Dried vegetable and fruit product identifying and sorting system and control method

Abstract

The invention belongs to the technical field of intelligent material sorting, and particularly relates to a dried vegetable and fruit product recognizing and sorting system and a control method, and the system comprises a vibration feeding mechanism, a conveying device, a photographing device, a horizontal joint mechanical arm, a weighing and makeup mechanism and a control device; according to the recognition result of the control device, the horizontal joint mechanical arm is controlled to suck the recognized dried vegetable and fruit products. The production efficiency of the dried vegetable and fruit products can be improved.

Description

Technical Field

[0001] The present application belongs to the technical field of intelligent material sorting, and specifically relates to a dried fruit and vegetable product identification and sorting system and a control method. Background Art

[0002] Dried fruit and vegetable products are dried foods made from fruits and vegetables through cooking, freezing and other techniques. In the sorting and bagging process of dried fruit and vegetable products, taking pumpkin slices as an example, it is required that when they are packaged into bags, the number of pumpkin slices in each bag is the same and the total weight has a fixed requirement, and no adhesion occurs. However, after slicing, cooking and refrigeration and other processing techniques, there will be a problem of multiple pumpkin slices sticking together, which affects the difficulty and efficiency of subsequent sorting and bagging; and in order to avoid the frozen pumpkin slices from melting and deforming, the sorting process cannot consume too much time. Therefore, the automatic breaking up and identification of pumpkin slices, and the realization of rapid sorting and packaging have become technical difficulties that limit the automated production of pumpkin slices. At present, there is no complete set of products on the market that can accurately sort out individual dried fruits and vegetables from the pile and combine them into fixed weights. The technology proposed in the present invention can make up for this deficiency. Summary of the invention

[0003] In order to solve the deficiencies of the prior art, the purpose of the present invention is to provide a dried fruit and vegetable product identification and sorting device and a control method thereof, so as to realize the automatic breaking up, identification and rapid sorting of dried fruit and vegetable products, and improve the production efficiency of dried fruit and vegetable products. The technical solution is as follows:

[0004] A dried fruit and vegetable product identification and sorting system includes a vibrating feeding mechanism, a conveying device, a photographing device, a horizontal joint mechanical arm, a weighing and counting mechanism, and a control device;

[0005] The vibrating feeding mechanism is used to output the material to the conveying device;

[0006] The horizontal joint mechanical arm adjusts its direction, cooperates with the end mechanism to pick up materials, and places them in the weighing and number-making mechanism;

[0007] The camera device is used to obtain images of materials on the conveyor device;

[0008] The weighing and numbering mechanism weighs the obtained materials, and when the weight of the materials reaches a set threshold, the control device pushes the materials out;

[0009] The control device is used to control the vibration feeding mechanism, the conveying device, the horizontal joint robot arm and the movement of the weighing and making-up mechanism as well as the image processing.

[0010] Preferably, the vibrating feeding mechanism includes a feeding tray, a vibrator and a primary conveyor belt; a vibrator is provided under the feeding tray, and the vibrator is fixed to the conveyor belt base through a feeding base; a breaking wheel is provided on the primary conveyor belt, and a plurality of groups of rubber soft rods with uniform gaps are provided on the breaking wheel.

[0011] Preferably, the conveying device includes a secondary conveyor belt, a scattering wheel is provided above the secondary conveyor belt, a plurality of rubber soft plates distributed at intervals are installed on the scattering wheel, and the scattering wheel is fixed above the second conveyor belt base.

[0012] Preferably, the horizontal joint robot arm is located on one side of the secondary conveyor belt of the conveying device, and the end of the horizontal joint robot arm has at least four degrees of freedom: translation along the X, Y, and Z axes and rotation around the Z axis; the end of the horizontal joint robot arm is also provided with an end mechanism, which includes a suction cup, an elastic buffer rod, a suction cup connecting plate and an air pipe; the suction cup is connected to the end of the elastic buffer rod, and the elastic buffer rod is fixed on the suction cup connecting plate; a spring is provided on the elastic buffer rod, and the suction cup connecting plate includes an L-shaped plate and a straight plate, and a through hole is formed between the two, which is used to clamp and fix the elastic buffer rod together with the end sleeve of the robot arm.

[0013] Preferably, the weighing and making up mechanism comprises a plurality of identical weighing units, each weighing unit comprises, from top to bottom, a discharge hopper, a buffer hopper, a weighing tray, a push slider, and a weighing sensor; the discharge hopper and the buffer hopper are both provided with a fan-shaped switch plate, which is opened and closed by a rotating cylinder, and the material falls from the discharge hopper to the buffer hopper below, and then falls to the weighing tray below. The weighing sensor below the weighing tray obtains the weight value of the material and feeds it back to the control device, which solves the optimal weight combination and controls the push slider to push out the group of materials.

[0014] Preferably, the control device includes a host computer and a slave computer, which are communicatively connected. The slave computer controls the movement of the vibration feeding mechanism, the conveying device, the horizontal joint robot arm, the end mechanism and the weighing and making-up mechanism; the host computer divides the area according to the image processing structure and controls the movement trajectory of the horizontal joint robot arm.

[0015] A method for identifying and sorting dried fruit and vegetable products comprises the following steps:

[0016] S100, acquiring an image of a working area on a conveying device;

[0017] S200, when it is identified that there is material to be identified in the working area, the conveying device is controlled to move forward a set distance and then stop;

[0018] S300, after the vibrating feeding mechanism and the conveying device stop, an image of the working area on the conveying device is acquired again to identify the material to be sucked in the image;

[0019] S400, control the horizontal joint robot arm to sequentially absorb the identified materials; sequentially place them in several weighing and numbering mechanisms, which feed back the weight values ​​to the host computer, and the host computer solves the optimal weight combination and controls the push slider to push out the materials;

[0020] S500, when the number of dried fruit and vegetable products in the working area is lower than the set threshold, the vibration feeding mechanism and the conveying device are controlled to move forward a set distance again and then stop, and steps S300-S500 are repeated.

