High-throughput microbe smart cloning and information analysis device and use method
By designing a high-throughput intelligent microbial cloning device, a fully automated microbial inoculation operation is achieved, solving the problems of low efficiency and contamination of existing equipment, improving inoculation accuracy and data recording capabilities, and making it suitable for scientific research and production needs.
Patent Information
- Application Number
- PCT/CN2024/124665
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2024-05-06
- Filing Date
- 2024-10-14
- Publication Date
- 2025-11-13
AI Technical Summary
Existing microbial inoculation equipment is inefficient, incomplete sterilization of inoculation needles leads to contamination, it cannot automatically record the status of microorganisms, and the proximity of the inoculation needle to the culture dish is difficult to control, which may damage the culture medium.
A high-throughput intelligent microbial cloning and information analysis device was designed, which includes a strain feeding, cap opening, photography, inoculation mechanism and a control and management system. It utilizes a liquid level detection and pressure detection system, combined with a DETR model, to automatically identify and record the strain status, realizing fully automated transport, cap opening, photography, inoculation, capping and sterilization operations.
It improves inoculation efficiency and accuracy, avoids bacterial contamination and damage, enables automated replacement of inoculation needles, and allows for data recording and multiple experimental analyses for each group of bacterial strains.
Smart Images

Figure CN2024124665_13112025_PF_FP_ABST
Abstract
Description
A high-throughput intelligent microbial cloning and information analysis device and its usage method Technical Field
[0001] This invention relates to the field of experimental equipment technology, and in particular to a high-throughput intelligent microbial cloning and information analysis device and its usage method. Background Technology
[0002] In microbial research, microbial inoculation is a common basic operation. The microorganisms used to produce strains are usually high-quality microorganisms that have been repeatedly screened by humans. Due to the small size and wide variety of microorganisms, inoculation needs to be repeated many times. In actual research, there are many difficulties, which require a lot of time and manpower. It also requires high technical skills from the operators. The strains must not be contaminated. During the inoculation process, the inoculation needle or loop must be sterilized at high temperature to prevent contamination. Incorrect inoculation operations can puncture the culture medium and damage the experimental samples. Moreover, the status information of the strains cannot be recorded in real time for each experiment, making it inconvenient to compare multiple strain experiments. This may lead to experimental duplication or make the experimental results difficult to analyze.
[0003] Currently, existing microbial inoculation equipment patents such as CN107118950A, CN107418888A, and CN115109687A all require multiple actions such as moving the cap to open and dip the inoculation plate when completing multiple inoculation tasks, resulting in relatively slow inoculation efficiency. Furthermore, the inoculation needle only undergoes sterilization after each inoculation, which may lead to low sterilization efficiency and contamination of other microbial strains. Additionally, the proximity of the needle to the culture dish cannot be measured during inoculation, potentially puncturing the culture medium and disrupting the experiment. Moreover, the inability to automatically record and analyze the current state of the microbial strains for each experiment may slow down subsequent research and analysis.
[0004] In conclusion, with the continuous development and improvement of automated equipment and rapid testing methods, only equipment that can complete inoculation work efficiently and accurately, automatically change inoculation needles, and record and analyze the status of each experimental strain can meet the needs of scientific research and production. Summary of the Invention
[0005] To address the aforementioned problems, this invention proposes a high-throughput intelligent microbial cloning and information analysis device and its usage method. This device can replace tedious and repetitive manual operations in microbial inoculation, accurately identify and record the status information of bacterial strains, automatically classify and compare bacterial strains from multiple experiments, avoid sample damage and non-standard operation, improve the accuracy and success rate of inoculation, and automatically replace the inoculation needle.
[0006] The objective of this invention can be achieved through the following technical solution: a high-throughput intelligent microbial cloning and information analysis device, comprising a body, a strain feeding mechanism, a petri dish feeding mechanism, a lid opening mechanism, a lid opening and photography mechanism, an inoculation mechanism, an inoculation needle replacement mechanism, a liquid level detection system, a pressure detection system, a high computing power module, and a control and management system;
[0007] The inoculum feeding mechanism is used to transport inoculum culture dishes to the first target location;
[0008] The lid-opening and photography mechanism is used to open the lid of the bacterial culture dish by adsorbing the lid of the bacterial culture dish that has been moved to the first target position through the first moving vacuum suction cup; and then to take pictures of the bacterial culture by moving the industrial camera above the bacterial culture.
[0009] The petri dish feeding mechanism is used to transport empty petri dishes to the second target location;
[0010] The lid-opening mechanism is used to open the lid of the empty culture dish by moving the second vacuum suction cup to adsorb the lid that has been moved to the second target position.
[0011] The inoculation device includes an inoculation needle apparatus for inoculating bacterial strains into empty petri dishes;
[0012] A liquid level detection system is used to detect the distance between the inoculation needle and the empty culture dish during inoculation.
[0013] The pressure detection system is used during inoculation to detect whether the inoculation needle is compressing and damaging the culture dish;
[0014] The control and management system is used to control the movement of the inoculum feeding mechanism, the petri dish feeding mechanism, the lid opening mechanism, the lid opening and photographing mechanism, and the inoculation mechanism; it also includes receiving information from the liquid level detection system and the pressure detection system based on the control and management system to control the movement of the inoculation mechanism.
[0015] The high-performance computing module is used to receive images of microorganisms captured by industrial cameras and identify the microorganisms.
[0016] Furthermore, the microbial feeding mechanism includes two parallel sets of first synchronous pulleys, which are driven by a first motor to move synchronously; a first transverse positioning push rod is provided below the two parallel sets of first synchronous pulleys, and during operation, the first transverse positioning push rod is in an extended stroke state to intercept the microbial culture dish to the first target position;
[0017] A first longitudinal positioning push rod is provided on one side of the culture dish located at the first target position, and a rear baffle of the culture dish is provided on the other side. During operation, the first longitudinal positioning push rod extends its stroke to push the culture dish to the rear baffle of the culture dish so that the culture dish remains stationary.
[0018] The side with the first longitudinal positioning push rod is also provided with a front baffle of the culture dish, which works in conjunction with the rear baffle of the culture dish to ensure that the culture dish will not slide out of the first synchronous pulley group;
[0019] The culture dish feeding mechanism includes two parallel sets of second synchronous pulleys, which are driven by a second motor to move synchronously. A second lateral positioning push rod is provided below the two parallel sets of second synchronous pulleys. During operation, the second lateral positioning push rod is in an extended stroke state to intercept the empty culture dish to the second target position.
