the display to be inspected, preventing them from affecting the accuracy of subsequent inspections.

[0052] like Figure 11-12As shown, optionally, the dust removal module 1 is fixed to the conveying module 3 by bolts, the self-centering clamping tray 2 is placed on the conveying module 3, the automatic sorting module 4 is fixed to the ground by anchor bolts, and the defect detection module 5 is fixed to the conveying module 3 by bolts; wherein the dust removal module 1 is used to remove dust from the LCD screen to be inspected; the conveying module 3 is used to convey the self-centering clamping tray with the LCD screen; the self-centering clamping tray 2 is used to fix the display screen on the tray to ensure the safety of the display screen during transportation on the tray, and is used to connect the signal and power supply between the display screen and the tray; the defect detection module 5 includes: an image acquisition module and a controller, the image acquisition module is used to capture the image of the LCD screen and transmit the captured image to the controller, the controller is used to control the normal operation of the entire equipment; the automatic sorting module 4 is used for automatic sorting of defective products.

[0053] Alternatively, as Figure 14-15 As shown in , 24 and 25, the self-centering clamping tray 2 includes a tray 10 and a self-centering clamping mechanism; wherein, the self-centering clamping mechanism is installed on the tray 10, and the self-centering clamping mechanism includes a transverse retractable rack 16, a longitudinal retractable rack 17, a buckle 24, and a trigger device 25; wherein, the buckle 24 is installed at one end of the transverse retractable rack 16 and the longitudinal retractable rack 17, and the other end of the transverse retractable rack 16 and the longitudinal retractable rack 17 cooperates with the trigger device 25 to realize the clamping of the LCD screen between the buckles 24. The self-centering clamping machine can also include a mounting shell 23; wherein, the transverse shrinkage rack 16 and the longitudinal shrinkage rack 17 are installed in the guide groove of the mounting shell 23, and do not interfere with each other, and the anti-slip buckle 24 is fixed to the end of the transverse shrinkage rack 16 and the longitudinal shrinkage rack 17 by screws / welding (that is, the two transverse shrinkage racks 16 are respectively installed with anti-slip buckles 24 at one end away from the trigger device 25, and the two longitudinal shrinkage racks 17 are respectively installed with anti-slip buckles 24 at one end away from the trigger device 25), the trigger device 25 is installed in the center of the mounting shell 23, and the end of the fixed shaft 22 of the trigger device 25 is cross-shaped and fixed to the mounting shell 23 to prevent the entire trigger device 25 from rotating around the axis direction. The self-centering clamping mechanism is designed with an upper cover 98 to prevent internal parts from shifting. Furthermore, a transverse retractable rack 16 is provided, which includes an upper and lower layer with a spacing between the two layers. A longitudinal retractable rack 17 is installed between the upper and lower layers, and a protrusion is provided at one end of the longitudinal retractable rack 17 where the anti-slip buckle 24 is not installed. The protrusion cooperates with the transverse retractable rack 16 for limiting positioning.

[0054] Alternatively, as Figure 26-33As shown, the trigger device 25 includes a trigger cover 18, a gear 15, a one-way ratchet mechanism, a disc coil spring 21, a fixed shaft 22, a return spring 92, and a fixed pin 95; the trigger cover 18 is essentially a spatial cam, which is installed on one end of the fixed shaft 22 through the fixed pin 95, and a strip through hole is opened on the cylindrical body of the trigger cover 18 for cooperating with the fixed pin 95, so that the trigger cover 18 can achieve a certain angle of circumferential rotation and a certain distance of axial movement (in the embodiment of the present invention, 4 strip through holes are set at intervals of 90 degrees, and any two strip through holes set opposite to each other cooperate with the fixed pin 95, that is, any two strip through holes at intervals of 180 degrees cooperate with the fixed pin 95), a return spring 92 is installed between the fixed shaft 22 and the trigger cover 18, and the entire one-way ratchet mechanism is fixed to the fixed shaft 22 by a spline, and the gear 15 is meshed with the transverse contraction rack 16 and the longitudinal contraction rack 17. At the same time, the one-way ratchet 19 of the one-way ratchet mechanism is fixedly connected to the gear 15 by a spline. When the gear 15 rotates, the one-way ratchet 19 of the one-way ratchet mechanism is driven to rotate together. One end of the rotating coil spring 21 is installed in the fixed groove of the gear 15. At the same time, the other end of the rotating coil spring 21 is fixed on the mounting shell 23 of the self-centering clamping mechanism; the one-way ratchet mechanism includes a one-way ratchet 19, a pawl 20, a pawl mounting frame 94, and a spring II 96; wherein, the pawl mounting frame 94 is installed in the one-way ratchet 19, one end of the pawl mounting frame 94 is splined with the fixed shaft 22 and is fastened by a fixed fixing nut 93, the pawl 20 is installed on the outer ring of the pawl mounting frame 94, and a spring II 96 is arranged between the outer ring of the pawl mounting frame 94 and the pawl 20, and the insert arranged at one end of the trigger pressure cover 18 cooperates with the pawl 20. The one-way ratchet fixing screw 99 is used to connect the one-way ratchet 19 and the pawl mounting frame 94 in the one-way ratchet mechanism. The one-way ratchet fixing screw is installed in the threaded hole of the one-way ratchet 19 through a thread, and cooperates with the annular groove of the pawl mounting frame 94, allowing the one-way ratchet 19 to rotate, but there will be no axial slippage.

[0055] When the user pulls out the horizontal or vertical retraction rack, the gear 15 is driven to rotate, and the one-way ratchet 19 connected thereto rotates and tightens the disc-shaped coil spring 21; when the user releases his hands, the one-way ratchet is stuck by the pawl 20 through the one-way ratchet 19 and cannot rotate. Figure 27 As shown in state a, after the display screen is placed, the display screen is pressed against the trigger cover 18 due to gravity, and the trigger cover 18 moves downward and rotates at a certain angle, so that the insert provided at one end of the trigger cover 18 is inserted between the one-way ratchet 19 and the one-way ratchet pawl 20, thereby pushing the pawl 20 inward. Figure 27 As shown in state b, the one-way ratchet 19 is disengaged from the one-way ratchet pawl 20. As the disc-shaped coil spring 21 is relaxed, the one-way ratchet 19 drives the rack to contract inwards and thus clamps the LCD screen.