[0021] Preferably, in step S200, the secondary conveyor belt of the conveying device carries the scattered materials and advances at a uniform speed, and is initially scattered and flattened by the scattering wheel, and the camera continuously acquires the image of the working area below. When the front material enters the edge of the working area, the camera captures the image with the first material target. At this time, the host computer recognizes the material target at the front side of the working area in the image, and controls the conveying device to advance a set distance and then stop; the set distance here is the length of the working area in the advancing direction of the secondary conveyor belt of the conveying device. When the secondary conveyor belt runs at a uniform speed, the advancing distance of the secondary conveyor belt can be controlled by controlling the advancing time of the secondary conveyor belt. When the front material on the secondary conveyor belt just covers the working area, the secondary conveyor belt stops advancing;

[0022] In step S300, based on the Yolov5 model, the image of the enhanced material on the conveyor belt with annotations is input, and the model is supervised trained. At the same time, the confidence threshold of the target frame is adjusted during the training, the anchor frame selection mechanism is improved, and the probability of overlapping frames and multiple target frames is reduced. The trained recognition model is obtained and deployed to the host computer; the camera collects the current frame image in real time and uploads it to the host computer. The host computer identifies the materials in the collected image of the working area, marks each target with a rectangular frame, and plans the coordinate sequence of all materials. The planned coordinate sequence forms the picking path of the robot arm;

[0023] In step S400, the upper computer plans the activity trajectory of the horizontal joint robot arm according to the weighing units at different positions of the weighing and making up mechanism, and numbers the rotary cylinders corresponding to the discharge hopper and the buffer hopper; when the horizontal joint robot arm places the material into the set discharge hopper of the weighing and making up mechanism, the lower computer controls the rotary cylinder of the discharge hopper to open the fan-shaped switch plate at the bottom of the discharge hopper, and the material falls into the buffer hopper of the lower layer. According to the set time, the upper discharge hopper closes the fan-shaped switch plate, and the lower buffer hopper opens its fan-shaped switch plate, and the material falls into the material bag on the tray again;

[0024] The horizontal joint robot arm uses a "door"-shaped trajectory when picking and placing. The fixed-ratio Lam′e curve is used for optimization at the turning points of vertical height and lateral movement, and then the horizontal straight line segment is eliminated and directly replaced by the Lam′e curve.

[0025] Preferably, the communication mechanism adopts a mutual feedback mechanism of the "hook principle", specifically: for the first hook, the camera acquires an image, the upper computer recognizes the presence of material, the first conveyor belt and the second conveyor belt stop after advancing a set distance, the upper computer obtains the stop information of the motors of the first conveyor belt and the second conveyor belt, and the camera takes a picture of a still and stable picture at this time, and obtains accurate information of all targets in the picture. When all targets in the picture are picked up by the horizontal joint mechanical arm, the camera shoots and recognizes that there are no targets in the picture, and feedback is given to the upper computer, and the conveyor belt continues to advance to the next working distance; for the second hook, a single material is picked up by the mechanical arm and placed in the hopper of each weighing unit. In the initial state of weight addition, each weighing unit contains one material, and the material is continuously put into the weighing unit. When it is calculated that the material weight combination of a certain weighing unit meets the requirements, the lower computer obtains a signal that the material on the weighing tray is pushed out, and controls the action of the push slider; the buffer hoppers of other units are opened to replenish the weighing tray below;

[0026] The third time the hand is hooked, during the material discharge process, the discharge hopper and the buffer hopper have opening and closing signals, and the suction cup exhales during the material discharge. The upper computer knows that the horizontal joint robotic arm has put down the picked up material, and the upper computer can control the horizontal joint robotic arm to absorb the next piece of material; when all the discharge hoppers do not generate opening and closing signals for material dropping, the horizontal joint robotic arm grabs the material and waits in the air above the weighing and making up the number mechanism.

[0027] Preferably, in the recognition of dried fruits and vegetables, the modified Yolov5 model is used for recognition, and the steps are as follows: Improve the expression form of the original IoU:

[0028]

[0029] in:

[0030] λ1,λ2,λ3: dynamic weights, adjusting the geometric contribution weights of IoU according to the confidence; B and B gt The predicted box and the real box in the detection are respectively; b and b gt Defined as B and B gt Center of diag is the diagonal length of the smallest rectangle that covers both boxes;

[0031] d and d gt It's B and B gt Vector representation of the diagonal; |d| and |d gt | represents d and d gt Length;

[0032] b sc_max and b i Represent the maximum confidence score box and other redundant boxes respectively;

[0033] ρ(b,b gt ): The Euclidean distance between the center of the predicted box and the true box, indicating the center offset;

[0034] ρ(d,d gt ): diagonal vector difference, measuring the shape difference of the box;

[0035] θ: Diagonal vector angular difference, used to capture the directional changes of the frame;

[0036] Improved loss function Loss DR for:

[0037] Loss DR =1-DR_IoU

[0038]

[0039] in:

[0040] IoU term: represents the intersection-over-union ratio between the predicted box and the real box:

[0041]

[0042] Center shift penalty:

[0043] is the offset distance of the frame center;

[0044] C diag is the minimum bounding box diagonal length, used to normalize the offset distance;

[0045] Diagonal Difference Term:

[0046] represents the diagonal vector difference, |d||d gt |The product of the diagonal lengths is used to represent the normalized weight, measuring the similarity of the box sizes;

[0047] Direction angle difference term:

[0048] sin 2 θ represents the square of the angle difference, normalizing the orientation difference of the weighted box;

[0049] Dynamic weight adjustment:

[0050] λ1,λ2,λ3: weight factors, dynamically adjusted to adapt to different data sets and detection tasks;

[0051]

[0052] Improved dense region suppression function DR_NMS: It comprehensively considers the overlapping area ratio of the box, the diagonal direction vector of the box, and the confidence, defines a direction-aware IoU suppression strategy, and designs the similarity score S dir , to improve the accuracy of dense target detection:

[0053] S dir =αR area +βR diag +γR angle

[0054] α+β+γ=1

[0055] in:

[0056] Area ratio, a measure of the similarity of the areas of two boxes:

[0057]

[0058] Diagonal ratio, which measures the similarity of the diagonal lengths of boxes:

[0059]

[0060] The direction angle ratio reflects the consistency of the frame direction through the ratio of the cosine values ​​of the direction angles:

[0061]

[0062] α, β, and γ are the weights of these three ratios, and their sum is 1;

[0063] The direction-aware similarity S dir Combined with the improved DR_IoU, it is used to dynamically adjust the final suppression of the box. The final IoU after DR_NMS adjustment is:

[0064] Adjusted-IoU=DR_IoU·S dir .