[0020] A second longitudinal positioning push rod is provided on one side of the empty culture dish located at the second target position, and an empty culture dish front baffle is provided on the other side. During operation, the extension stroke of the second longitudinal positioning push rod pushes the empty culture dish to the empty culture dish front baffle so that the empty culture dish remains stationary.
[0021] The side with the second longitudinal positioning push rod is also equipped with a rear baffle for empty culture dishes, which works in conjunction with the front baffle for empty culture dishes to ensure that the culture dishes do not slide out of the first synchronous pulley group.
[0022] Furthermore, the cover-opening and photographing mechanism includes a synchronous belt module, which is driven by a third motor to move left and right; a movable support plate is provided on the synchronous belt module, and a movable cover plate is connected to the movable support plate; an industrial camera is installed on one side of the movable cover plate, and a light shield is provided at one end of the camera head of the industrial camera, with a light source inside the light shield; a first lifting device is installed on the other side of the movable cover plate, and the output end of the first lifting device is connected to a first vacuum suction cup;
[0023] The opening mechanism includes a rotary cylinder, the output end of which is fixedly connected to one end of a rotating frustum, the other end of which is fixedly connected to one end of a frustum support column, and the other end of the frustum support column is fixedly connected to a rotating fixing plate; the rotating fixing plate is provided with a second lifting device, and the output end of the second lifting device is connected to a second vacuum suction cup.
[0024] Furthermore, the first lifting device includes a first lifting cylinder, and the output end of the first lifting cylinder is connected to the first vacuum suction cup.
[0025] The second lifting device includes a second lifting cylinder, and the output end of the second lifting cylinder is connected to the second vacuum suction cup.
[0026] Furthermore, the inoculation mechanism includes two parallel first linear modules; a second linear module is disposed between the two parallel first linear modules; the second linear module is provided with a third lifting device; the output end of the third lifting device is connected to the inoculation needle device.
[0027] The inoculation needle device includes a fourth motor and two rollers with a gap between them for passing through and squeezing the inoculation needle material; the fourth motor is used to drive the rollers to rotate; when changing the inoculation needle, the fourth motor drives the rollers to rotate to carry out the inoculation needle material.
[0028] Furthermore, the third lifting device includes a lead screw module, and the slider of the lead screw module is connected to the inoculation needle device.
[0029] Furthermore, the liquid level detection system includes an ultrasonic liquid level detection sensor; the ultrasonic liquid level detection sensor is mounted on the lead screw module;
[0030] The pressure detection system includes a patch pressure resistance sensor, which is installed at the inoculation needle interface in the inoculation needle device.
[0031] Furthermore, it includes an inoculation needle replacement mechanism for replacing inoculation needles; the inoculation needle replacement mechanism includes a waste collection box, a gate, and an inoculation needle material box; the gate is used to cut off the inoculation needle, and a waste collection box is provided below the gate; the inoculation needle material box contains inoculation needle material and is provided with an inoculation needle material outlet; the inoculation needle material exits from the inoculation needle material outlet and enters the inoculation needle device along the inoculation needle material guide tube;
[0032] The inoculation needle replacement mechanism also includes a ring-shaped ultraviolet lamp, which is located at the inoculation needle material outlet.
[0033] Furthermore, the high-computing-power module includes a DETR model; the bacterial strain images captured by the industrial camera are preprocessed and then input into the DETR model to identify the bacterial strains;
[0034] The DERT model has been improved in the following ways: a local attention module is introduced on the basis of the original multi-head attention module; the number of attention heads in the multi-head attention module is increased; and a residual structure is introduced.
[0035] The present invention also provides a method of using the aforementioned high-throughput intelligent microbial cloning and information analysis device, comprising:
[0036] The culture dishes and empty culture dishes are transported to the first target position and the second target position respectively via the culture feeding mechanism and the culture dish feeding mechanism;
[0037] The caps of the culture dishes and empty culture dishes at the first and second target locations are opened by the cap-opening and imaging mechanisms, respectively. The cap-opening and imaging mechanism also uses an industrial camera to photograph and record the bacteria in the culture dishes. The high-computing module receives the images of the bacteria captured by the industrial camera and uses a target detection algorithm to locate and classify the bacteria distributed in the culture dishes. The morphological characteristics of the bacteria are obtained by the target detection, which can determine the growth maturity of each colony. The algorithm selects mature colonies as inoculation targets. In the subsequent inoculation action of the equipment, the inoculation structure selects mature colonies to pick up and inoculate.
[0038] After opening the top cover, the inoculation mechanism moves the inoculation needle to pick up each bacterial strain in turn and inoculate it into an empty culture dish. The liquid level detection system and the pressure detection system work together to detect whether the height of the inoculation needle is reasonable.
[0039] After all inoculation is completed, the capping and photographing mechanism and the capping mechanism close the caps on the culture dishes and empty culture dishes (at this time, the empty culture dishes are already inoculated with bacteria), and the culture feeding mechanism and the culture dish feeding mechanism deliver them out respectively.
[0040] After each inoculation is completed, the following steps are also included: the inoculation device moves the inoculation needle to the gate to cut off the inoculation needle, and the inoculation needle material extracted by the inoculation needle device becomes the next inoculation needle; the inoculation needle material is sterilized by ultraviolet light;
[0041] The step of locating and classifying the various bacterial species distributed in the bacterial culture dish using a target detection algorithm includes: preprocessing the image; inputting the preprocessed image into the DETR model to identify the bacterial species; the preprocessing includes noise reduction, filtering, and sharpening.
[0042] The step of inputting the preprocessed image into the DETR model to identify bacterial species includes:
[0043] The preprocessed image is first input into the feature extraction module of the ResNet-50 backbone network. Features are extracted through convolution to obtain the original target feature map, which is then normalized. Following the ReLU activation function, four residual blocks are applied, with the last residual block, which has the richest semantic information, selected. A 1×1 convolutional layer is used to reduce the dimension of the selected residual block, obtaining independent unit tokens in the sequence. Before inputting these tokens into the encoder, they are converted into vector representations. Simultaneously, positional encoding vectors are generated using sine and cosine functions. A local attention module is introduced during this process to add positional information to each token. These positional encodings are added to the original target feature map and transposed, then input into the Transformer encoder to generate multiple learnable object queries, where the number of object queries is greater than the number of colonies distributed in the petri dish.
[0044] By introducing a linear transformation, object queries are generated into query, key, and value representations corresponding to the input of the multi-head attention module. This module performs similarity calculations on each key to obtain the weight coefficients of the value corresponding to each key. These weight coefficients are processed by the softmax activation function to obtain the weights of each attention head. Finally, these weights are multiplied by their corresponding values and summed to obtain the output of each attention head. The introduction of the multi-head attention module enables the model to focus on different regional scales when extracting image features, thereby effectively capturing multi-scale feature information. At the same time, this module is adaptive and can select key information according to different scenarios.
