Visual detection method for assessing the drainage regularity in a filtration operation

Abstract

This invention discloses a visual detection method for evaluating the standardization of drainage in filtration operations. By employing a state machine dynamic triggering mechanism, the optimal observation window is defined as the stable existence of the funnel and the absence of glass rod obstruction. The method accurately extracts and caches the single-channel binary mask of the three filter paper layers to avoid real-time segmentation noise. Then, based on the mask, the orientation of the three filter paper layers and the contact position between the glass rod and the filter paper layers are determined in real time. Relying on two-dimensional conditions, the method identifies when the glass rod is ready. Real-time interactive analysis and timed segmentation are implemented to achieve extremely high real-time response performance and reduce hardware requirements. Finally, the method calls the opposite-side drainage rules to adjudicate compliance, achieving full-process automation. This method ensures segmentation accuracy and judgment accuracy while reducing real-time computing power consumption, efficiently solving the core problems of existing technologies and providing reliable technical support for evaluating the standardization of experimental operations.

Description

Technical Field

[0001] This invention relates to the field of image recognition technology, and in particular to a visual detection method for evaluating the standardization of traffic flow in filtering operations. Background Technology

[0002] The precision and standardization of experimental procedures are core elements in ensuring the validity and safety of experimental results. With the development of artificial intelligence and smart education, utilizing computer vision technology for the automatic supervision, evaluation, and guidance of experimental processes has become a current research hotspot. However, existing visual analysis techniques still face significant challenges when applied to complex and delicate experimental operations.

[0003] For example, in biochemical experiments, filtration is a common experimental step, and the folding of the filter paper is crucial, directly affecting the filtration efficiency and speed. The four-fold method is a standard folding method, where a circular filter paper is folded twice to create a structure with one side being a single layer and the other three layers, facilitating placement into a funnel for efficient filtration. Currently, visual analysis solutions for locating fine areas like "three-layer filter paper" in experiments can be mainly categorized as follows:

[0004] First, a solution based on general object detection: This method attempts to treat the "three-layer filter paper area" as an independent object and use a rectangular frame for positioning. However, this is difficult to implement in practice. The fundamental problem is "shape mismatch" because the fan-shaped or triangular area of ​​the three-layer filter paper cannot be accurately defined by the rectangular frame, resulting in a rough and unreliable positioning result.

[0005] Second, the traditional real-time semantic segmentation scheme: In pursuit of accuracy, this method usually first locates the funnel, and then runs the segmentation model for each frame of the region to identify pixels. However, this scheme is very costly, not only because the annotation cost is high, but also because it has extremely high requirements for the real-time computing power of the hardware, making it unsuitable for large-scale, low-cost deployment.

[0006] Third, the key point detection scheme based on the region: This method also requires first locating the funnel, and then predicting the two key vertices of the crease in the region to form the dividing line. This scheme is a compromise between the first two methods in terms of cost and accuracy, but it still has limitations.

[0007] However, when faced with real-world, complex experimental scenarios, all of the aforementioned solutions collectively revealed a series of deep-seated technical challenges. These challenges are not simply matters of model accuracy, but rather structural flaws that existing technological paradigms cannot fundamentally address. Specifically, they manifest as follows:

[0008] Key functional areas cannot be identified: In the filtration experiment, the glass rod must lean against the three layers of filter paper. The three layers of filter paper are visually very similar to the single-layer edge. Traditional visual methods cannot effectively distinguish and locate them, making it impossible to determine whether the first step of the experimental operation is correct.

[0009] Information is obscured at critical moments: During the crucial moment of pouring liquid, the operator's hand, the beaker being held, and the liquid flowing down will completely or severely obstruct the contact area between the glass rod and the filter paper. Therefore, all solutions that rely on "real-time analysis" will fail due to the loss of visual information at this moment when analysis is most needed.