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] The overall structure of the device is mainly composed of horizontal joint robotic arms, which are suitable for high-speed sorting of materials (pumpkin slices and other dried fruit and vegetable products) on the conveyor belt plane. At the same time, the end of the robotic arm is designed with an elastic buffer shaft and a flexible suction cup to avoid collision in high-speed picking, ensuring that the pumpkin slices are not damaged; the conveyor belt is designed with a scattering wheel, which can realize the automatic scattering of pumpkin slices for subsequent suction by the suction cup. When the suction cup sucks the pumpkin slices, a pause grabbing method is adopted. The conveyor belt with a large number of pumpkin slices will pause when it reaches the working area of ​​the camera. The camera only needs to take a picture once to identify all the targets in the current working area at once, and then plan all the paths to guide the robotic arm to pick up and place all the targets in order. After all the targets are picked up, the conveyor belt continues to move forward one working area and starts working in the next area, ensuring the quality of the picked pumpkin slices. In the path planning of the robotic arm, the arc-optimized "door"-shaped trajectory is designed, which greatly improves the picking speed and stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 A perspective view of the present invention;

[0068] Figure 2 It is a front view of the present invention;

[0069] Figure 3 It is a left side view of the present invention;

[0070] Figure 4 A top view of the present invention;

[0071] Figure 5 It is a partial schematic diagram of the terminal mechanism;

[0072] Figure 6 It is a schematic diagram of a scattering wheel and a leveling wheel;

[0073] Figure 7 A schematic diagram of the local effect of identifying dried fruit and vegetable products in the working area;

[0074] Figure 8 Schematic diagram of the movement trajectory of the end of the robotic arm.

[0075] Among them, 1. Vibration feeding mechanism; 2. Feeding plate; 3. Vibrator; 4. Feeding base; 5. Control cabinet; 6. Conveyor belt base 1; 7. Bottom plate; 8. Industrial camera; 9. End sleeve of robot arm; 10. End mechanism; 11. Horizontal joint robot arm; 12. Camera bracket; 13. Make up and weighing mechanism; 14. Robot arm base; 15. Base fixing plate; 16. Breaking wheel; 17. Discharging hopper; 18. Buffer hopper; 1 9. Weighing tray; 20. Pushing slide block; 21. Weighing sensor; 22. Primary conveyor belt; 23. Breaking wheel motor; 24. Leveling wheel; 25. Secondary conveyor belt; 26. Leveling wheel motor; 27. Robot arm working range track; 28. Suction cup connecting plate; 29. ​​Elastic buffer rod; 30. Suction cup; 31. Rubber soft column; 32. Leveling soft plate; 33. Fan-shaped switch plate; 34. Conveyor belt base II; 35. Weighing fixed plate. DETAILED DESCRIPTION

[0076] In order to facilitate the understanding of the present invention, the present invention is further described below through embodiments in conjunction with the accompanying drawings, but the present invention is not limited thereto and is not intended to be limited thereto.

[0077] The embodiment of the present invention, as Figure 1-Figure 7 As shown, firstly, a device for identifying and sorting dried fruit and vegetable products is provided, comprising:

[0078] The vibrating feeding mechanism 1 includes a feeding plate 2 and a vibrator 3, which is used to drop the piled dried fruits and vegetables onto the primary conveyor belt 22 in a relatively uniform manner. The vibrating feeding mechanism 1 is fixed to the conveyor belt base 6 through a feeding base 4 and is located above the primary conveyor belt 22. The conveyor belt base 6 is fixed to the bottom plate 7.

[0079] The primary conveyor belt 22 is provided with a scattering wheel 16 , and the scattering wheel 16 is designed with multiple groups of rubber soft rods 31 with uniform gaps, which can scatter the piled dried fruits and vegetables without damaging them.

[0080] A leveling wheel 24 and a leveling wheel motor 26 are arranged above the secondary conveyor belt 25. Two rubber soft plates 32 are installed on the leveling wheel 24, which can level the dried fruits and vegetables that still have individual vertical spirals after being scattered into a layer; the leveling wheel 24 is fixed on the conveyor belt base 34, and is fixed to the bottom plate 7 through the conveyor belt base 34.

[0081] The primary conveyor belt 22 is located above the secondary conveyor belt 25, and the ends of the primary conveyor belt 22 and the secondary conveyor belt 25 are arranged to overlap each other.

[0082] The camera mechanism includes an industrial camera 8 and a camera bracket 12. The industrial camera 8 is located above the articulated mechanical arm 11 and is fixedly arranged toward the conveyor belt to obtain images of the scattered dried fruit and vegetable products on the conveyor belt.

[0083] The horizontal articulated robot arm 11 is located on one side of the secondary conveyor belt 25 and is fixed on the robot arm base 14. The robot arm base 14 is fixed on the bottom plate 7 through the base fixing plate 15. The end of the articulated robot arm 11 has at least four degrees of freedom: translation along the X, Y, and Z axes and rotation around the Z axis. The end of the articulated robot arm 11 is also provided with an end mechanism 10. The end mechanism 10 includes a suction cup 30 for sucking dried fruit and vegetable products on the conveyor belt. The suction cup 30 is connected to the end of the articulated robot arm through an elastic buffer shaft 29. The elastic buffer shaft 29 is fixed to the end sleeve 9 of the robot arm through an end connecting plate 28. The end connecting plate 28 is divided into a straight plate and an L-shaped plate.

[0084] The weighing and making up mechanism 13 is located on the same side of the horizontal joint mechanical arm 11 and is fixed to the mechanical arm base 14 through a weighing fixing plate 35. The weighing and making up mechanism 13 includes four identical weighing units, each of which includes a discharge hopper 17, a buffer hopper 18, a weighing tray 19, a pusher slide 20, and a weighing sensor 21; the discharge hopper 17 and the buffer hopper 18 are equipped with a fan-shaped switch plate 33, which is opened and closed by a rotating cylinder, and the dried fruits and vegetables can fall onto the buffer hopper below; with the same design, the dried fruits and vegetables in the buffer hopper can fall onto the weighing tray below, and the weighing sensor below the weighing tray can quickly obtain the weight value of the dried fruits and vegetables and feed it back to the host computer. The host computer solves the optimal weight combination and controls the pusher slide to push out the dried fruits and vegetables through the lower computer.