[0045] The object queries and the original target feature map are input into the Transformer decoder to generate multiple feature outputs.
[0046] The feature output is input into the MLP prediction head. The feature output of each object query is processed by the FFN layer to predict the morphological feature category and pixel location information of all bacterial species in the culture dish and record them in the bacterial species information database. Furthermore, the bacterial species in the current experiment are compared with the bacterial species information database. Bacterial species of the same category are classified and their information for this experiment is recorded to identify the differences or commonalities in the cultivation status and growth characteristics of the same bacterial species in multiple experiments.
[0047] When calculating the loss function of the DETR model, the background elements plus the bacterial target output need to be used as the label set, the feature output of each object query needs to be used as the prediction set, and the Hungarian algorithm is used for matching. The loss value is calculated based on the matching results, and then the algorithm model parameters are optimized.
[0048] Continuous model training: The DETR model is used to detect newly acquired bacterial species images in the current experiment, obtaining multiple possible predicted targets. These targets are then filtered using set confidence intervals to obtain a set of pseudo-labels. The pseudo-labels and the real labels of the current new bacterial species images are distilled together to obtain a set containing both real and pseudo-labels as the dataset. A binary matching method is used, and a knowledge distillation loss function is employed. Training dataset, where The cross-entropy of the outputs from the teacher network and the student network. This is the cross-entropy between the dataset labels and the network output during student network training, which can be adjusted during training. The weight values are used to enhance learning efficiency.
[0049] Compared with the prior art, the beneficial effects of the present invention are:
[0050] This invention provides a high-throughput intelligent microbial cloning and information analysis device. Through the automated coordination of various mechanisms, it achieves fully automated operations such as transport, opening, photographing, inoculation, capping, sterilization, needle replacement, and information collection and recording. This improves inoculation efficiency and accuracy, and eliminates the need for operators to have close contact with the microorganisms, avoiding contamination and damage. After each inoculation, the device can remove the used needle and automatically replace it with a sterilized new needle, preventing cross-contamination caused by multiple inoculations. During inoculation, the distance between the inoculation needle and the culture dish is controlled by a liquid level detection system and a pressure detection system, preventing the needle from pressing and damaging the culture dish and the culture medium. It can photograph and record data for each group of microorganisms, compare and analyze data from multiple experiments, and continuously learn to identify new microorganisms. Attached Figure Description
[0051] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.
[0052] Figure 1 is a schematic diagram of the structure of a high-throughput intelligent microbial cloning and information analysis device provided in an embodiment of the present invention.
[0053] Figure 2 is a schematic diagram of the internal structure of a high-throughput intelligent microbial cloning and information analysis device provided in an embodiment of the present invention.
[0054] Figure 3 is a schematic diagram of the structure of a microbial inoculum feeding mechanism and a culture dish feeding mechanism provided in an embodiment of the present invention.
[0055] Figure 4 is a structural schematic diagram of an opening-and-photographing mechanism provided in an embodiment of the present invention.
[0056] Figure 5 is a schematic diagram of the structure of an inoculation needle replacement mechanism provided in an embodiment of the present invention.
[0057] Figure 6 is a schematic diagram of the internal structure of an inoculation needle device provided in an embodiment of the present invention.
[0058] Figure 7 is a top view of the bottom of the body of an embodiment of the present invention.
[0059] Figure 8 is a flowchart of the usage method of a high-throughput intelligent microbial cloning and information analysis device provided in an embodiment of the present invention.
[0060] Figure 9 is a structural diagram of the bacterial strain image recognition algorithm model provided in an embodiment of the present invention.
[0061] The labels in the diagram represent: 1. Machine body; 2. Second lifting cylinder; 3. Second vacuum suction cup; 4. Rotating fixing plate; 5. Rotating frustum; 6. Frustum support column; 7. Waste collection box; 8. Gate; 9. Middle baffle; 10. Linear fixing plate; 11. First linear module; 12. X-axis fixing plate; 13. Linear module transverse fixing block; 14. Second linear module; 15. Right angle adapter plate; 16. Vertical guide rail fixing component; 17. Motor fixing plate; 18. Lead screw module; 19. Inoculation needle fixing plate; 20. Inoculation needle device; 21. Inoculation needle interface; 211. Roller; 212. Rotating electric... Machine 214, screen 22, ring ultraviolet lamp 23, inoculation needle material 24, inoculation needle material box 25, patch pressure resistance sensor 26, lid opening and photography mechanism 27, ultrasonic liquid level detection sensor 28, inoculum feeding mechanism 29, petri dish feeding mechanism 30, first motor fixing frame 2901, first two-phase stepper motor 2902, first coupling 2903, first synchronous belt transition block 2904, first synchronous belt fixing shaft seat 2905, first optical axis 2906, first synchronous belt pulley set 2907, inoculum rear baffle 2908, first conveyor baffle 2909, front baffle of the culture medium; 2910, first longitudinal positioning push rod; 2911, first transverse positioning push rod; 2912, culture medium petri dish; 2913, second longitudinal positioning push rod; 3001, second transverse positioning push rod; 3002, rear baffle of the petri dish; 3003, second conveyor baffle; 3004, front baffle of the petri dish; 3005, second optical axis; 3006, second synchronous belt fixing bearing; 3007, second synchronous belt pulley assembly; 3008, second synchronous belt transition block; 3009, second coupling; 3010, second two-phase stepper motor; 3011, second motor mounting... 3012 fixed frame, 3013 empty culture dish, 2701 synchronous belt module, 2702 first vacuum suction cup, 2703 first lifting cylinder, 2704 movable support plate, 2705 movable cover plate, 2706 industrial camera, 2707 light source, 2708 light shield, 31 industrial control computer, 32 servo driver, 33 switching power supply, 34 solenoid valve, 35 first pressure regulating valve, 36 solenoid valve, 37 air switch, 38 vacuum pump, 39 second pressure regulating valve, 40 micro air compressor, 41 cooling fan, 42 high computing power module, 43 inoculation needle guide tube. Detailed Implementation
[0062] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0063] As shown in Figures 1 to 5, the high-throughput intelligent microbial cloning and information analysis device of the present invention includes a body 1 (equipped with a chassis, an inlet, and an outlet). The body 1 is equipped with a bacterial culture feeding mechanism 29 and a petri dish feeding mechanism 30, used to transport bacterial culture dishes and empty petri dishes to the inoculation position, respectively. A lid-opening and photographing mechanism 27 is installed on the side and rear of the bacterial culture feeding mechanism 29. When the bacterial culture dish is transported to the position, the lid of the bacterial culture dish can be adsorbed by a first vacuum suction cup 2702, and then photographed by an industrial camera 2706. A lid-opening mechanism is installed on the side and front of the petri dish feeding mechanism 30, where a second vacuum suction cup 30 adsorbs the empty petri dish. The top cover of the culture dish is opened and moved to another position by the rotary cylinder 5. An inoculation mechanism is installed inside the upper part of the machine body 1 to inoculate the bacteria in the culture dish into the empty culture dish. An ultrasonic liquid level detection sensor 28 is installed on its lead screw module 19 to detect the distance between the inoculation needle and the culture dish. A patch pressure resistance sensor 26 is installed on the inoculation needle interface of the inoculation needle device 21 to detect whether the inoculation needle is pressing and damaging the culture dish. The inoculation needle material box 25 is installed on the side of the machine body. When the inoculation needle material 24 is discharged, the ring ultraviolet lamp 23 performs a sterilization operation. After each inoculation, the inoculation needle is cut off by the gate 9 installed below the middle baffle 10.