[0010] Functional error posture cannot be identified: A sophisticated and subtle operational error exists during filtration experiments: "unidirectional pouring." Specifically, assuming the three layers of filter paper are on the operator's right side, the operator uses their right hand to place the end of the glass rod against the filter paper, while simultaneously using their left hand to hold the beaker and pour the liquid. In this ergonomically inefficient posture, the operator can only hold the glass rod vertically, meaning the spout of the beaker, the glass rod, and the three layers of filter paper are all on the same vertical plane. This causes the glass rod to lose its inclined drainage surface, and the liquid directly impacts the filter paper, completely disabling the core function of "drainage." Therefore, all existing technologies that only focus on "positioning" cannot detect this functional error. Summary of the Invention

[0011] To address the aforementioned issues, this invention proposes a visual inspection method for evaluating the standardization of drainage during filtration operations. This method aims to solve the core challenges of existing technologies in precise experimental evaluation, namely, the inability to accurately identify sub-object-level functional areas (such as three layers of filter paper), the inability to effectively handle severe occlusion during dynamic processes, and the inability to deeply evaluate the functionality and intent of the operation (such as the "same-direction tilting" posture).

[0012] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0013] A visual inspection method for evaluating the standardization of diversion in filtering operations includes the following steps:

[0014] Acquire a real-time video stream of the target scene, identify key objects in each frame of the real-time video stream, and encapsulate them as key information associated with the corresponding frame images;

[0015] The key information is regarded as a state machine. When the state machine meets the optimal observation window condition for the traffic diversion task, a timed segmentation is triggered. When the state machine does not meet the optimal observation window condition for the traffic diversion task but meets the interactive event condition, a real-time interactive analysis is triggered. When the state machine has completed at least one real-time interactive analysis and meets the tilting action condition, a functional posture evaluation is triggered.

[0016] The timing-based segmentation includes: using the funnel detection box as a reference, cropping the target image region containing the funnel from the image, inputting it into the segmentation network, outputting a single-channel binary mask with the same area as the target image region, and the corresponding confidence score, wherein the non-zero pixels in the single-channel binary mask represent the three-layer filter paper region, and caching the funnel detection box, the single-channel binary mask and the confidence score as the optimal segmentation result;

[0017] The real-time interactive analysis includes: reading the optimal segmentation result, using the funnel detection frame as a reference, transforming the single-channel binary mask to the global coordinate system of the corresponding frame image, comparing the centroid of the transformed single-channel binary mask with the center point of the funnel detection frame to obtain the orientation features of the three-layer filter paper, detecting whether the corner pixels of the glass rod detection frame overlap with the single-channel binary mask, and if so, outputting the orientation features of the contact point between the glass rod detection frame and the three-layer filter paper, and encapsulating the orientation features of the three-layer filter paper and the orientation features of the contact point into an interactive analysis result;

[0018] The functional posture assessment includes: extracting the latest hand detection frame, beaker detection frame, and glass rod detection frame from the key information; calculating the intersection-union ratio (IUU) of the two hand detection frames with the beaker detection frame and the glass rod detection frame, respectively; determining the hand orientation attributes of the beaker and the glass rod; invoking preset opposite-side traffic diversion rules; using the three-layer filter paper orientation features and the hand orientation attributes of the beaker in the interaction analysis results as the criteria to determine whether the user's current traffic diversion operation is compliant; and outputting the determination result as the final assessment result.

[0019] In some implementations, the key information includes at least the funnel detection frame, the glass rod detection frame, the beaker detection frame, and the hand detection frame.

[0020] In some implementations, the optimal observation window condition for the traffic diversion task includes continuously detecting whether the current frame image simultaneously satisfies a first condition and a second condition, wherein the first condition includes the funnel detection box being stably present in the image, and the second condition includes the glass rod detection box not being present, or the intersection-over-union ratio of the glass rod detection box and the funnel detection box being less than a preset threshold.

[0021] In some implementations, the interaction event condition includes that the intersection-union ratio of the glass rod detection frame and the funnel detection frame is greater than a preset threshold.

[0022] In some implementations, the conditions for the pouring action include detecting that the distance between the beaker detection frame and the funnel detection frame is less than a preset distance threshold, and that the change in the aspect ratio of the beaker detection frame is greater than a preset change threshold.

[0023] In some implementations, during the opportunistic segmentation, if the confidence score of the single-channel binary mask obtained from other image segmentations is higher than the cached confidence score, then all variables in the optimal segmentation result are updated.