[0056] Alternatively, as Figure 14 、 23 As shown, the self-centering clamping pallet 2 also includes a conductive copper sheet 11, a signal interface 12, a guide path 13 and a guide wheel 14. The conductive copper sheet 11, the signal interface 12, the guide path 13 and the guide wheel 14 are all installed on the pallet 10. The self-centering clamping mechanism is fixed to the center of the pallet 10 by screws. The conductive copper sheet 11 and the signal interface 12 are located on both sides of the self-centering clamping mechanism that are perpendicular to the transmission direction of the upper synchronous belt conveyor. The conductive copper sheet 11 is installed in a rectangular groove at a fixed position on the pallet 10 and is fixed by screws. The signal interface 12 is fixed to the pallet 10 by bonding. The guide path 13 is installed in the fixed installation groove of the pallet 10 by interference fit, and the guide wheel 14 is installed at the four corners of the pallet 10 by screws. Conductive copper sheet 11 and signal interface 12 serve as the tray's power and signal interfaces. Conductive copper sheet 11 connects to power via conductive roller 89, and signal interface 12 connects signals via a signal connector. Tray 10 is equipped with fixed power interface 90 and signal interface 91 for signal and power connections between the display and tray 10. The entire tray 10 is equipped with various necessary wires.

[0057] The signal connector is as follows Figure 21As shown, the connector primarily comprises: an interface 78, a connecting pin 79, a magnet 80, a signal line 81, a spring I 82, a wire-pressing plug 83, a connecting block 84, and an electromagnet 85. Signal line 81 passes through spring I 82 and the connecting block 84, with one end extending out of the connecting block 84. The wire-pressing plug 83 is mounted in the connecting block 84, and the electromagnet 85 is secured to the connecting block 84 with a nut. The entire signal connector is mounted on a signal connector bracket and secured to the upper synchronous belt conveyor with screws. The connecting pin 79 and magnet 80 fit into the fixing slot of the interface 78, and the connecting pin 79 connects to the signal line 81. Interface 78 is designed as a funnel, mating with the signal interface 12 on the tray. Adjusting the compression of the spring can provide a certain degree of rigidity to the signal line while maintaining a certain degree of flexibility, ensuring reliable signal connection. Combined with the magnetic attraction, this allows for more efficient automatic signal connection. Furthermore, the signal interface utilizes a magnetically assisted connection method, which reduces positioning accuracy requirements. When connecting the signal, the signal line must be pre-tightened to compress spring I 82 and compress the wire-pressing screw 83. The tension applied to signal line 81 by spring I 82 imparts a certain degree of rigidity, yet also a certain degree of flexibility. Magnetic attraction and the bell-shaped design of the interface, combined with the signal interface, facilitate signal connection. When the tray 10 arrives at the inspection position and stops, the controller energizes electromagnet 85 and moves it downward, completing the signal connection to the display to be inspected. The transversely retractable rack 16 and longitudinally retractable rack 17 are mounted in the guide grooves of the mounting housing 23 without interfering with each other. The housing is designed with an upper cover to hold the racks in place. Anti-slip clips 24 are welded to the ends of the racks to clamp the display and prevent it from sliding. The tray 10 is equipped with a fixed power interface 90 and a signal interface 91 for signal and power connections between the display and the tray 10. The necessary wires are distributed throughout the tray 10.

[0058] The conductive wheel set is as follows Figure 22 As shown, it mainly includes: a mounting frame 86, a rocker arm 87, a clamping device 88, and a conductive roller 89; the rocker arm 87 is mounted on the mounting frame 86 via screws, and the conductive roller 89 is also mounted on the rocker arm 87 via screws. A clamping device 88 is installed between the rocker arm 87 and the mounting frame 86 to compress the conductive roller 89, thereby ensuring a reliable connection between the conductive roller and the conductive copper sheet. The clamping device 88 is T-shaped, with both ends of the horizontal pin connected to the rocker arm and the free end of the vertical telescopic rod connected to the mounting frame 86. The vertical telescopic rod includes a thick rod and a thin rod, and the thin rod is covered with a spring. The entire set of conductive wheels is mounted on an insulating connecting block. The three sets of conductive wheels are mounted on the bracket through the insulating connecting block. The three sets of conductive wheels are respectively the neutral wire, the live wire, and the ground wire.

[0059] Optionally, the conveying module 3 includes an upper synchronous belt conveyor and a lower double-speed chain conveyor; the upper synchronous belt conveyor includes a first conveying module and a first adjusting module, the first conveying module is used to provide power through the first power system to drive the synchronous belt 27 to move, and the movement of the synchronous belt 27 in the first conveying module drives the self-centering clamping pallet 2 to follow the movement; the first adjusting module is used to drive the nut block Ⅰ41 installed on the adjusting screw rod Ⅰ40 to move up and down through the adjusting screw rod Ⅰ40, and the up and down movement of the nut block Ⅰ41 drives the angle of the two first support arms 39 connected to the nut block Ⅰ41 to change, and the change in the angle of the first support arm 39 drives the first conveying module at the other end of the two first support arms 39 to move in the opposite direction with the first support arm 39, and the movement direction is the same as the conveying direction of the first conveying module. Vertical; the lower-level double-speed chain conveyor is installed at the lower part of the upper-level synchronous belt conveyor, and includes a second conveying module and a second adjusting module. The second conveying module is used to provide power through the second power system to drive the double-speed chain 61 to move, and the movement of the double-speed chain 61 of the second conveying module drives the self-centering clamping tray 2 sorted out by the automatic sorting module 4 to follow the movement; the second adjusting module is used to drive the nut block Ⅱ52 installed on the adjusting screw rod Ⅱ53 to move up and down through the adjusting screw rod Ⅱ53, and the up and down movement of the nut block Ⅱ52 drives the angle of the two second support arms 51 connected to the nut block Ⅱ52 to change, and the change of the angle of the second support arm 51 drives the second conveying module at the other end of the two second support arms 51 to move in the opposite direction with the first support arm 39, and the movement direction is perpendicular to the conveying direction of the second conveying module.