[0064] The inoculum feeding mechanism 29 is installed inside the machine body 1 and includes a first motor mounting bracket 2901, a first two-phase stepper motor 2902 (i.e., the first motor), a first coupling 2903, a first synchronous belt transition block 2904, a first synchronous belt fixed shaft seat 2905, a first optical shaft 2906, a first synchronous belt pulley group 2907, an inoculum rear baffle 2908, a first conveying baffle 2909, an inoculum front baffle 2910, a first longitudinal positioning push rod 2911, and a first transverse positioning push rod 2912. The inoculum front baffle 2910 and the inoculum rear baffle 2908 are respectively fixed to a fixed chassis inside the machine body. The first synchronous belt pulley group 2907 is divided into front and rear parts, wherein the pulley is movably connected to and supported by the first optical shaft 2906. The first synchronous belt transition block 2904 consists of four blocks, which are respectively fixed to the left and right sides of the inoculum front baffle 2910 and the inoculum rear baffle 2908. The first synchronous belt pulley group 2907 protects the synchronous belt on the outside. The first transmission baffle 2909 is installed at the four corners on the synchronous belt transition block 2904 to protect the synchronous belt pulley group 2907. The height of the first transmission baffle 2909 is lower than that of the first synchronous belt pulley group 2907. The first motor fixing bracket 2901 and the first two-phase stepper motor 2902 are installed on the right side of the first synchronous belt pulley group 2907. The output shaft of the first two-phase stepper motor 2902 is connected to the first synchronous belt pulley group 2907 via the first coupling 2903 and the first optical shaft 2906. The first synchronous belt fixing shaft seat 2905 is installed between the two pulleys of the first synchronous belt pulley group 2907. It has a through hole through which the first optical shaft 2906 can pass and provide support. The first transverse positioning push rod 2912 is installed below the first transmission baffle 2909 on the front side of the inoculation position. The first longitudinal positioning push rod 2911 is installed below the middle baffle 10 on the side of the inoculation position.
[0065] The culture dish 2913 enters through the feed port on the right side of the machine body 1 and falls onto the first synchronous pulley group 2907. The first two-phase stepper motor 2902 drives the first coupling 2903 and the first optical shaft 2906 to rotate, thereby driving the first synchronous pulley group 2907 forward. The culture dish 2913 is conveyed forward. At this time, the first transverse positioning push rod 2912 is in the extended stroke state. When the culture dish 2913 is conveyed to the inoculation position (i.e. the first target position), it is intercepted by the first transverse positioning push rod 2912. The first two-phase stepper motor 2902 stops operating. The first longitudinal positioning push rod 2911 extends its stroke to push the culture dish 2913 to the rear baffle 2908, thereby fixing the position of the culture dish 2913 at the inoculation position.
[0066] The culture dish feeding mechanism 30 is located in front of and parallel to the culture feeding mechanism 29. It consists of a second longitudinal positioning push rod 3001, a second transverse positioning push rod 3002, a culture dish rear baffle 3003, a second conveying baffle 3004, a culture dish front baffle 3005, a second optical shaft 3006, a second synchronous belt fixing shaft seat 3007, a second synchronous belt pulley group 3008, a second synchronous belt transition block 3009, a second coupling 3010, a second two-phase stepper motor 3011 (i.e., the second motor), and a second motor mounting bracket 3012. The connection method of each component is basically the same as that of the culture feeding mechanism 29. The second two-phase stepper motor 3011 and the second motor mounting bracket 3012 are installed on the left side of the second synchronous belt pulley group 3008. The output shaft of the second two-phase stepper motor 3011 is connected to the second optical shaft 3006 via the second coupling 3010. 06 Connection: The second optical axis 3006 passes through the pulley bearing of the second synchronous belt pulley group 3008 and connects to it, supporting it. A second synchronous belt fixing shaft seat 3007 is installed between the two pulleys, with a through hole. The second optical axis 3006 is supported by the through hole. Four second synchronous belt transition blocks 3009 are located on the outside of the pulleys of the second synchronous belt pulley group 3008, and are installed on the rear baffle 3003 and the front baffle 3005 of the culture dish to protect the synchronous belt. The rear baffle 3003, the second conveying baffle 3004, and the front baffle 3005 of the culture dish are installed on both sides and in the middle of the second synchronous belt pulley group 3008 for protection. The second transverse positioning push rod 3002 is installed below the second conveying baffle 3004 on the front side of the inoculation position. The second longitudinal positioning push rod 3001 is installed below the middle baffle 10 on the side of the inoculation position.
[0067] Empty culture dish 3013 falls onto the second synchronous pulley group 3008 through the feed port on the right side of the machine body. The second two-phase stepper motor 3011 drives the second synchronous pulley group 3008 to transport the empty culture dish 3013 forward. The second lateral positioning push rod 3002 extends its stroke to position the lateral coordinate of the empty culture dish 3013 (i.e., to the second target position). After the second synchronous pulley group 3008 stops, the second longitudinal positioning push rod 3001 extends its stroke to position the longitudinal coordinate of the empty culture dish 3013. The empty culture dish 3013 is then fixed at the inoculation position.
[0068] The cover-opening and photography mechanism 27 is installed behind the inoculum feeding mechanism 29. It consists of a synchronous belt module 2701, a first vacuum suction cup 2702, a first lifting cylinder 2703, a movable support plate 2704, a movable cover plate 2705, an industrial camera 2706, a light source 2707, and a light shield 2708. The synchronous belt module 2701 is installed parallel to the inoculum feeding mechanism 29 and is fixed on a chassis behind it. The movable support plate 2704 has an L-shaped structure. Its horizontal end is fixedly connected to the slider of the synchronous belt module 2701, and its vertical end is connected to the movable cover plate 2705. The industrial camera 2706, the light source 2707, and the light shield 2708 are installed on the left side of the movable cover plate 2705. The first lifting cylinder 2703 is installed on the right side of the movable cover plate 2705, and its output end is connected to the first vacuum suction cup 2702.