[0085] The control device includes an upper computer and a lower computer. The lower computer is located in the control cabinet 5. The upper computer adopts a computer. The computer is used to identify each dried fruit and vegetable product in the image obtained by the industrial camera and to run the weight combination calculation of the four hoppers to determine the tray number of the pushed material. The lower computer is used to control the horizontal joint mechanical arm to absorb each identified dried fruit and vegetable product according to the recognition result of the computer and to control the opening and closing of the cylinder of the hopper and the push slide of the numbering and weighing mechanism.

[0086] During the production and processing of pumpkin slices and other dried fruit and vegetable products, after slicing, cooking and refrigeration, pumpkin slices of different shapes, sizes and disorderly stacking will be conveyed to the sorting and packaging station via a conveyor belt. The traditional way of sorting and packaging pumpkin slices is to manually break apart the sticky pumpkin slices, and manually count and take out the required number of individuals and put them into the packaging bag, which has the disadvantages of low efficiency and poor sanitary conditions. To this end, this embodiment provides a dried fruit and vegetable product identification and sorting device, which automatically breaks up the sticky pumpkin slices during the conveyor belt transportation process by setting a breaking wheel on the conveyor belt, and then the sucker on the horizontal joint robot arm sucks the broken pumpkin slices and packs them.

[0087] Specifically, the joint mechanical arm 11 in this embodiment is a four-degree-of-freedom mechanical arm, and the end has four degrees of freedom along the X, Y, and Z axes and four degrees of freedom of rotation around the Z axis, so as to grab pumpkin slices at different positions in the working area, and can quickly rise and fall along the Z axis to meet the rhythm of high-speed pick-up and placement. The joint mechanical arm is connected to the mechanical arm base 11 by bolts, and the mechanical arm has a guide hole connected to the air path. The end of the mechanical arm is provided with a mechanical arm end sleeve 9, which is a hollow sleeve for installing a suction cup and an air pipe. The end mechanism 10 includes a suction cup 30, an elastic buffer rod 29, a suction cup connecting plate 28 and an air pipe. The suction cup 30 is a food-grade skirt film suction cup made of silicone material, which is characterized by being able to be closely attached to the surface of irregular pumpkin slices, thereby stably sucking and not falling. The suction cup 30 is threadedly connected to the end of the elastic buffer rod 29, and the elastic buffer rod is fixed to the suction cup connecting plate 28 by bolts. The elastic buffer rod is provided with a spring, which can have a certain buffer floating range to avoid excessive pressure on the pumpkin stem from the end. The suction cup connecting plate 28 includes an L-shaped plate and a straight plate, and the two plates are combined to form a through hole in the middle, which is used to clamp and fix the elastic buffer shaft with the end sleeve of the mechanical arm.

[0088] The suction cup 30 is connected to the pneumatic component, and the pumpkin stem is sucked under the control of the pneumatic component. The pneumatic component includes an air filter, a vacuum generator, and an air pump, which are sequentially connected to the suction cup and the rear of the elastic buffer shaft with an air pipe as a connector. The air pipe on the terminal mechanism is connected to the upper end of the elastic buffer shaft, and is connected to the air filter after passing through the end sleeve of the robot arm. The air filter is connected to the vacuum generator through the air pipe, and then connected to the air pump through the air pipe. The air pump serves as an air source to supply air to the entire device. The vacuum generator can generate positive pressure and negative pressure, and can be controlled by a solenoid valve to achieve rapid switching of positive and negative pressure. Negative pressure is generated when sucking pumpkin slices to suck the target, and positive pressure is generated when placing pumpkin slices to make the pumpkin slices fall quickly.

[0089] The secondary conveyor belt 25 is provided with a scattering wheel 16 and a scattering wheel motor 23. The scattering wheel 16 is connected to the scattering wheel motor 23. Driven by the scattering wheel motor, the pumpkin slices on the secondary conveyor belt are continuously moved, thereby playing the role of initially flattening the piled pumpkin slices. For the stuck pumpkin slices, the algorithm will exclude them after identification. At the subsequent position, personnel will pick them out and break them apart, and put them back to the beginning of the production line and sorted again by the horizontal joint robot arm 11.

[0090] The surfaces of the secondary conveyor belt 25 and the primary conveyor belt 22 are green PVC materials, which are used to transport the scattered and disordered pumpkin slices after processing techniques such as slicing, cooking and refrigeration.

[0091] The camera mechanism includes an industrial camera 7 and a camera support frame 8. The industrial camera 7 faces the central area of ​​the conveyor belt below to obtain real-time images of pumpkin slices on the conveyor belt; the bottom of the camera support frame 8 is fixed on the floor, and the industrial camera 7 is fixed on the top. In some embodiments, the height of the industrial camera is 700 mm higher than the plane of the workbench.

[0092] A plurality of weighing units are also provided on the base 11 of the horizontal joint robot arm. In this embodiment, the number of weighing units is four. The plurality of weighing units are installed in parallel on the side of the base of the horizontal joint robot arm and fixed on the base 14 of the robot arm. Each time the end of the horizontal joint robot arm places a pumpkin slice on a weighing unit, the control device will obtain the coordinates of the robot arm at this time and the signal of the solenoid valve controlling the suction cup to exhale. The control device marks the number of material slices in each weighing unit according to these two signals. The fruit and vegetable slices of appropriate weight fall onto the gathering production line.

[0093] The hardware of the control device includes a control cabinet 5 and a computer. The control cabinet includes a robot arm motion control module, a conveyor belt motion control module and a suction cup control module. The computer is a GPU computer.

[0094] The GPU computer is used to process images taken by industrial cameras, run recognition algorithms, calculate coordinate transformation relationships, and send the coordinate information in the processing results to the horizontal joint robot arm motion control module through TCP communication.

[0095] The horizontal joint robot arm motion control module is used to control the start and stop, speed adjustment and path planning of the joint robot arm. In this embodiment, the horizontal joint robot arm 11 is a four-axis robot arm, so the robot arm motion control module can control the rotation angles J1, J2, J3, and J4 of the four axes of the robot arm separately, or jointly control the linear motion of the end of the horizontal joint robot arm along the X, Y, and Z axes and the rotational motion around the Z axis.

[0096] The primary conveyor belt and the secondary conveyor belt motion control module are used to receive signals from the lower computer and control the rapid start and stop and uniform speed advance of the primary conveyor belt and the secondary conveyor belt.