[0060] like Figure 16-18As shown, the first transmission module in the upper synchronous belt conveyor includes a frame II 26, a synchronous belt 27, a synchronous pulley 28, an auxiliary wheel 29, a pulley mounting support 30, an auxiliary wheel mounting support 31, a beam 32, a pulley mounting plate 33, a tensioning wheel 34, a drive shaft 35, a gear box I 36, a servo motor I 37, and a mounting plate I 38. The first adjustment module includes a first support arm 39, an adjustment screw rod I 40, a nut block I 41, a slider link support 42, a slider I 43, a slide rail I 44, and a screw rod mounting block 97; the servo motor 37 and the gear box as the first power system Ⅰ36 is fixed by bolts, gear box Ⅰ36 is installed on mounting plate Ⅰ38 and mounting plate Ⅰ38 is welded to frame Ⅱ26, drive shaft 35 is fixed to gear box Ⅰ36 by flat key and bearing end cover, synchronous pulley 28 is installed on both ends of drive shaft 35 by spline, and synchronous pulley 28 can move axially on drive shaft 35, slider Ⅰ43 and slide rail Ⅰ44 cooperate with each other; slider link support 42 is installed with slider Ⅰ43 by bolts, synchronous pulley 28 is installed between two pulley mounting supports 30 through pulley rotating spindle, pulley mounting support 30 One end is connected to the slider link support 42, and the other end of the pulley mounting support 30 is fixedly connected to the beam 32; multiple auxiliary wheels 29 are installed between the two auxiliary wheel mounting supports 31; the synchronous pulley 28, auxiliary wheel 29 and tensioning wheel 34 on the same side are connected by a synchronous belt 27; a pulley mounting plate 33 is installed on the beam 32 by bolts for installing the tensioning wheel 34 and limiting the drive shaft 35; bearings are installed in the bearing holes at both ends of the first support arm 39, and the inner rings of the bearings at one end of the two first support arms 39 are connected to the nut block I 41 through the shaft, and the bearings at the other end of the first support arm 39 are connected to the belt of the side The wheel rotation spindle is matched with the shaft, and the axial movement of the adjusting screw rod Ⅰ40 is limited by the screw rod mounting block 97, and only circumferential rotation exists; the servo motor 37 drives the lower synchronous pulley to move the synchronous belt 27, driving the synchronous pulleys at both ends to rotate, thereby completing the transportation task; by rotating the handle to drive the adjusting screw rod Ⅰ40 to move, the distance between the two beams of the synchronous belt conveyor can be changed to adapt to pallets of different sizes. The synchronous belt conveyor is driven by a servo motor with high movement accuracy, which can make the pallet stop at a specified place, ensuring the reliability of the subsequent signal connection of the display screen to be tested. The lower-level speed chain includes a frame III 45, a slide rail II 46, a slider II 47, a slider mounting support 48, a chain guide 49, a sprocket II 50, a second support arm 51, a nut block II 52, an adjusting screw 53, a sprocket mounting block 54, a driven sprocket shaft 55, a driving sprocket shaft 56, a support arm mounting seat 57, an adjusting screw fixing seat 58, a gear box II 59, a servo motor II 60, and a speed chain 61; the servo motor II 60 serving as the second power system is fixed on the chain guide through a bracket, and the gear box II 59 and the driving connecting wheel shaft are key-connected to drive the driving sprocket to move, and the servo motor II 60 rotates to drive the driving sprocket, thereby driving the entire conveyor through the speed chain 61 to complete the transportation work.Its overall installation is similar to that of the upper synchronous belt conveyor. The lower double-speed chain conveyor also features centering adjustment, and the adjustment principle is the same as for the synchronous belt conveyor, so I won't go into detail here. This structure uses a double-speed chain and synchronous pulleys for transmission. The upper synchronous belt conveyor and the lower double-speed chain conveyor combine to achieve cost savings. The gearbox can be a turbine gearbox.

[0061] Alternatively, as Figure 20 As shown, the automatic sorting module 4 includes: an anti-skid platform 62, a gear shaft 63, a rack 64, a 90° reversing cylinder 65, a mounting plate II 66, a vertical lifting cylinder 67, and a mounting base plate 68; one end of the gear shaft 63 is mounted on the mounting plate II 66 through a thrust ball bearing, which effectively reduces the friction during the rotation of the gear shaft 64; the 90° reversing cylinder 65 is fixedly connected to the mounting plate II 66, and the end of the push rod of the 90° reversing cylinder 65 is installed with a rack 64, which meshes with the outer ring gear of the gear shaft 63, and the anti-skid platform 62 is fixed to the other end of the gear shaft 63 by a threaded connection, and the axial direction of the gear shaft 63 is perpendicular to the movement direction of the push rod end of the 90° reversing cylinder 65; the output end of the vertical lifting cylinder 67 is threadedly connected to the mounting plate II 66. By applying the above technical solution, the vertical lifting cylinder 67 drives the mounting base 68 and the anti-slip rubber platform 62, gear shaft 63, rack 64, and 90° reversing cylinder 65 installed on the upper part of the mounting base 68 to move vertically along with the mounting base. The 90° reversing cylinder 65 drives the rack 64 to move linearly. The engagement of the rack and the gear shaft drives the gear shaft to rotate, thereby driving the anti-slip rubber platform 62 on the gear shaft to rotate synchronously. By rotating the pallet 90°, the function of product separation is achieved, thereby realizing the separation of good and defective products.