[0069] When the culture dish 2913 is fixed at the inoculation position (i.e., the first target position), the synchronous belt module 2701 drives the entire mechanism to move laterally, moving the first vacuum suction cup 2702 above the culture dish 2913. The first lifting cylinder 2703 lowers the first vacuum suction cup 2702 and adsorbs the top cover of the culture dish 2913, then returns to its original position to move the top cover of the culture dish 2913 upward, completing the opening action. The synchronous belt module 2701 continues to move, moving the industrial camera 2706 above the opened culture dish 2913 and performing a photo-taking operation. The light source 2707 is installed inside the light shield 2708 to provide brightness. After the photo is taken, it is transmitted to the high-computing-power module.
[0070] The bacterial data collected by the industrial camera 2706 is recorded in the high-computing module 42. First, the image is preprocessed, including noise reduction, filtering, and sharpening. The preprocessed image is then fed into the DETR model, and features are extracted through the ResNet-50 backbone network and convolutional layers. Position encoding is added to each feature. Then, the feature map is output through the Transformer encoder, generating multiple object queries. These queries are then processed by the multi-head attention module and, together with the output feature map, are passed through the Transformer decoder to generate multiple feature outputs. Each feature output is then used by the MLP prediction head to identify all bacterial species.
[0071] The DETR model identifies the target bounding boxes of the bacterial strains from the images. By combining the center pixel coordinates of the target bounding boxes with the coordinates of the bacterial culture dish within the apparatus, the relative coordinate relationship between the bacterial strain and the apparatus can be obtained. The DETR model can also classify and record the bacterial strains based on information such as colony distribution morphology, colony size, colony height, colony color, and transparency by comparing them with the bacterial strain information database and the current experimental strains. It can also identify morphological differences between multiple experimental bacterial strains. Based on the morphological characteristics of the colonies, mature colonies are selected and inoculated in subsequent inoculation actions. The algorithm uses the bacterial strain images obtained in the inoculation task as a dataset to directly train the model on a high-computing machine. Incremental learning on images of both known and unknown bacterial strains can further increase the detection accuracy and the number of bacterial strains that can be identified.
[0072] The opening mechanism consists of a second lifting cylinder 2, a second vacuum suction cup 3, a rotating fixing plate 4, a rotating cylinder 5, a rotating frustum 6, and a frustum support column 7. The rotating cylinder 5 is installed in front of the front baffle 3005 of the culture dish, and its rotating platform is fixedly connected to the rotating frustum 6. The other end of the rotating frustum 6 is fixedly connected to the frustum support column 7 and the rotating fixing plate 4. The second lifting cylinder 2 and the second vacuum suction cup 3 are fixed on the rotating fixing plate 4, wherein the second vacuum suction cup 3 is fixed on the output shaft of the second lifting cylinder 2.
[0073] When the empty culture dish 3013 is transported to the inoculation position (second target position), the rotary cylinder 5 drives the rotary frustum 6, the frustum support column 7 and the rotary fixing plate 4 to rotate, moving the second lifting cylinder 2 and the second vacuum suction cup 3 above the empty culture dish 3013. The second lifting cylinder 2 lowers and the second vacuum suction cup 3 adsorbs the top cover of the empty culture dish 3013. Then the second lifting cylinder 2 is raised again, and the rotary cylinder 5 continues to rotate to transfer the top cover of the empty culture dish 3013 to other positions.
[0074] As shown in Figure 6, the inoculation mechanism consists of a linear fixing plate 11, a first linear module 12, an x-axis fixing plate 13, a linear module transverse fixing block 14, a second linear module 15, a right-angle adapter plate 16, a vertical guide rail fixing component 17, a motor fixing plate 18, a lead screw module 19, an inoculation needle fixing plate 20, and an inoculation needle device 21. Two linear fixing plates 11 are installed on the upper part of the machine body, one on the left and one on the right. One first linear module 12 is fixed on each of the left and right linear fixing plates 11 for longitudinal movement. The x-axis fixing plate 13 is fixed to the first linear module 12. On the slider, a horizontal fixing block 14 and a second linear module 15 are fixed above it for horizontal movement. The lead screw module 19 is connected to the second linear module 15 through a right-angle adapter plate 16, a vertical guide rail fixing piece 17, and a motor fixing plate 18. The inoculation needle fixing plate 20 has an L-shaped structure, with its vertical end fixedly connected to the slider of the lead screw module 19 and its lower horizontal end equipped with an inoculation needle device 21. An ultrasonic liquid level detection sensor 28 is installed on the side of the lead screw module 19, and a patch pressure resistance sensor 26 is installed on the needle head of the inoculation needle device 21.
[0075] After the inoculum feeding mechanism 29 and the petri dish feeding mechanism 30 fix the inoculum culture dish 2913 and the empty petri dish 3013 at the inoculation position, respectively, the first linear module 12 and the second linear module 15 operate respectively, moving the inoculation needle device 21 above the inoculum culture dish 2913 according to the acquired inoculum coordinates (at this time, the top cover of the inoculum culture dish 2913 has been removed). The lead screw module 19 slowly descends until the inoculation needle on the inoculation needle device 21 dips into the inoculum. The ultrasonic liquid level detection sensor 28 is used on the side of the inoculation needle device 21 to detect the descent height and prevent the inoculation needle from being lowered. If the culture medium is damaged due to insufficient height or pressure, the patch pressure resistance sensor 26 can detect whether the inoculation needle is compressing the culture medium based on the deformation information of the inoculation needle and the needle tip. The sensor receiver transmits the information to the control and management system (industrial computer 31). After the culture is dipped, the first linear module 12 and the second linear module 15 move the inoculation needle device 21 above the empty culture dish 3013. The lead screw module 19 lowers the height to inoculate the culture medium onto the empty culture dish 3013. Repeating the above actions can complete the inoculation task for all the cultures that need to be inoculated.
[0076] After inoculation, the synchronous belt module 2701 moves the top cover of the culture dish 2913 above the culture dish 2913. The first lifting cylinder 2703 lowers the first vacuum suction cup 2702 to complete the capping action of the culture dish 2913. The rotating cylinder 5 moves the top cover of the empty culture dish 3013 above the empty culture dish 3013, and the second lifting cylinder 2 lowers the second vacuum suction cup 3 to complete the capping action of the empty culture dish 3013.