[0024] In some implementations, the opposite-side drainage rule includes: if the orientation characteristics of the three-layer filter paper are consistent with the hand orientation of the beaker, the current drainage operation is determined to be in compliance with the specification; if the orientation characteristics of the three-layer filter paper are opposite to the hand orientation of the beaker, the current drainage operation is determined to have a same-direction pouring error.

[0025] The beneficial effects of this invention are as follows: By employing a state machine dynamic triggering mechanism, the optimal observation window is the stable existence of the funnel and the absence of glass rod obstruction. The single-channel binary mask of the three layers of filter paper is accurately extracted and cached, avoiding real-time segmentation noise. Then, the orientation of the three layers of filter paper and the contact position between the glass rod and the three layers of filter paper are determined in real time based on the mask. The ready state of the glass rod is identified by relying on dual-dimensional conditions. Real-time interactive analysis and timed segmentation are performed in staggered shifts, achieving extremely high real-time response performance and reducing hardware requirements. Finally, the opposite-side drainage rules are invoked to adjudicate compliance, achieving full-process automation. This ensures segmentation accuracy and judgment accuracy while reducing real-time computing power consumption, efficiently solving the core problems of existing technologies and providing reliable technical support for the evaluation of experimental operation standardization. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a visual inspection method for evaluating the standardization of flow in a filtering operation, as disclosed in an embodiment of the present invention.

[0027] Figure 2a This is a schematic diagram of the funnel detection frame;

[0028] Figure 2b This is a schematic diagram of the segmented single-channel binary mask;

[0029] Figure 3a A diagram illustrating a compliant lead generation procedure;

[0030] Figure 3b This is a schematic diagram of a drainage operation that suffers from an error of pouring water in the same direction. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the content of this invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to this invention are shown in the accompanying drawings, not all of them.

[0032] like Figure 1 As shown, this embodiment proposes a visual inspection method for evaluating the standardization of diversion in filtering operations, including the following steps:

[0033] Step 1: Obtain the real-time video stream of the target scene, identify the key objects in each frame of the real-time video stream, and encapsulate them as key information associated with the corresponding frame images.

[0034] In one example, after the host computer starts up, it controls the image acquisition device to acquire real-time video streams at a resolution of no less than 1920x1080 and a frame rate of 30fps. After the real-time video stream is sent to the host computer, target recognition is performed on a per-frame basis. In the target recognition step, real-time target detection algorithms such as YOLOv8 can be trained on a specific experimental equipment dataset to output information on all key objects in the scene in real time. In this case, because it is necessary to evaluate the diversion specifications in the filtering operation, the input training data needs to include objects such as funnels, glass rods, beakers, and hands. The key information output should include at least the detection boxes for the funnel, glass rod, beaker, and hand. The key information of these objects will be called in subsequent steps. In addition, this key information generally needs to include the image type, the corresponding detection box name, and the coordinates of the detection box, for example: [{'class': 'funnel', 'box': [x1, y1, x2, y2]}, {'class': 'glass_rod', 'box': [...]}, {'class': 'left_hand', 'box': [...]}].

[0035] Step 2: Treat the key information as a state machine. When the state machine meets the optimal observation window conditions for the traffic acquisition task, trigger the timing segmentation. When the state machine does not meet the optimal observation window conditions for the traffic acquisition task but meets the interaction event conditions, trigger the real-time interaction analysis. When the state machine has completed at least one real-time interaction analysis and meets the tilting action conditions, trigger the functional posture evaluation.

[0036] Step 2 of this solution is the core of the solution. Unlike existing technologies that use semantic segmentation for brute-force, continuous real-time segmentation, resulting in high costs and inability to handle occlusion, this solution first treats key information in the image frame as a state machine as input to logic gates. Then, it sets three conditions to enter different processes, enabling state transitions based on trigger conditions. The specific trigger conditions are as follows:

[0037] The optimal observation window conditions for the lead generation task include continuously detecting whether the current frame image simultaneously meets the first and second conditions. The first condition includes the stable presence of the funnel detection box in the image. The second condition includes the absence of the glass rod detection box, or the intersection-union ratio (IU) of the glass rod detection box and the funnel detection box being less than a preset threshold (e.g., 0.01). The core design goal of the optimal observation window is to acquire the features of the three layers of filter paper without obstruction. This window only exists in the stage before the glass rod interacts with the funnel. At this time, the experimental operation has not yet entered the core interaction stage of the glass rod contacting the funnel, which is the idle period before the operation. Its timing is completely staggered from subsequent interaction events, avoiding resource conflicts between the high-energy-consuming timing segmentation task in step 3 and the real-time interactive analysis task in step 4, while ensuring that feature extraction occurs in the operation stage most suitable for accurate segmentation.