[0062] Optionally, the defect detection module 5 includes a photoelectric stop switch and one or more defect detection components. Each defect detection component group has the same structure, including an image acquisition module and a controller. The photoelectric counter in each image acquisition module is used to count the number of passing pallets and transmit the count number to the respective controller. The controller drives the camera 75 to capture images of the LCD screen based on the trigger signal of the photoelectric stop switch and the count number of the photoelectric counter, and then performs defect detection on the LCD screen based on the captured images. It should be noted that the number of automatic sorting modules 4 is the same as the number of defect detection components.

[0063] Alternatively, as Figure 19As shown, the image acquisition module includes a crossbeam 69, a column 70, a slider III 71, a slide rail III 72, a camera 75, a light shield 76, a servo motor III 77, and a photoelectric counter; a moving device is installed on the crossbeam 69 and the column 70, and the moving device includes a screw rod 73 and a nut 74; mounting plates are respectively installed at both ends of the crossbeam 69 and the column 70 by screws to support the screw rod 73 and place the bearing, and the bearing and the screw rod 73 adopt an interference fit, and the nut 74 on the column 70 is fixedly connected to the slider III 71 by screws. Slide block III71 cooperates with slide rail III72 to accommodate changes in the spacing between two uprights 70 fixed to beam 32. (Once slider III71 is moved to the desired position, it can be bolted to slide rail III72.) Beam 69 and slide rail III72 are screwed together. Camera 75 is mounted on a camera bracket and screwed to nut 74 on beam 69. Servo motor III77 is screwed to the mounting plate and connected to screw rod 73 via a coupling, providing power for its rotation. A photoelectric counter counts the number of pallets passing through and transmits the count to the controller. The controller captures images for the LCD screen based on the trigger signal from the photoelectric stop switch and the count number from the photoelectric counter. The screw-nut mechanism gives camera 75 two degrees of freedom, enabling the capture of the clearest images. Camera 75 is secured to nut 74 via a connector. Bearings are installed at both ends of screw 73, and servo motor III 77 is connected to screw 73 via a coupling. A light shield 76 prevents interference from other light sources on the camera. Servo motor III 77 is mounted to the corresponding crossbeam 69 and column 70 via screws and mounting plates. The controller can be a host computer. The controller's detection model uses a liquid crystal display defect detection model. Considering that spot mura is the most representative and difficult defect type among LCD screen defects, the present invention's detection model demonstrates significant improvements in the average precision, missed detection rate, false detection rate, and accuracy of spot mura detection compared to the original network. This allows for more effective spot mura detection, providing a practical basis for the detection effectiveness of the liquid crystal display defect detection model.

[0064] According to another aspect of an embodiment of the present invention, a liquid crystal display defect detection model is provided for use in a defect detection module 5 in a liquid crystal display defect detection and sorting production line. Based on the YOLOV4-Tiny model, the output of the first CSP module in the backbone feature extraction network is added as a shallow feature map. A convolution operation is performed on the original deep feature map to adjust the number of channels and then divided into two outputs: one output is upsampled and then fused with the middle feature map. The fused result is then upsampled and fused with the newly added shallow feature map. The resulting feature map is then processed by a Ghost module and input to a YOLO head for classification and regression prediction. The other output is processed by a Ghost module and then input to a YOLO head for classification and regression prediction. The newly introduced shallow feature map [76,76,128] has more detailed information, which is beneficial for detecting small targets. The addition of the Ghost module to the feature map after feature fusion yields a similar feature map with concentrated features. Using the Ghost module instead of ordinary convolution reduces the amount of computation. Although the improved YOLOV4-Tiny target detection algorithm has a lower frame rate, its average precision, missed detection rate, false detection rate and accuracy are greatly improved compared to the original network. It can more effectively detect spot mura. The improved algorithm structure is as follows: Figure 6 As shown, the submodule structure is as follows Figure 7 shown.

[0065] Call the LCD defect detection model and provide the categories and names required for training. This model is a lightweight target detection network with low computer resource usage and fast real-time detection, making it suitable for industrial implementation. It uses CSPdarknet53_tiny as the backbone feature extraction network. Three effective feature layers are derived from this backbone feature extraction network to strengthen the feature extraction network. These three effective feature layers are [19, 19, 512], [38, 38, 256], and [76, 76, 128]. [19, 19, 512] is a deep feature layer with rich semantic information, while [76, 76, 128] is a shallow feature layer suitable for small target detection. In the Neck part, the effective feature layer of [19,19,512] is convolved once to adjust the number of channels. The feature map after channel adjustment is processed by the Ghost module and input into the yolo head for regression prediction and classification prediction; in addition, the convolved [19,19,512] effective feature layer is upsampled once and fused with the feature map of [38,38,256] to obtain a feature map with richer feature information; the result of feature fusion is upsampled again and then fused with the feature map of [76,76,128] to obtain an effective feature map with deep, middle and shallow semantic information, which is then input into the Ghost module for processing. Finally, the feature map is input into the yolo head for regression prediction and classification prediction; the prediction information includes the target type name, type confidence, target center point coordinates and the height and width of the target box. The role of the Ghost module is to replace the ordinary convolution and reduce the amount of calculation. The Ghost module consists of a 1×1 convolution and a depthwise separable convolution. The 1×1 convolution adjusts the number of channels in the input feature map to achieve the necessary feature concentration of the input features; the depthwise separable convolution generates a similar feature map of the concentrated features. This improvement introduces shallow features to fuse shallow and deep features. The Ghost module then generates a similar feature map of the concentrated features, which is then input into the YOLO head for regression and classification prediction, enhancing the YOLOV4-Tiny object detection algorithm's ability to detect spot mura defects.