[0097] The suction cup control module is used to quickly control the suction and blowing of the end suction cup by controlling the on and off of the solenoid valve of the vacuum generator, corresponding to the picking and placing of the targets respectively.

[0098] When the pumpkin slices move to the working area below the industrial camera through the conveyor belt, the industrial camera will obtain the current frame image and send it to the host computer, that is, the GPU computer. The host computer extracts the pumpkin slice target information of the current frame based on the recognition algorithm, and then obtains the position information of the target required by the robot arm through mathematical derivation and inverse solution, and finally sends it to the control cabinet through socket communication to control the robot arm to absorb the pumpkin slice target in sequence according to the planned path and point information. The camera obtains each frame of the image in real time and processes it according to the above process. The horizontal joint robot arm 11 will pick up and place the pumpkin slices in the working area on the conveyor belt at high speed. Due to the large number and density of pumpkin slices on the conveyor belt in the required industrial scenario, the dynamic grasping of grabbing on the go cannot pick up all the pumpkin slices in the current working area. The remaining pumpkin slices will flow forward with the faster secondary conveyor belt. Since the pumpkin slices are in a frozen state, the pumpkin slices that flow away will thaw and melt in the next cycle, reducing the quality. Therefore, the present invention adopts a pause grabbing method. The secondary conveyor belt with a large number of pumpkin slices will pause when it reaches the working area. The camera only needs to take a picture once, identify all the targets in the current area at one time, and send all the coordinate sequences to the lower computer of the control cabinet. After the lower computer plans all the paths, the horizontal joint controls the mechanical arm to pick up and place all the targets in order. After all the targets are grabbed, the upper computer recognizes that the number of targets is zero, sends a signal, and the conveyor belt continues to move forward one working area to start working in the next area.

[0099] The lower computer controls the start and stop of the conveyor belt through the conveyor belt start and stop control device. Figure 6 As shown, in this embodiment, the conveyor belt motion control module in the lower computer is a programmable controller PLC, and the output end of the PLC acts on the coil of the relay KA4 to control the on and off of the normally open end of the relay KA4, thereby controlling the start and stop of the conveyor belt motor 10. When the PLC controls the relay KA4 coil to be energized, the normally open end of the relay KA4 is closed, the motor rotates forward, and the conveyor belt moves forward at a constant speed; when the relay KA4 coil is de-energized, the normally open end of the relay KA4 is disconnected, the motor stops, and the conveyor belt stops running.

[0100] After receiving the image of the working area captured by the industrial camera, the host computer identifies and annotates the dried fruit and vegetable products in the image. A dried fruit and vegetable product recognition model is built based on the deep learning algorithm, and each dried fruit product in the image is identified and framed based on the dried fruit and vegetable product recognition model.

[0101] Taking pumpkin slices as an example, 5,000 pictures of pumpkin slices on a conveyor belt were collected in advance to create a data set, including images taken at different angles, different lighting, and different distances. Data enhancement methods such as inversion, cropping, splicing, and color change were used to expand the final data set to 20,000. Make the learned images richer and more complex than the actual situation, further improving the recognition accuracy. Subsequently, after manually annotating the image samples containing various situations, the deep learning intelligent algorithm was used, based on the Yolov5 model, and the image data set was input for training to learn the case semantic information of edge contact. Since the maximum suppression NMS used by the model is not suitable for the dense small target detection method of dried fruits and vegetables in this scheme, the mechanism of NMS is to sort the anchor frame scores in the score set S by high and low, traverse and calculate the overlapping area ratio of each anchor frame and the highest-scoring candidate frame, and directly remove the anchor frame if it is greater than the threshold a. This method may cause the detection of targets in the overlapping area to fail; at the same time, if a is too small, adjacent targets may be missed, and if a is too large, a target may have multiple repeated detection frames. Therefore, based on this, the anchor frame selection mechanism was improved to target the distribution characteristics of dried fruits and vegetables on the conveyor belt, which are numerous and closely spaced. The SoftNMS algorithm was used to select anchor frames to reduce the probability of overlapping frames and frames containing multiple targets. When the overlapping area ratio of two anchor frames is greater than a, the score will not be directly reset to zero, but will be reduced. The higher the overlapping area ratio, the more the score will be reduced. The trained recognition model can intelligently identify and select a single target from the closely packed pumpkin slices.

[0102] According to the action characteristics of the device sucking up the pumpkin slices from the flat position and placing them in the weighing unit, this implementation uses a "door"-shaped trajectory ( Figure 8 ), and add arc interpolation to optimize the corners to improve smoothness, reduce vibration and travel time, such as Figure 5 As shown. At the turning points of vertical height and lateral movement, we use the fixed-ratio Lam'e curve for optimization, and then cancel the horizontal straight line segment and directly use the Lam'e curve instead. This design can reduce the acceleration and deceleration during lifting-translation-landing, greatly reducing the time spent on the reciprocating pick-and-place journey and reducing jitter. In actual industrial production, 40 pick-and-place operations can be performed per minute, meeting the needs of industrial scenarios.

[0103] The embodiment of the present invention, based on the above-mentioned dried fruit and vegetable product identification and sorting device, further provides a control method of the identification and sorting device, including:

[0104] S100, acquiring an image of the working area on the secondary conveyor belt;

[0105] S200, when it is identified that there are dried fruit and vegetable products to be identified in the working area, the secondary conveyor belt is controlled to move forward a set distance and then stop;

[0106] S300, after the secondary conveyor belt stops, an image of the working area is acquired again to identify the dried fruit and vegetable products to be sucked in the image;

[0107] S400, controlling the joint robotic arm to sequentially absorb the identified dried fruit and vegetable products;

[0108] S500, when the amount of dried fruit and vegetable products in the working area is lower than the set threshold, the secondary conveyor belt is controlled to move forward a set distance again and then stop, and steps S300-S500 are repeated.

[0109] In step S100, the working area is the field of view of the industrial camera on the conveyor belt. In this embodiment, the industrial camera is set 700 mm above the conveyor belt, and the working range is a semicircular area with a diameter of 550 mm.

[0110] The working area of ​​the horizontal joint robot is a circle, but the conveyor belt is on the right side of the robot arm. The picking working area is the overlap of the semicircular area on the conveyor belt and the rectangular area of ​​the camera, and the placement area is the semicircular area on the left side of the conveyor belt. Figure 4 shown.