[0038] The first condition mentioned above ensures that the cropping range always revolves around a valid and stable funnel region when performing the cropping task in the subsequent step 3. If the funnel target is unstable (e.g., its position shakes violently, or it partially falls out of frame), the cropped area may miss the funnel or contain a large amount of irrelevant background, causing the segmentation network to be unable to accurately locate the three-layer filter paper region within the funnel, thus outputting an invalid single-channel binary mask. This condition, from the perspective of spatial baseline stability, defines a reliable operating range for feature extraction. The essence of the second condition is to prevent the glass rod from obscuring the three layers of filter paper area inside the funnel by restricting its appearance in the image or by limiting the spatial overlap between the glass rod and the funnel. The three layers of filter paper are the core segmentation target of the segmentation network. If the glass rod obscures this area (e.g., the glass rod is against the funnel wall), the segmentation network will be unable to fully identify the pixel distribution of the three layers of filter paper. The output single-channel binary mask may miss the three-layer area or mistakenly identify the glass rod pixels as the three layers of filter paper, directly compromising the accuracy of the calculation of the contact point orientation features in the subsequent step 4. This condition, from the perspective of eliminating obstruction interference, provides an unobstructed image environment for feature extraction in step 4.

[0039] In summary, the timing-based segmentation triggered by the optimal observation window conditions of this traffic diversion task is a staggered capture mechanism. This cleverly shifts the originally energy-intensive segmentation task to a non-critical, unobstructed idle period before the experimental operation. This means that in the subsequent real-time interactive analysis phase, the hardware system does not need to perform any high-cost segmentation calculations, thus achieving extremely high real-time response performance and extremely low hardware requirements.

[0040] The interaction event conditions include that the intersection-union ratio (IoU) of the glass rod detection frame and the funnel detection frame is greater than a preset threshold (e.g., 0.05). A key prerequisite for the drainage operation is that the glass rod must be against the inner wall of the funnel (or close to the funnel opening) to guide the liquid into the funnel along the glass rod. If the glass rod only moves near the funnel (without contact, IoU ≤ threshold), it is considered an invalid approach, and therefore analysis is unnecessary. Thus, the essence of this interaction event condition is to filter out scenarios where the glass rod and funnel have made substantial contact or are in close proximity using the IoU threshold, ensuring that the real-time interaction analysis in subsequent step 4 only initiates analysis for interaction actions that may affect the compliance of the drainage process, avoiding misjudgments of meaningless glass rod movements. The triggering of real-time interactive analysis is based on the premise that "the state machine does not meet the optimal observation window condition of the lead generation task". In other words, the system will only check whether "the IoU between the glass rod and the funnel meets the standard" when the glass rod starts to approach or contact the funnel. Therefore, the interactive event condition and the optimal observation window condition of the lead generation task are completely mutually exclusive in the triggering logic and will not occur at the same time, ensuring that the system switches clearly between step 3 and step 4 and avoids process conflicts.

[0041] The conditions for a tilting action include: the distance between the beaker detection frame and the funnel detection frame is less than a preset distance threshold, and the change in the aspect ratio of the beaker detection frame is greater than a preset change threshold. The trigger condition of the distance between the beaker and funnel detection frames being less than the preset distance threshold ensures that the beaker is within a reasonable spatial range for pouring liquid into the funnel. If the beaker is too far from the funnel (distance ≥ threshold), even tilting will not pour the liquid into the funnel, constituting an invalid tilt. Furthermore, the change in the aspect ratio of the beaker detection frame being greater than the preset change threshold quantifies the degree of tilt of the beaker. Assuming the aspect ratio is stable when the beaker is upright (e.g., 2:1), tilting during liquid pouring (visually shortening the height and increasing the width) will significantly decrease the aspect ratio (e.g., to 1.2:1). A change exceeding the threshold indicates that the beaker has tilted. The combination of the two triggering conditions mentioned above can exclude non-tipping scenarios such as "the beaker is close to the funnel but not tilted" and "the beaker is tilted but too far from the funnel". Only the core actions that are in a reasonable position and are indeed being tilted will trigger the functional posture assessment in step 5.