[0066] like Figure 5According to another aspect of an embodiment of the present invention, a method for training a liquid crystal display defect detection model is provided, including: constructing a liquid crystal display mura defect image dataset; calling the liquid crystal display defect detection model, inputting the model's backbone weights, and pre-training the model using a training set in the defect image dataset to obtain multiple weight parameters; inputting multiple weight files obtained from the pre-training into the liquid crystal display defect detection model, and using a test set in the defect image dataset to screen out optimal pre-training weights, inputting the optimal pre-training weights into the liquid crystal display defect detection model, and using the training set in the mura defect image dataset to perform reinforcement training on the model to obtain multiple weights after reinforcement training; testing the obtained weights after reinforcement training using the test set in the mura defect image dataset and selecting the optimal weights after reinforcement training, and inputting the optimal weights after reinforcement training into the liquid crystal display defect detection model to generate a frozen model.

[0067] like Figure 1-10 According to another aspect of an embodiment of the present invention, a method for detecting and sorting defects in a liquid crystal display screen is provided, comprising:

[0068] Start the LCD screen visual real-time detection and sorting system; the transmission module 3 supplies the display screen to be detected, starts the image acquisition module and adjusts the camera parameters, and the entire device starts running; for example Figure 10 When there are two sets of defect detection components, conveyor module 3 is turned on. Pallet 1 first passes counter 2, then photoelectric counter 1. When pallet 1 passes counter 1, pallet 2 just passes counter 2. At this time, the increment of counter 2 reaches 2 and the stop switch is triggered, meeting the stop condition of the upper synchronous belt conveyor. At this time, the upper synchronous belt conveyor stops. Image acquisition begins, and inspection and automatic sorting are performed. After one inspection is completed, the count results of counters 1 and 2 are cleared. Wait for pallet 3 to pass counter 2 first and then counter 1. When pallet 3 passes counter 1, pallet 4 passes counter 2. At this time, the increment of counter 2 reaches 2 and the stop switch is triggered, meeting the stop condition of the upper synchronous belt conveyor. The conveyor stops, and acquisition begins, and inspection and automatic sorting are performed. Subsequent pallets are inspected in the same way.

[0069] The LCD screen placed on the self-centering clamping tray 2 is driven by the conveying module 3 to move toward the defect detection module 5. The defect detection module 5 performs defect detection on the LCD screen on the self-centering clamping tray 2. The automatic sorting module 4 is used to sort the LCD screen according to the recognition results of the defect detection module 5.

[0070] Optionally, defect detection is performed on the LCD screen on the self-centering clamping tray 2 through the defect detection module 5, including: the photoelectric counter in the image acquisition module in the defect detection module 5 is used to count the passing pallets and transmit the count number to the respective controllers, the controller drives the camera 75 to capture the image of the LCD screen based on the trigger signal of the photoelectric sensing stop switch and the count number of the photoelectric counter, and performs defect detection on the LCD screen based on the captured image.

[0071] Optionally, the controller drives the camera 75 to capture images of the LCD screen according to the trigger signal of the photoelectric sensor stop switch and the count number of the photoelectric counter, and performs defect detection on the LCD screen according to the captured image. Specifically, if the defect detection components in the defect detection module 5 are a group, for the current detection, the photoelectric counter counts to 1 and the photoelectric sensor stop switch is triggered, then the controller drives the upper synchronous belt conveyor in the transmission module 3 to stop moving, drives the camera 75 to capture images of the LCD screen, and calls the freeze model to perform defect detection on the captured LCD screen image. When the detection is completed, the photoelectric counter The counters are reset to zero and the next test is performed until the test is completed. If there are n groups of defect detection components in the defect detection module 5, for the current test, if the maximum increment of the n photoelectric counters is m and the photoelectric sensor stop switch is triggered, the mth controller drives the upper synchronous belt conveyor in the transmission module 3 to stop moving, and the first m controllers respectively drive the camera 75 to capture the image of the LCD screen and call the freeze model to perform defect detection on the captured LCD screen image. When the test is completed, all photoelectric counters are reset to zero and the next test is performed until the test is completed. Among them, the n photoelectric counters correspond to the n controllers one by one. For example, there are three groups of defect detection components. During a test, if the increment of photoelectric counter No. 3 is 3 and the stop switch is triggered, the upper synchronous belt conveyor stops. If the maximum increment of the three counters is 2, it is determined that only the last two displays are left. At this time, the maximum increment of counter No. 2 is used as the condition for determining the stop of the upper synchronous belt conveyor. If the maximum increment of the three counters is 1, it is determined that only the last display is left. At this time, the maximum increment of counter 1 is used as the condition for determining the stop of the upper synchronous belt conveyor. It should be noted that after completing a test, a counter will be reset to zero. At this time, the maximum value of the three counts is the maximum increment of the test situation. For example, in three groups of defect detection components, the maximum value of the three counters is the maximum increment.

[0072] Furthermore, in the case where there are two sets of defect detection components, the following optional implementation process is given:

[0073] S1. Randomly classify the collected 1000 LCD screen Mura images into training set images and validation set images in a 9:1 ratio; Figure 2 、3 shown; Figure 2 Only some of the Mura defect images of the crystal display are shown;

[0074] S2. Use LabelImg labeling software to label the anchor frame foreign objects in the training set images and the validation set images. The labeling includes: the horizontal and vertical coordinates of the target information center position, the length and width of the bounding box, and the foreign object category. When labeling, it is necessary to determine the category of the LCD screen mura defect, which will be named as the mura class, as shown in the attached figure. Figure 4 The figure shows the anchor diagram of the LCD screen spot mura defect using the LabelImg annotation tool; Figure 4 Grayscale display of images with six different backgrounds: red, blue, green, black, white, and gray;

[0075] S3. Convert all the labeled Mura defect information of the LCD screens to a unified format. That is, use a Python script to convert the label file from .xml file format to .txt file format. This format conversion makes the data more suitable for deep learning in the subsequent steps.

[0076] S4. Put the results of the annotation conversion into the corresponding training set image folder and verification set image folder respectively, together forming a liquid crystal display mura defect image dataset consisting of a training set and a verification set.