[0111] In step S200, the secondary conveyor belt carries the scattered pumpkin slices forward at a constant speed, and is initially scattered and flattened by the scattering wheel. The industrial camera keeps acquiring the image of the working area below. When the front pumpkin slice enters the edge of the working area, the camera captures the image with the first pumpkin slice target. At this time, the host computer recognizes the pumpkin slice target at the front of the working area in the image, and controls the secondary conveyor belt to move forward for a set distance before stopping. The set distance here is the length of the working area in the forward direction of the secondary conveyor belt. When the secondary conveyor belt runs at a constant speed, the forward distance of the secondary conveyor belt can be controlled by controlling the forward time of the secondary conveyor belt. In this embodiment, after the secondary conveyor belt is controlled to move forward for another 0.5 seconds, the front pumpkin slice, that is, the pumpkin slice identified first, just moves to the back of the working area. At this time, the front pumpkin slice on the secondary conveyor belt just covers the working area, and the secondary conveyor belt stops moving forward.

[0112] In step S300, a dried fruit and vegetable product recognition model is constructed based on a deep learning algorithm, and each dried fruit and vegetable product in the image is identified and framed based on the dried fruit and vegetable product recognition model. Taking the dried fruit and vegetable product as pumpkin slices as an example, the dried fruit and vegetable product recognition model is the pumpkin slice recognition model, which is constructed based on the Yolov5 model. The image of the enhanced pumpkin slices on the conveyor belt with annotations is input, and the model is supervised for training. At the same time, the confidence threshold of the target frame is adjusted during training, and the NMS algorithm selects the optimal frame to reduce the probability of overlapping frames and containing multiple target frames. The trained pumpkin slice recognition model is obtained and deployed to the host computer. The industrial camera collects the current frame image in real time and uploads it to the host computer. The host computer identifies each pumpkin slice in the image of the collected working area, marks each target with a rectangular frame, and plans the coordinate sequence of all pumpkin slices. The planned coordinate sequence forms the robot arm picking path.

[0113] In step S400, the horizontal joint robot arm motion control module picks up each pumpkin slice in turn according to the robot arm picking path planned by the host computer. During the picking process, the motion trajectory of the robot arm is a gate-shaped trajectory, and the circular arc interpolation method is used at the top turning point of the gate-shaped trajectory to optimize the trajectory to improve the smoothness of the robot arm movement. The robot arm puts the pumpkin slices into four weighing units in order. After the picking is completed, the industrial camera acquires the current frame image again to identify whether there are any remaining pumpkin slices. When the remaining number is less than the set threshold, it is qualified. Otherwise, the above algorithm is repeated to continue picking until all pumpkin slices in the working area are picked up.

[0114] In step S500, after all pumpkin slices in the current working area are picked up, the host computer issues a command to control the conveyor belt to continue to move forward at a constant speed for 0.5 seconds to obtain a new working area range, so as to sort and pick up the pumpkin slices in the next working area. Repeat steps S300-S500 to achieve continuous picking of pumpkin slices on the conveyor belt.

[0115] After step S500 is completed, it also includes: S600, when no dried fruit and vegetable products are identified in the working area within the set time, a completion signal is generated to prompt to replenish materials or shut down the device. If the host computer does not identify pumpkin slices in the image of the conveyor belt for 3 consecutive seconds, that is, the range of 6 working areas, it is considered that the pumpkin slices on the conveyor belt have been sorted and a completion signal is generated. In the identification of dried fruits and vegetables, since dried fruits and vegetables usually have highly dense arrangement characteristics, and their shapes have certain directionality and size differences, the original IoU and NMS algorithms in the Yolov5 model are prone to overlap and confusion, resulting in reduced detection accuracy. For the application scenario of dense dried fruits and vegetables recognition, we proposed Dense Region IoU (DR_IoU), Dense RegionNMS (DR_NMS) and the corresponding loss function Loss DR_IoU . To better cope with the detection problem of dense targets, especially in the detection task of dried fruits and vegetables with complex arrangement and directional differences, the final algorithm model can accurately identify and select a single target from closely packed materials. The design mainly includes the following four contents:

[0116] 1. Dense target separation: Reduce false detection and missed detection of overlapping targets, suitable for situations where dried fruits and vegetables are closely arranged.

[0117] 2. Shape and orientation differences: Consider that dried fruits and vegetables may have orientation or shape differences (such as irregular slice shapes).

[0118] 3. Dynamic weight adjustment: Dynamically adjust the weight of the detection algorithm according to the confidence difference to make high-confidence targets more prominent.

[0119] 4. More precise loss optimization: optimize box position, shape, orientation and size simultaneously in the regression stage.

[0120] The formula is as follows:

[0121] (i) DR_IoU: It improves the original IoU expression by dynamically adjusting the contribution weight of IoU, considering the geometric similarity characteristics of the candidate boxes (area, diagonal length ratio and angle difference) and the confidence of the box.

[0122]

[0123] in:

[0124] λ1,λ2,λ3: dynamic weights, adjust the geometric contribution weights of IoU according to the confidence.

[0125] B and B gt The predicted box and the real box in the detection are respectively b and b gt Defined as B and B gt The center of Cdiag is the diagonal length of the smallest rectangle that covers both boxes.

[0126] d and d gt It's B and B gt Vector representation of the diagonal. |d| and |d gt | represents d and d gt Length.

[0127] b sc_max and b i Represent the maximum confidence score box and other redundant boxes respectively.

[0128] ρ(b,b gt ): The Euclidean distance between the center of the predicted box and the true box, indicating the center offset.

[0129] ρ(d,d gt ): Diagonal vector difference, measuring the shape difference of the box.

[0130] θ: Difference in the direction angle of the diagonal vectors, used to capture changes in the orientation of the box.

[0131] (II) Propose an improved loss function Loss DR It can minimize the geometric differences between boxes, maximize the overlapping area, and penalize the center position offset, box size difference, and direction difference. It can adapt to highly densely arranged targets, capture differences in position, shape, and direction, and optimize the performance of the detection model in complex scenes.

[0132] Loss DR =1-DR_IoU

[0133]

[0134] in:

[0135] 1. IoU item: It represents the intersection over union (IoU) of the predicted box and the true box, and is used to measure the degree of regional overlap.

[0136]

[0137] 2. Center shift penalty:

[0138]

[0139] is the offset distance of the frame center.