[0077] After the lid is closed, the first longitudinal positioning push rod 2911 and the second longitudinal positioning push rod 3001, the first transverse positioning push rod 2912 and the second transverse positioning push rod 3002 retract their strokes respectively, the inoculum culture dish 2913 and the empty culture dish 3013 are unrestrained, the first two-phase stepper motor 2902 and the second two-phase stepper motor 3011 start to operate and drive the first synchronous pulley group 2907 and the second synchronous pulley group 3008 to move respectively, transporting the inoculum culture dish 2913 and the empty culture dish 3013 (at this time the empty culture dish 3013 has been inoculated with inoculum) to the left to the discharge port on the left side of the machine body.
[0078] The inoculation needle replacement mechanism consists of a waste collection box 8, a gate 9, a middle baffle 10, a ring-shaped ultraviolet lamp 23, inoculation needle material 24, and an inoculation needle material box 25. The inoculation needle material box 25 is installed on the right side of the machine body 1 and contains the ring-wound inoculation needle material 24. A ring-shaped ultraviolet lamp 23 is installed at the outlet of the inoculation needle material box 25. When the inoculation needle material 24 is discharged from the outlet, the ring-shaped ultraviolet lamp 23 performs sterilization. The inoculation needle material 24 is a soft material. It flows from the outlet along the inoculation needle material guide tube 43, then through the through holes of various components of the inoculation mechanism, and connects to the inoculation needle device 21. It also exits from the inoculation needle interface of the inoculation needle device 21. 211 extends (the inoculation needle material extending from the inoculation needle interface 211 is the inoculation needle). After each inoculation action, the inoculation mechanism moves the inoculation needle device 21 to the gate 9 and cuts off the used inoculation needle. The waste inoculation needle is placed in the waste collection box 8. The inoculation needle device 21 is equipped with a rotary motor 214 and is connected to two rollers 212. The inoculation needle material 24 passes through the two rollers 212 and is squeezed. When changing the needle, the rotary motor 214 rotates and drives the rollers 212 to squeeze the inoculation needle material 24 downward. The inoculation needle material 24 is brought out from the inside of the inoculation needle device 21 and replaced with a new inoculation needle.
[0079] As shown in Figure 7, the control and management system includes hardware devices and software control systems such as a screen 22, an industrial computer 31, servo drivers 32, a switching power supply 33, solenoid valves 34, pressure regulating valves 35, solenoid valves 36, air switches 37, a vacuum pump 38, a miniature air compressor 40, and cooling fans 41. This system coordinates the overall operation and status monitoring of the control equipment. The screen 22 is installed on the front of the machine body 1 and connected to the industrial computer 31 for control and interaction between the operator and the equipment. The cooling fan 41 is installed on the bottom rear cover of the machine body 1 to dissipate heat generated by the hardware. Five servo drivers 32 are located on the bottom side plate and front and rear covers of the machine body for heat dissipation, serving as the cooling source for each linear module (…). Two first linear modules 12, a second linear module 15, a synchronous belt module 2701, and a lead screw module 19 provide drive support. A miniature air compressor 40 is used to compress air. A first pressure regulating valve 35 and a second pressure regulating valve 39 adjust the output air pressure of the first lifting cylinder 2703 and the second lifting cylinder 2. A solenoid valve 34 acts as a switch to output gas to the second lifting cylinder 2 and the first lifting cylinder 2703 of the cover opening mechanism and the cover opening and photographing mechanism. A vacuum pump 38 draws gas for the second vacuum suction cup 3 and the first vacuum suction cup 2702 of the cover opening mechanism and the cover opening and photographing mechanism. A switching power supply 33 and an air switch 37 provide power for the equipment pressure regulation. All devices are installed at the bottom of the machine body.
[0080] This invention also provides a method for using a high-throughput intelligent microbial cloning and information analysis device, as shown in Figure 8, including:
[0081] The culture dishes and empty culture dishes are transported to the first target position and the second target position respectively via the culture feeding mechanism and the culture dish feeding mechanism;
[0082] The caps of the culture dishes and the empty culture dishes at the first and second target locations are opened by the cap-opening and imaging mechanisms, respectively. The cap-opening and imaging mechanism also uses an industrial camera to photograph and record the bacteria in the culture dishes. The high-computing module receives the images of the bacteria captured by the industrial camera and uses a target detection algorithm to locate and classify the various bacteria distributed in the culture dishes.
[0083] After opening the top cover, the inoculation mechanism moves the inoculation needle to pick up each bacterial strain in turn and inoculate it into an empty culture dish. The liquid level detection system and the pressure detection system work together to detect whether the height of the inoculation needle is reasonable.
[0084] After all inoculation is completed, the capping and photographing mechanism and the capping mechanism close the caps on the culture dishes and empty culture dishes (at this time, the empty culture dishes are already inoculated with bacteria), and the culture feeding mechanism and the culture dish feeding mechanism deliver them out respectively.
[0085] After each inoculation is completed, the following steps are also included: the inoculation device moves the inoculation needle to the gate to cut off the inoculation needle, and the inoculation needle material extracted by the inoculation needle device becomes the next inoculation needle; the inoculation needle material is sterilized by ultraviolet light;
[0086] The step of locating and classifying the various bacterial species distributed in the bacterial culture dish using a target detection algorithm includes: preprocessing the image; inputting the preprocessed image into the DETR model to identify the bacterial species; the preprocessing includes noise reduction, filtering, and sharpening.
[0087] The process of inputting the preprocessed image into the DETR model to identify bacterial species, as shown in Figure 9, includes:
[0088] The preprocessed image is first input into the feature extraction module of the ResNet-50 backbone network. Features are extracted through convolution to obtain the original target feature map, which is then normalized. Following the ReLU activation function, four residual blocks are applied, with the last residual block, which has the richest semantic information, selected. A 1×1 convolutional layer is used to reduce the dimension of the selected residual block, obtaining independent unit tokens in the sequence. Before inputting these tokens into the encoder, they are converted into vector representations. Simultaneously, positional encoding vectors are generated using sine and cosine functions. A local attention module is introduced during this process to add positional information to each token. These positional encodings are added to the original target feature map and transposed, then input into the Transformer encoder to generate multiple learnable object queries, where the number of object queries is greater than the number of colonies distributed in the petri dish.
[0089] By introducing a linear transformation, object queries are generated into query, key, and value representations corresponding to the input of the multi-head attention module. This module performs similarity calculations on each key to obtain the weight coefficients of the value corresponding to each key. These weight coefficients are processed by the softmax activation function to obtain the weights of each attention head. Finally, these weights are multiplied by their corresponding values and summed to obtain the output of each attention head. The introduction of the multi-head attention module enables the model to focus on different regional scales when extracting image features, thereby effectively capturing multi-scale feature information. At the same time, this module is adaptive and can select key information according to different scenarios.