[0077] S5. Before pre-training begins, modify the hyperparameters of the configuration file. The hyperparameters of the configuration file mainly include the learning rate, number of extracted images, number of iterations, and weight decay coefficient. The remaining hyperparameters are the default values. In this embodiment, the learning rate is 0.01, the number of extracted images is 8, the number of iterations is 250, the weight decay coefficient is 0.0005, the number of backbone network unfreezing iterations is 50, and the remaining hyperparameters are the default values. The experimental equipment is a desktop GPU with NVIDIA GeForce RTX3060, 8GB of memory, and an Intel Core i7-12700F CPU.

[0078] S6. Call the training set and configuration file, input the backbone weights of the model, and pre-train the deep learning network model in the deep learning network framework Pytorch; according to the hyperparameters of the configuration file, first freeze the parameters of the backbone network Backbone of the model, and freeze the iterations for 50 times. Then, randomly extract 8 pictures from the training set and put them into the deep model network model for pre-training, and update the weight parameters of other parts of the model. When the number of iterations of a single batch is set to 250, the weight parameter update is completed and the final weight parameters are generated; the weight update is specifically as follows: use the randomly initialized weight parameters to perform forward propagation calculation of the convolutional neural network and obtain a set of intermediate parameters, and then use the intermediate parameters for back propagation to obtain a new set of weight parameters; the new weight parameters will be the old weight parameters used for calculating the forward propagation before the iteration;

[0079] S7. Input a validation set and use the deep learning network framework Pytorch to perform a performance evaluation on all pre-trained weight parameters obtained in step S5, and screen out the optimal pre-trained weight parameters. The validation set used for quantitative performance evaluation is the validation set of the LCD display mura defect image dataset. The evaluation can be judged by average precision, frame rate, missed detection rate, false detection rate, and accuracy rate.

[0080] S8. Change the number of iterations to 350. Other parameters remain the same as in step S5. Load the optimal pre-trained weights into the model as the weights for the entire model. Perform intensive training on the model. First, freeze the parameters of the backbone network 50 times. During training, update the parameters, and perform forward and backward propagation the same as in step S5.

[0081] S9. Use the deep learning network framework Pytorch to input the validation set to perform a performance evaluation on all the obtained reinforcement training weight parameters, and screen out the optimal reinforcement training weight parameters. The validation set used for quantitative performance evaluation is the validation set of the LCD display mura defect image dataset. The evaluation can be judged by average precision, frame rate, missed detection rate, false detection rate, and accuracy rate.

[0082] S10, loading the optimal weight parameters into the target detection algorithm to generate a frozen model;

[0083] S11. Start the LCD display visual real-time inspection and sorting system, supply the display to be inspected, activate the image acquisition module, adjust the camera parameters, and the entire system begins operation. The upper synchronous belt conveyor speed is 0.2 m / s, and the lower double-speed chain speed is 0.25 m / s. The spacing between pallets is 200 mm. Camera parameters include image acquisition speed, pixel size, and memory space. Specifically, the industrial camera's power-on and power-off time is 1000 milliseconds, and it captures five 2590×1942 JPG images per second. The image acquisition card has 5000 MB of memory space. The industrial camera is a Basler color camera with a resolution of 2.3 MP and a frame rate of 51 fps. The industrial camera lens is a Basler Standard C-mount lens with a maximum image circle of 1 inch, an 8.0 mm fixed focal length, an aperture range of F1.4-F16, and a standard resolution of 2 megapixels. The lens hood is used to overcome ambient light interference and ensure stable lighting during image acquisition, resulting in optimal imaging results for image processing.

[0084] S12. The LCD screen placed on the self-centering clamping tray 2 is driven by the conveying module 3 to move toward the defect detection module 5. The defect detection module 5 performs defect detection on the LCD screen on the self-centering clamping tray 2. The automatic sorting module 4 is used to sort the LCD screen according to the recognition result of the defect detection module 5. Specifically: the LCD screen to be inspected on the production line is dusted by the dust removal module 1 and then conveyed by the synchronous belt conveyor through the photoelectric counter; the tray on the synchronous belt conveyor triggers the photoelectric sensor counter, and the photoelectric counter numbers the passing trays in a counting manner and transmits the number to the controller. When the tray triggers the photoelectric sensor at the specified position, it stops. The controller drives the upper synchronous belt conveyor in the transmission module 3 to stop moving according to the trigger signal of the photoelectric sensor stop switch and the count number of the photoelectric counter, and connects the signal and power of the display screen to be inspected. The controller controls the electromagnet to move downward to push the connector to complete the wiring under the auxiliary attraction of the magnetic block; so that the signal interface can be connected reliably. The power connection is connected through the conductive roller and the conductive copper bar. The controller controls the power on and off to complete the connection; the controller drives the camera 75 to shoot the area to be photographed to collect image information of the LCD screen to be inspected, and calls the frozen model to perform real-time detection of point mura (spot mura) defects on the image of the LCD screen to be inspected captured by the industrial camera. According to the result of the target real-time detection, it is determined whether there is a point mura (spot mura) defect: if the images captured by the two industrial cameras do not contain point mura (spot mura) defects, the controller controls the dust removal module 1, the upper synchronous belt conveyor and the image acquisition module to continue working; if the image captured by any industrial camera contains a spot mura defect, the controller controls the dust removal module 1, the upper synchronous belt conveyor and the image acquisition module to stop working; the real-time detection will generate each spot mura defect. The controller controls the automatic sorting module 4 directly below the pallet based on the detection output results. First, the vertical lift cylinder 67 rises to the specified height and stops. After waiting for 0.5 seconds, the 90° rotation cylinder 65 moves the rack 64, which drives the gear shaft 63 and the anti-slip rubber platform 62 to rotate 90 degrees before stopping. After waiting for 0.5 seconds, the vertical lift cylinder 67 descends to the specified height. The pallet is placed on the lower-level double-speed chain conveyor line, completing the automatic sorting of defective products. After sorting is complete, the automatic sorting module 4 resets and awaits the next controller instruction. The controller then controls the dust removal module 1, the upper-level synchronous belt conveyor, and the image acquisition module to continue operating.