[0140] c diag is the minimum bounding box diagonal length, used to normalize the offset distance.

[0141] 3. Diagonal difference terms:

[0142]

[0143] represents the diagonal vector difference, |D||d gt |The product of the diagonal lengths is used to represent the normalized weight, measuring the similarity of box sizes.

[0144] 4. Direction angle difference term:

[0145]

[0146] sin 2 θ represents the square of the angle difference, which normalizes the orientation difference of the weighted box.

[0147] 5. Dynamic adjustment of weights:

[0148] λ1,λ2,λ3: weight factors, dynamically adjusted to adapt to different data sets and detection tasks.

[0149]

[0150] (III) Improved dense region suppression function DR_NMS: Taking into account the overlapping area ratio of the box, the diagonal direction vector of the box, and the confidence, a direction-aware IoU suppression strategy is defined and the similarity score S is designed. dir , in order to improve the accuracy of dense target detection.

[0151] S dir =αr area +βr diag +γR angle

[0152] α+β+γ=1

[0153] in:

[0154] 1. Area ratio, which measures the similarity of the areas of two boxes.

[0155]

[0156] 2. Diagonal ratio, which measures the similarity of the diagonal lengths of boxes.

[0157]

[0158] 3. Direction angle ratio: reflects the consistency of the frame direction through the ratio of the cosine values ​​of the direction angles.

[0159]

[0160] 4 .α , β, and γ are the weights of these three ratios, and their sum is 1.

[0161] The direction-aware similarity S dir Combined with the improved DR_IoU, it is used to dynamically adjust the final suppression of the box. The final IoU after DR_NMS adjustment is:

[0162] Adjusted-IoU=DR_IoU·S dir

[0163] DR_NMS advantages:

[0164] 1. More accurate overlapping box suppression: Introducing direction-aware similarity to reduce false suppression between dense objects.

[0165] 2. Stronger adaptability: Combined with dynamic weight adjustment, it can adapt to different dense target scenarios.

[0166] 3. Improve recall rate: Improve detection frame integrity by moderately retaining low IoU frames

[0167] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A dried fruit and vegetable product identification and sorting system, characterized in that: It includes a vibrating feeding mechanism, a conveying device, a photographing device, a horizontal joint mechanical arm, a weighing and making up mechanism and a control device; The vibrating feeding mechanism is used to output the material to the conveying device; The horizontal joint mechanical arm adjusts its direction, cooperates with the end mechanism to pick up materials, and places them in the weighing and number-making mechanism; The camera device is used to obtain images of materials on the conveyor device; The weighing and making up the number mechanism weighs the obtained materials, and when the weight of the materials reaches a set threshold, the control device pushes the materials out; The control device is used to control the vibration feeding mechanism, the conveying device, the horizontal joint robot arm and the movement of the weighing and making-up mechanism as well as image processing.

2. A dried fruit and vegetable product identification and sorting system according to claim 1, characterized in that: The vibrating feeding mechanism includes a feeding tray, a vibrator and a primary conveyor belt; a vibrator is provided under the feeding tray, and the vibrator is fixed on the conveyor belt base through a feeding base; a breaking wheel is provided on the primary conveyor belt, and a plurality of groups of rubber soft rods with uniform gaps are provided on the breaking wheel.

3. A dried fruit and vegetable product identification and sorting system according to claim 1, characterized in that: The conveying device comprises a secondary conveyor belt, a scattering wheel is arranged above the secondary conveyor belt, a plurality of rubber soft plates distributed at intervals are installed on the scattering wheel, and the scattering wheel is fixed above the second conveyor belt base.

4. The dried fruit and vegetable product identification and sorting system according to claim 1, characterized in that: The horizontal joint robot arm is located on one side of the secondary conveyor belt of the conveying device. The end of the horizontal joint robot arm has at least four degrees of freedom: translation along the X, Y, and Z axes and rotation around the Z axis; the end of the horizontal joint robot arm is also provided with an end mechanism, which includes a suction cup, an elastic buffer rod, a suction cup connecting plate and an air pipe; the suction cup is connected to the end of the elastic buffer rod, and the elastic buffer rod is fixed on the suction cup connecting plate; a spring is provided on the elastic buffer rod, and the suction cup connecting plate includes an L-shaped plate and a straight plate, and a through hole is formed between the two, which is used to clamp and fix the elastic buffer rod together with the end sleeve of the robot arm.

5. The dried fruit and vegetable product identification and sorting system according to claim 1, characterized in that: The weighing and making up mechanism includes a plurality of identical weighing units, each weighing unit includes, from top to bottom, a discharge hopper, a buffer hopper, a weighing tray, a push slide block, and a weighing sensor; the discharge hopper and the buffer hopper are both provided with a fan-shaped switch plate, which is opened and closed by a rotating cylinder, and the material falls from the discharge hopper to the buffer hopper below, and then falls to the weighing tray below. The weighing sensor below the weighing tray obtains the weight value of the material and feeds it back to the control device, and the control device solves the optimal weight combination and controls the push slide block to push out the group of materials.

6. The dried fruit and vegetable product identification and sorting system according to claim 1, characterized in that: The control device includes an upper computer and a lower computer, which are communicatively connected. The lower computer controls the movement of the vibration feeding mechanism, the conveying device, the horizontal joint robot arm, the end mechanism and the weighing and making-up mechanism; the upper computer divides the area according to the image processing structure and controls the movement trajectory of the horizontal joint robot arm.

7. A dried fruit and vegetable product identification and sorting control method, configured in the dried fruit and vegetable product identification and sorting system according to any one of claims 1 to 6, characterized in that: The following steps are involved: S100, acquiring an image of a working area on a conveying device; S200, when it is identified that there is material to be identified in the working area, the conveying device is controlled to move forward a set distance and then stop; S300, after the vibrating feeding mechanism and the conveying device stop, an image of the working area on the conveying device is acquired again to identify the material to be sucked in the image; S400, control the horizontal joint robot arm to sequentially absorb the identified materials; sequentially place them in several weighing and numbering mechanisms, which feed back the weight values ​​to the host computer, and the host computer solves the optimal weight combination and controls the push slider to push out the materials; S500, when the number of dried fruit and vegetable products in the working area is lower than the set threshold, the vibration feeding mechanism and the conveying device are controlled to move forward a set distance again and then stop, and steps S300-S500 are repeated.