[0090] The object queries and the original target feature map are input into the Transformer decoder to generate multiple feature outputs.
[0091] The feature output is input into the MLP prediction head. The feature output of each object query is processed by the FFN layer to predict the morphological feature category and pixel location information of all bacterial species in the culture dish and record them in the bacterial species information database. Furthermore, the bacterial species in the current experiment are compared with the bacterial species information database. Bacterial species of the same category are classified and their information for this experiment is recorded to identify the differences or commonalities in the cultivation status and growth characteristics of the same bacterial species in multiple experiments.
[0092] When calculating the loss function of the DETR model, the background elements plus the bacterial target output need to be used as the label set, the feature output of each object query needs to be used as the prediction set, and the Hungarian algorithm is used for matching. The loss value is calculated based on the matching results, and then the algorithm model parameters are optimized.
[0093] Continuous model training: The DETR model is used to detect newly acquired bacterial species images in the current experiment, obtaining multiple possible predicted targets. These targets are then filtered using set confidence intervals to obtain a set of pseudo-labels. The pseudo-labels and the real labels of the current new bacterial species images are distilled together to obtain a set containing both real and pseudo-labels as the dataset. A binary matching method is used, and a knowledge distillation loss function is employed. Training dataset, where The cross-entropy of the outputs from the teacher network and the student network. This is the cross-entropy between the dataset labels and the network output during student network training, which can be adjusted during training. The weight values are adjusted to enhance learning efficiency. Incremental learning is used to acquire the ability to detect new bacterial species while maintaining the ability to detect older categories.
[0094] Although the present invention has been disclosed in the above embodiments, it is not intended to limit the invention. Any person skilled in the art can make many possible variations and improvements to the technical solutions of the present invention, or modify them into equivalent implementations, without departing from the scope of the present invention. Therefore, any simple modifications, equivalent changes, or improvements made to the embodiments described above in accordance with the technical essence of the present invention are within the protection scope of the present invention, but do not depart from the content of the present invention.
Claims
1. A high-throughput intelligent microbial cloning and information analysis device, characterized in that, It includes the main body, inoculum feeding mechanism, petri dish feeding mechanism, lid opening mechanism, lid opening and photography mechanism, inoculation mechanism, inoculation needle replacement mechanism, liquid level detection system, pressure detection system, high computing power module and control and management system; The inoculum feeding mechanism is used to transport inoculum culture dishes to the first target location; The lid-opening and photography mechanism is used to open the lid of the bacterial culture dish by adsorbing the lid of the bacterial culture dish that has been moved to the first target position through the first moving vacuum suction cup; and then to take pictures of the bacterial culture by moving the industrial camera above the bacterial culture. The petri dish feeding mechanism is used to transport empty petri dishes to the second target location; The lid-opening mechanism is used to open the lid of the empty culture dish by moving the second vacuum suction cup to adsorb the lid that has been moved to the second target position. The inoculation device includes an inoculation needle apparatus for inoculating bacterial strains into empty petri dishes; A liquid level detection system is used to detect the distance between the inoculation needle and the empty culture dish during inoculation. The pressure detection system is used during inoculation to detect whether the inoculation needle is compressing and damaging the culture dish; The control and management system is used to control the movement of the inoculum feeding mechanism, the petri dish feeding mechanism, the lid opening mechanism, the lid opening and photographing mechanism, and the inoculation mechanism; it also includes receiving information from the liquid level detection system and the pressure detection system based on the control and management system to control the movement of the inoculation mechanism. The high-performance computing module is used to receive images of microorganisms captured by industrial cameras and identify the microorganisms.
2. The device according to claim 1, characterized in that, The inoculum feeding mechanism includes two parallel first synchronous pulley groups, which are driven by a first motor to move synchronously; a first transverse positioning push rod is provided below the two parallel first synchronous pulley groups, and during operation, the first transverse positioning push rod is in an extended stroke state to intercept the inoculum culture dish to the first target position. A first longitudinal positioning push rod is provided on one side of the culture dish located at the first target position, and a rear baffle of the culture dish is provided on the other side. During operation, the first longitudinal positioning push rod extends its stroke to push the culture dish to the rear baffle of the culture dish so that the culture dish remains stationary. The side with the first longitudinal positioning push rod is also provided with a front baffle of the culture dish, which works in conjunction with the rear baffle of the culture dish to ensure that the culture dish will not slide out of the first synchronous pulley group; The culture dish feeding mechanism includes two parallel sets of second synchronous pulleys, which are driven by a second motor to move synchronously. A second lateral positioning push rod is provided below the two parallel sets of second synchronous pulleys. During operation, the second lateral positioning push rod is in an extended stroke state to intercept the empty culture dish to the second target position. A second longitudinal positioning push rod is provided on one side of the empty culture dish located at the second target position, and an empty culture dish front baffle is provided on the other side. During operation, the extension stroke of the second longitudinal positioning push rod pushes the empty culture dish to the empty culture dish front baffle so that the empty culture dish remains stationary. The side with the second longitudinal positioning push rod is also equipped with a rear baffle for empty culture dishes, which works in conjunction with the front baffle for empty culture dishes to ensure that the culture dishes do not slide out of the first synchronous pulley group.
3. The device according to claim 1, characterized in that, The cover-opening and photography mechanism includes a synchronous belt module, which is driven by a third motor to move left and right. A movable support plate is provided on the synchronous belt module, and a movable cover plate is connected to the movable support plate. An industrial camera is installed on one side of the movable cover plate, and a light shield is provided at one end of the camera head of the industrial camera. A light source is provided inside the light shield. A first lifting device is installed on the other side of the movable cover plate, and the output end of the first lifting device is connected to a first vacuum suction cup. The opening mechanism includes a rotary cylinder, the output end of which is fixedly connected to one end of a rotating frustum, the other end of which is fixedly connected to one end of a frustum support column, and the other end of the frustum support column is fixedly connected to a rotating fixing plate; the rotating fixing plate is provided with a second lifting device, and the output end of the second lifting device is connected to a second vacuum suction cup.
4. The device according to claim 3, characterized in that, The first lifting device includes a first lifting cylinder, and the output end of the first lifting cylinder is connected to the first vacuum suction cup; The second lifting device includes a second lifting cylinder, and the output end of the second lifting cylinder is connected to the second vacuum suction cup.