[0085] The present invention tested 100 images containing some spot mura, collected from actual factories. Since spot mura occurs frequently in actual production processes, its detection is of great significance. Furthermore, the present invention can also identify other types of defects, as long as the network is trained with image data of other types of defects. The present invention tested spot mura using four different network models trained on them. The specific results are shown in Table 1. Among them, the Map value of the original YOLOV4-Tiny network is as low as 0.8438, which shows that the network performance is the worst. In actual detection, the lowest detection accuracy of the network is 0.94. After adding a Ghost module in front of the two prediction heads yolo head, the detection performance of YOLOV4-Tiny has been improved. The Map value has reached 0.8957 and the detection accuracy has reached 0.96, which proves that one of my improvements is effective. In addition, YOLOV4-Tiny adds a 76×76×128 effective feature layer on the basis of the original network to construct a feature pyramid. The Map value of the network has reached 0.8802 and the detection accuracy has reached 0.96, which is an improvement compared with the original network. This proves that the improvement is effective. Finally, we combine the two improvements to construct the final detection network. The final detection network structure is shown in the figure below. Figure 6 As shown in the figure, the experimental results show that our improved algorithm has the best performance. The Map value of the network reaches 0.9885 and the detection accuracy reaches 0.98. Compared with the original YOLOV4-tiny target detection network, the performance is greatly improved. Figure 9 This is a comparison chart of network AP0.5. The larger the area enclosed by the AP curve, the better the network performance. Figure 9 It can be seen that the performance of the improved YOLOV4-tiny target detection algorithm of the present invention is better than the original YOLOV4-tiny target detection algorithm. Figure 8 This is a comparison chart of the actual detection results of the improved YOLOV4-tiny target detection algorithm of the present invention and the original YOLOV4-tiny target detection algorithm. From the third and last rows of the figure, it can be clearly seen that the detection accuracy of the algorithm of the present invention is significantly improved compared with the original algorithm. Various experimental results show that the improvement of the algorithm by the present invention is effective.

[0086] Table 1 Comparison of experimental results

[0087] Network Model Missed detection False detection Detection accuracy Map0.5 YOLOV4-Tiny original network 6(100) 0(100) 94％ 84.38％ YOLOV4-Tiny+Ghoust module 4(100) 0(100) 96％ 89.57％ YOLOV4-Tiny+ adds feature layer 4(100) 0(100) 96％ 88.02％ YOLOV4-Tiny+Add Feature Layer+Ghoust Module (this invention) 2(100) 0(100) 98％ 98.85％

[0088] The application of the above technical solution effectively solves the problem of mura defects in LCD screens in the current automated assembly line operation. The defects are small in size, low in contrast, irregular in shape, and appear in random locations, which makes detection difficult. The method of the present invention uses deep learning for analysis, outputs identification and sorting results, and combines it with an automatic sorting module to separate unqualified products, effectively improving the detection efficiency of LCD screen mura defects and also improving the detection accuracy.

[0089] The specific embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in this field without departing from the scope of the present invention.

Claims

1. A method for detecting and sorting defects in liquid crystal display screens, characterized in that: include: Turn on the LCD screen visual real-time detection and sorting system; The liquid crystal display screen placed on the self-centering clamping tray (2) is driven by the conveying module (3) to move toward the defect detection module (5), and the defect detection module (5) performs defect detection on the liquid crystal display screen on the self-centering clamping tray (2), and the automatic sorting module (4) is used to sort the liquid crystal display screens according to the recognition results of the defect detection module (5); The defect detection module (5) is used to detect defects on the liquid crystal display screen on the self-centering clamping tray (2), comprising: The photoelectric counter in the image acquisition module in the defect detection module (5) is used to count the pallets passing through and transmit the count number to the respective controllers. The controller drives the camera (75) to acquire an image of the liquid crystal display screen according to the trigger signal of the photoelectric sensor stop switch and the count number of the photoelectric counter, and performs defect detection on the liquid crystal display screen according to the acquired image. The controller drives the camera (75) to collect images of the liquid crystal display screen according to the trigger signal of the photoelectric sensor stop switch and the count number of the photoelectric counter, and performs defect detection on the liquid crystal display screen according to the collected images, specifically: If the defect detection components in the defect detection module (5) are a group, and for the current detection, the photoelectric counter counts to 1 and the photoelectric sensor stop switch is triggered, the controller drives the upper synchronous belt conveyor in the conveying module (3) to stop moving, drives the camera (75) to collect images of the liquid crystal display, and calls the frozen model obtained according to the liquid crystal display defect detection model to perform defect detection on the collected liquid crystal display image. When the detection is completed, the photoelectric counter is reset to zero, and the next detection is carried out until it is completed; If there are n groups of defect detection components in the defect detection module (5), for the current detection, if the maximum increment value of the n photoelectric counters is m and the photoelectric sensor stop switch is triggered, the mth controller drives the upper synchronous belt conveyor in the conveying module (3) to stop moving, and the first m controllers respectively drive the camera (75) to collect images of the liquid crystal display screen, and call the frozen model obtained according to the liquid crystal display screen defect detection model to perform defect detection on the collected liquid crystal display screen image. When the current detection is completed, all photoelectric counters are reset to zero, and the next detection is carried out until the end; wherein, the n photoelectric counters correspond to the n controllers one by one; The LCD defect detection model is based on the YOLOV4-Tiny model. The output of the first CSP module in the backbone feature extraction network is added as a shallow feature map. The original deep feature map is convolved once to adjust the number of channels and then divided into two outputs: one output is upsampled and then fused with the middle feature map. The fused result is upsampled once and then fused with the newly added shallow feature map. The obtained feature map is processed by a Ghost module once and then input into the YOLO head for classification prediction and regression prediction; the other output is processed by the Ghost module once and then input into the YOLO head for classification prediction and regression prediction.