8. The method for identifying and sorting dried fruit and vegetable products according to claim 7, characterized in that: In step S200, the secondary conveyor belt of the conveying device carries the scattered materials and moves forward at a uniform speed, and is initially broken up and flattened by the breaking wheel. The camera continuously acquires the image of the working area below. When the front material enters the edge of the working area, the camera captures the image with the first material target. At this time, the host computer recognizes the material target at the front side of the working area in the image, and controls the conveying device to move forward a set distance and then stop. The set distance here is the length of the working area in the forward direction of the secondary conveyor belt of the conveying device. When the secondary conveyor belt runs at a uniform speed, the forward distance of the secondary conveyor belt can be controlled by controlling the forward time of the secondary conveyor belt. When the front material on the secondary conveyor belt just covers the working area, the secondary conveyor belt stops moving forward. In step S300, based on the Yolov5 model, the image of the enhanced material on the conveyor belt with annotations is input, and the model is supervised trained. At the same time, the confidence threshold of the target frame is adjusted during the training, the anchor frame selection mechanism is improved, and the probability of overlapping frames and multiple target frames is reduced. The trained recognition model is obtained and deployed to the host computer; the camera collects the current frame image in real time and uploads it to the host computer. The host computer identifies the materials in the collected image of the working area, marks each target with a rectangular frame, and plans the coordinate sequence of all materials. The planned coordinate sequence forms the picking path of the robot arm; In step S400, the upper computer plans the activity trajectory of the horizontal joint robot arm according to the weighing units at different positions of the weighing and making up mechanism, and numbers the rotary cylinders corresponding to the discharge hopper and the buffer hopper; when the horizontal joint robot arm places the material into the set discharge hopper of the weighing and making up mechanism, the lower computer controls the rotary cylinder of the discharge hopper to open the fan-shaped switch plate at the bottom of the discharge hopper, and the material falls into the buffer hopper of the lower layer. According to the set time, the upper discharge hopper closes the fan-shaped switch plate, and the lower buffer hopper opens its fan-shaped switch plate, and the material falls into the material bag on the tray again; The horizontal joint robot arm uses a "door"-shaped trajectory when picking and placing. The fixed-ratio Lam′e curve is used for optimization at the turning points of vertical height and lateral movement, and then the horizontal straight line segment is eliminated and directly replaced by the Lam′e curve.

9. A method for identifying and sorting dried fruit and vegetable products according to claim 8, characterized in that: The communication mechanism adopts the mutual feedback mechanism of the "hook principle", which is as follows: The first time the hand is hooked, the camera acquires an image, the host computer recognizes the presence of materials, the first and second conveyor belts advance the set distance and then stop, the host computer obtains the information that the motors of the first and second conveyor belts have stopped, at this time the camera takes a picture of a still and stable picture, and obtains accurate information about all the targets in the picture, when all the targets in the picture are picked up by the horizontal joint robot arm, the camera takes a picture to identify that there are no targets in the picture, and feedback is given to the host computer, and the conveyor belt continues to advance to the next working distance; In the second hook, a single material is picked up by the robot arm and placed in the hopper of each weighing unit. In the initial state of weight addition, each weighing unit contains one material, and materials are continuously put into the weighing unit. When the material weight combination of a certain weighing unit is calculated to meet the requirements, the lower computer receives the signal that the material on the weighing tray is pushed out, and controls the action of the push slider; the buffer hoppers of other units are opened to refill the weighing tray below; The third time the hand is hooked, during the material discharge process, the discharge hopper and the buffer hopper have opening and closing signals, and the suction cup exhales during the material discharge. The upper computer knows that the horizontal joint robotic arm has put down the picked up material, and the upper computer can control the horizontal joint robotic arm to absorb the next piece of material; when all the discharge hoppers do not generate opening and closing signals for material dropping, the horizontal joint robotic arm grabs the material and waits in the air above the weighing and making up the number mechanism.

10. A method for identifying and sorting dried fruit and vegetable products according to claim 9, characterized in that: In the recognition of dried fruits and vegetables, the modified Yolov5 model is used for recognition. The steps are as follows: Improve the original IoU expression: in: λ1,λ2,λ3: dynamic weights, adjust the geometric contribution weights of IoU according to the confidence; B and B gt The predicted box and the real box in the detection are respectively; b and b gt Defined as B and B gt Center of diag is the diagonal length of the smallest rectangle that covers both boxes; d and d gt It's B and B gt Vector representation of the diagonal; |d| and |d gt | represents d and d gt The length of b sc_max and b i Respectively represent the maximum confidence score box and other redundant boxes; ρ(b,b gt ): The Euclidean distance between the predicted box and the real box center, indicating the center offset; ρ(d,d gt ): diagonal vector difference, measuring the shape difference of the box; θ: Diagonal vector angular difference, used to capture the directional changes of the frame; Improved loss function Loss DR for: Loss DR =1-DR_IoU in: IoU term: represents the intersection-over-union ratio between the predicted box and the real box: Center shift penalty: is the offset distance of the frame center; c diag is the minimum bounding box diagonal length, used to normalize the offset distance; Diagonal Difference Term: represents the diagonal vector difference, |d||d gt |The product of the diagonal lengths is used to represent the normalized weight, measuring the similarity of the box sizes; Direction angle difference term: sin 2 θ represents the square of the angle difference, normalizing the orientation difference of the weighted box; Dynamic weight adjustment: λ1,λ2,λ3: weight factors, dynamically adjusted to adapt to different data sets and detection tasks; Improved dense region suppression function DR_NMS: It comprehensively considers the overlapping area ratio of the box, the diagonal direction vector of the box, and the confidence, defines a direction-aware IoU suppression strategy, and designs the similarity score S dir , to improve the accuracy of dense target detection: S dir =αR area +βR diag +γR angle α+β+γ=1 in: Area ratio, a measure of the similarity of the areas of two boxes: Diagonal ratio, which measures the similarity of the diagonal lengths of boxes: The direction angle ratio reflects the consistency of the frame direction through the ratio of the cosine values ​​of the direction angles: α, β, and γ are the weights of these three ratios, and their sum is 1; The direction-aware similarity S dir Combined with the improved DR_IoU, it is used to dynamically adjust the final suppression of the box. The final IoU after DR_NMS adjustment is: Adjusted-IoU=DR_IoU·S dir 。

Images