5. The device according to claim 1, characterized in that, The inoculation mechanism includes two parallel first linear modules; a second linear module is disposed between the two parallel first linear modules; the second linear module is provided with a third lifting device; the output end of the third lifting device is connected to an inoculation needle device. The inoculation needle device includes a fourth motor and two rollers with a gap between them for passing through and squeezing the inoculation needle material; the fourth motor is used to drive the rollers to rotate; when changing the inoculation needle, the fourth motor drives the rollers to rotate to carry out the inoculation needle material.
6. The device according to claim 5, characterized in that, The third lifting device includes a lead screw module, and the slider of the lead screw module is connected to the inoculation needle device.
7. The device according to claim 6, characterized in that, The liquid level detection system includes an ultrasonic liquid level detection sensor; the ultrasonic liquid level detection sensor is mounted on the lead screw module. The pressure detection system includes a patch pressure resistance sensor, which is installed at the inoculation needle interface in the inoculation needle device.
8. The device according to claim 1, characterized in that, The device includes an inoculation needle replacement mechanism for replacing inoculation needles; the inoculation needle replacement mechanism includes a waste collection box, a gate, and an inoculation needle material box; the gate is used to cut off the inoculation needle, and a waste collection box is provided below the gate; the inoculation needle material box contains inoculation needle material and is provided with an inoculation needle material outlet; the inoculation needle material exits from the inoculation needle material outlet and enters the inoculation needle device along the inoculation needle material guide tube; The inoculation needle replacement mechanism also includes a ring-shaped ultraviolet lamp, which is located at the inoculation needle material outlet.
9. The device according to claim 1, characterized in that, The high-performance computing module includes a DETR model; the bacterial strain images captured by the industrial camera are preprocessed and then input into the DETR model to identify the bacterial strains; The DERT model has been improved in the following ways: a local attention module is introduced on the basis of the original multi-head attention module; the number of attention heads in the multi-head attention module is increased; and a residual structure is introduced.
10. A method of using a high-throughput intelligent microbial cloning and information analysis device according to any one of claims 1-9, characterized in that, include: The culture dishes and empty culture dishes are transported to the first target position and the second target position respectively via the culture feeding mechanism and the culture dish feeding mechanism; The caps of the culture dishes and the empty culture dishes at the first and second target locations are opened by the cap-opening and imaging mechanisms, respectively. The cap-opening and imaging mechanism also uses an industrial camera to photograph and record the bacteria in the culture dishes. The high-computing module receives the images of the bacteria captured by the industrial camera and uses a target detection algorithm to locate and classify the various bacteria distributed in the culture dishes. After opening the top cover, the inoculation mechanism moves the inoculation needle to pick up each bacterial strain in turn and inoculate it into an empty culture dish. The liquid level detection system and the pressure detection system work together to detect whether the height of the inoculation needle is reasonable. After all inoculation is completed, the capping and photographing mechanism and the capping mechanism close the caps on the culture dishes and empty culture dishes (at this time, the empty culture dishes are already inoculated with bacteria), and the culture feeding mechanism and the culture dish feeding mechanism deliver them out respectively. After each inoculation is completed, the following steps are also included: the inoculation device moves the inoculation needle to the gate to cut off the inoculation needle, and the inoculation needle material extracted by the inoculation needle device becomes the next inoculation needle; the inoculation needle material is sterilized by ultraviolet light; The step of locating and classifying the various bacterial species distributed in the bacterial culture dish using a target detection algorithm includes: preprocessing the image; inputting the preprocessed image into the DETR model to identify the bacterial species; the preprocessing includes noise reduction, filtering, and sharpening. The step of inputting the preprocessed image into the DETR model to identify bacterial species includes: The preprocessed image is first input into the feature extraction module of the ResNet-50 backbone network. Features are extracted through convolution to obtain the original target feature map, which is then normalized. Following the ReLU activation function, four residual blocks are applied, with the last residual block, which has the richest semantic information, selected. A 1×1 convolutional layer is used to reduce the dimension of the selected residual block, obtaining independent unit tokens in the sequence. Before inputting these tokens into the encoder, they are converted into vector representations. Simultaneously, positional encoding vectors are generated using sine and cosine functions. A local attention module is introduced during this process to add positional information to each token. These positional encodings are added to the original target feature map and transposed, then input into the Transformer encoder to generate multiple learnable object queries, where the number of object queries is greater than the number of colonies distributed in the petri dish. By introducing a linear transformation, object queries are generated into query, key, and value representations corresponding to the input of the multi-head attention module. This module performs similarity calculations on each key to obtain the weight coefficients of the value corresponding to each key. These weight coefficients are processed by the softmax activation function to obtain the weights of each attention head. Finally, these weights are multiplied by their corresponding values and summed to obtain the output of each attention head. The introduction of the multi-head attention module enables the model to focus on different regional scales when extracting image features, thereby effectively capturing multi-scale feature information. At the same time, this module is adaptive and can select key information according to different scenarios. The object queries and the original target feature map are input into the Transformer decoder to generate multiple feature outputs. The feature output is input into the MLP prediction head. The feature output of each object query is processed by the FFN layer to predict the morphological feature category and pixel location information of all bacterial species in the culture dish and record them in the bacterial species information database. Furthermore, the bacterial species in the current experiment are compared with the bacterial species information database. Bacterial species of the same category are classified and their information for this experiment is recorded to identify the differences or commonalities in the cultivation status and growth characteristics of the same bacterial species in multiple experiments. When calculating the loss function of the DETR model, the background elements plus the bacterial target output need to be used as the label set, the feature output of each object query needs to be used as the prediction set, and the Hungarian algorithm is used for matching. The loss value is calculated based on the matching results, and then the algorithm model parameters are optimized. Continuous model training: The DETR model is used to detect newly acquired bacterial species images in the current experiment, obtaining multiple possible predicted targets. These targets are then filtered using set confidence intervals to obtain a set of pseudo-labels. The pseudo-labels and the real labels of the current new bacterial species images are distilled together to obtain a set containing both real and pseudo-labels as the dataset. A binary matching method is used, and a knowledge distillation loss function is employed. Training dataset, where The cross-entropy of the outputs from the teacher network and the student network. This is the cross-entropy between the dataset labels and the network output during student network training, which can be adjusted during training. The weight values are used to enhance learning efficiency.
Citation Information
Patent Citations
Automatic bacteria inoculator
CN107118950A
Automatic plate inoculation instrument
CN108795730A
Microbial sample inoculation device
CN111286458A
Automatic microorganism sample treatment system
CN112725168A
High-throughput microorganism intelligent cloning and information analysis equipment and use method
CN118360136A
Cited By
Multimodal enhanced filamentous fungus microscopic image identification method and system
CN121903972A