2. A liquid crystal display screen defect detection and sorting production line using the liquid crystal display screen defect detection and sorting method according to claim 1, characterized in that: The invention comprises a self-centering clamping tray (2), a conveying module (3), an automatic sorting module (4), and a defect detection module (5); the self-centering clamping tray (2) is placed on the conveying module (3), the self-centering clamping tray (2) is driven to move by the conveying module (3), the defect detection module (5) performs defect detection on the liquid crystal display screen on the self-centering clamping tray (2), and the automatic sorting module (4) is used to sort the liquid crystal display screen according to the recognition result of the defect detection module (5).

3. The LCD defect detection and sorting production line according to claim 2, characterized in that: The utility model also includes a dust removal module (1), wherein the dust removal module (1) includes a dust removal roller (6), an air-blowing dust collector (7), a frame I (8), and a sprocket I (9); wherein the sprocket I (9) is installed at both ends of the roller shaft of the dust removal roller (6); the dust removal roller (6) and the air-blowing dust collector (7) are installed on the frame I (8) in a front-to-back manner, and the air-blowing dust collector (7) is close to the forward end of the conveying module (3) in the conveying direction; the sprocket I (9) is powered by the servo motor I (37) of the upper synchronous belt conveyor in the conveying module (3).

4. The LCD defect detection and sorting production line according to claim 2, characterized in that: The self-centering clamping tray (2) includes a tray (10) and a self-centering clamping mechanism; wherein the self-centering clamping mechanism is installed on the tray (10), and the self-centering clamping mechanism includes a transverse retractable rack (16), a longitudinal retractable rack (17), a buckle (24), and a trigger device (25); wherein the buckle (24) is installed at one end of the transverse retractable rack (16) and the longitudinal retractable rack (17), and the other end of the transverse retractable rack (16) and the longitudinal retractable rack (17) cooperates with the trigger device (25) to achieve clamping of the liquid crystal display screen between the buckles (24).

5. The LCD defect detection and sorting production line according to claim 4, characterized in that: The trigger device (25) comprises a trigger cover (18), a gear (15), a one-way ratchet mechanism, a disc coil spring (21), a fixed shaft (22), a return spring (92), and a fixed pin (95); the trigger cover (18) is mounted on one end of the fixed shaft (22) via the fixed pin (95); a strip-shaped through hole is provided on the cylindrical body of the trigger cover (18) for cooperating with the fixed pin (95), so that the trigger cover (18) can realize circumferential rotation and axial movement, and the fixed shaft (22) and the trigger cover (18) are connected to each other. ) is provided with a return spring (92), the entire one-way ratchet mechanism is fixed on the fixed shaft (22), the gear (15) is meshed with the transverse contraction rack (16) and the longitudinal contraction rack (17), and the one-way ratchet (19) of the one-way ratchet mechanism is fixedly connected to the gear (15), and when the gear (15) rotates, the one-way ratchet (19) of the one-way ratchet mechanism is driven to rotate together, one end of the disc-shaped coil spring (21) is installed in the fixed groove of the gear (15), and the other end of the disc-shaped coil spring (21) is fixed on the self-centering clamping mechanism; The one-way ratchet mechanism comprises a one-way ratchet (19), a pawl (20), a pawl mounting frame (94), and a spring II (96); wherein the pawl mounting frame (94) is mounted in the one-way ratchet (19), one end of the pawl mounting frame (94) is spline-matched with the fixed shaft (22) and is fastened by a fixed nut (93), the pawl (20) is mounted on the outer ring of the pawl mounting frame (94), a spring II (96) is arranged between the outer ring of the pawl mounting frame (94) and the pawl (20), and an insert arranged at one end of the trigger cover (18) is matched with the pawl (20).

6. The LCD defect detection and sorting production line according to claim 2, characterized in that: The conveying module (3) comprises an upper synchronous belt conveyor and a lower double-speed chain conveyor; The upper synchronous belt conveyor includes a first transmission module and a first adjustment module. The first transmission module is used to provide power-driven movement through a first power system, and the movement of the first transmission module drives the self-centering clamping tray (2) to follow the movement; the first adjustment module is used to drive the nut block I (41) installed on the adjusting screw rod I (40) to move up and down through the adjusting screw rod I (40), and the up and down movement of the nut block I (41) drives the angle of the two first support arms (39) connected to the nut block I (41) to change, and the change of the angle of the first support arm (39) drives the first transmission module at the other end of the two first support arms (39) to move in the opposite direction with the first support arm (39), and the movement direction is perpendicular to the transportation direction of the first transmission module; The lower-layer double-speed chain conveyor is installed at the lower part of the upper-layer synchronous belt conveyor, and includes a second conveying module and a second adjusting module. The second conveying module is used to provide power-driven movement through the second power system, and the self-centering clamping tray (2) sorted by the automatic sorting module (4) is driven to follow the movement through the movement of the second conveying module; the second adjusting module is used to drive the nut block II (52) installed on the adjusting screw rod II (53) to move up and down through the adjusting screw rod II (53), and the angle of the two second support arms (51) connected to the nut block II (52) is driven to change through the up and down movement of the nut block II (52), and the second conveying module at the other end of the two second support arms (51) is driven to move in the opposite direction with the first support arm (39) through the change of the angle of the second support arms (51), and the movement direction is perpendicular to the transportation direction of the second conveying module.

7. The LCD defect detection and sorting production line according to claim 2, characterized in that: The defect detection module (5) includes a photoelectric sensor stop switch and one or more groups of defect detection components. Each group of defect detection components has the same structure and includes an image acquisition module and a controller. The photoelectric counter in each group of image acquisition modules is used to count the pallets passing through and transmit the count number to the respective controllers. The controller drives the camera (75) to acquire an image of the liquid crystal display screen based on the trigger signal of the photoelectric sensor stop switch and the count number of the photoelectric counter, and performs defect detection on the liquid crystal display screen based on the acquired image.

Images