[0042] Step 3, time-bound segmentation, includes: using the funnel detection box as a reference, cropping the target image region containing the funnel from the image and inputting it into the segmentation network. The network outputs a single-channel binary mask with an area equal to the target image region, along with a corresponding confidence score (scored based on feature sharpness and / or segmentation accuracy). Non-zero pixels in the single-channel binary mask represent the three-layer filter paper region. The funnel detection box, single-channel binary mask, and confidence score are cached as the optimal segmentation result. More preferably, in time-bound segmentation, if the confidence score of the single-channel binary mask obtained from other image segmentations is higher than the cached confidence score, all variables in the optimal segmentation result are updated. The segmentation network described above can be the U-Net segmentation network. The U-Net segmentation network requires pre-training before use, learning the mapping rules from the input image to the pixel-level segmentation mask through a large amount of labeled data. This lays the foundation for subsequent fine-tuning or direct use for segmentation tasks of laboratory instruments such as funnels, three-layer filter paper, and glass rods. Specific details are not elaborated here.

[0043] In step 3, obtaining a complete single-channel binary mask for the funnel and three layers of filter paper provides a 1:1 reproduction of the true shape of the target area. The fidelity and richness of the single-channel binary mask enable the subsequent analysis process in step 4, with accuracy and robustness far exceeding existing solutions. Furthermore, this solution avoids the need for analysis during occlusion by memorizing the required high-precision information before occlusion occurs, completely solving the dynamic occlusion problem faced by all real-time analysis solutions. Moreover, when the system captures a better single-channel binary mask (referring to a single-channel binary mask with a higher confidence score), it can be updated through overlay. In one example, the funnel detection box is as follows: Figure 2a As shown, the segmented single-channel binary mask is as follows: Figure 2b As shown, the confidence score at this point is 0.98.

[0044] Step 4, real-time interactive analysis includes: reading the optimal segmentation result; using the funnel detection frame as a reference, transforming the single-channel binary mask to the global coordinate system of the corresponding frame image; comparing the centroid of the transformed single-channel binary mask with the center point of the funnel detection frame to obtain the orientation features of the three-layer filter paper; detecting whether the corner pixels of the glass rod detection frame overlap with the single-channel binary mask; if so, outputting the orientation features of the contact point between the glass rod detection frame and the three-layer filter paper; and encapsulating the orientation features of the three-layer filter paper and the contact point orientation features into an interactive analysis result. Clearly, the premise for outputting the interactive analysis result is that the glass rod has already made contact with the three-layer filter paper. Although the glass rod has made contact with the three-layer filter paper, the error of "tilting in the same direction" cannot be ruled out. Therefore, this interactive analysis result will serve as the basis for judgment in the subsequent step 5.

[0045] In step 4, the optimal segmentation result needs to be loaded from the cache, using the pre-obtained optimal segmentation result as the basis for analysis. The reference system for the single-channel binary mask after segmentation in step 3 is its own coordinates, not the original image. Therefore, it needs to be transformed to the global coordinate system of the original image before feature extraction. In this step, the single-channel binary mask is first used to quantify the placement of the three layers of filter paper in the funnel, i.e., the orientation features of the three layers of filter paper. Then, the region of interest is extracted using the single-channel binary mask. Specifically, the corner pixels of the detection frame are checked for overlap with the single-channel binary mask using a glass rod. If there is no overlap, it indicates that the glass rod is not in contact with the filter paper, possibly indicating a violation of the rule of not using the glass rod for flow guidance, and no information is output. If there is overlap, the orientation features of the contact point are output (e.g., the contact point is in the upper / middle / lower region of the three layers of filter paper, or its left / right offset), serving as an implicit prerequisite for proceeding to step 5 (because step 5...). The trigger condition is "when the state machine has completed at least one real-time interactive analysis and meets the tilting action condition". The core value of the contact point orientation feature is to determine whether the glass rod accurately contacts the three-layer filter paper area. This is a basic prerequisite for the compliance of the entire drainage operation. If the contact point orientation feature output fails, it means that the glass rod contact position is outside the three-layer filter paper area, and no further interactive analysis results need to be generated.

[0046] Step 5, Functional Posture Assessment, includes: extracting the latest hand detection frame, beaker detection frame, and glass rod detection frame from key information; calculating the intersection-union ratio (IU) of the two hand detection frames with the beaker and glass rod detection frames respectively; determining the hand orientation attributes of the beaker and glass rod; invoking preset opposite-side drainage rules (an expert rule base can be set up to store these opposite-side drainage rules); and using the three-layer filter paper orientation features and beaker hand orientation attributes from the interaction analysis results as the criteria to determine whether the user's current drainage operation is compliant. The determination result is output as the final assessment result. When a tilting action occurs, the expert rules are applied for assessment, combining the analysis results from Stage 2 with the real-time hand position.

[0047] The rules for lateral drainage include: if the orientation of the three layers of filter paper is consistent with the hand orientation of the beaker, the current drainage operation is deemed to be in compliance with the standard; if the orientation of the three layers of filter paper is opposite to the hand orientation of the beaker, the current drainage operation is deemed to have a unidirectional pouring error.

[0048] It should be noted that in biochemical experiments, the basic standard for funnel filtration is that the glass rod should be leaned against the three layers of filter paper on the inner wall of the funnel. Therefore, assuming that the orientation of the three layers of filter paper and the hand position of the beaker both indicate the left side, it means that the three layers of filter paper and the beaker are on the operator's left side. In this case, the single layer of filter paper is undoubtedly on the right side. When the operator uses their left hand to hold the beaker and pour liquid into the funnel, since step 4 has already determined that the glass rod is in contact with the three layers of filter paper, the glass rod held in the right hand must be tilted and in contact with the three layers of filter paper, which conforms to the drainage standard. Figure 3a As shown in the figure, the beaker detection frame, funnel detection frame, and glass rod detection frame are displayed (the detection frame for the three-layer filter paper area is not shown). Conversely, assuming the orientation of the three-layer filter paper is shown as left and the beaker's hand orientation is shown as right, it means the three-layer filter paper is on the operator's left and the beaker is on the operator's right. In this case, the single-layer filter paper is undoubtedly on the right. When the operator uses their right hand to hold the beaker and pour the liquid into the funnel, since step 4 has already determined that the glass rod is in contact with the three-layer filter paper, the glass rod held in the left hand will necessarily be in vertical contact with the three-layer filter paper, resulting in a unidirectional pouring error. Figure 3b As shown. The above-mentioned opposite-side drainage rule only determines whether the current drainage operation conforms to the specification by judging the orientation characteristics of the three layers of filter paper and the orientation characteristics of the beaker.

[0049] The above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made based on the essence of the content of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A visual inspection method for evaluating the standardization of diversion in filtering operations, characterized in that, Includes the following steps: Acquire a real-time video stream of the target scene, identify key objects in each frame of the real-time video stream, and encapsulate them as key information associated with the corresponding frame images; The key information is regarded as a state machine. When the state machine meets the optimal observation window condition for the traffic diversion task, a timed segmentation is triggered. When the state machine does not meet the optimal observation window condition for the traffic diversion task but meets the interactive event condition, a real-time interactive analysis is triggered. When the state machine has completed at least one real-time interactive analysis and meets the tilting action condition, a functional posture evaluation is triggered. The timing-based segmentation includes: using the funnel detection box as a reference, cropping the target image region containing the funnel from the image, inputting it into the segmentation network, outputting a single-channel binary mask with the same area as the target image region, and the corresponding confidence score, wherein the non-zero pixels in the single-channel binary mask represent the three-layer filter paper region, and caching the funnel detection box, the single-channel binary mask and the confidence score as the optimal segmentation result; The real-time interactive analysis includes: reading the optimal segmentation result, using the funnel detection frame as a reference, transforming the single-channel binary mask to the global coordinate system of the corresponding frame image, comparing the centroid of the transformed single-channel binary mask with the center point of the funnel detection frame to obtain the orientation features of the three-layer filter paper, detecting whether the corner pixels of the glass rod detection frame overlap with the single-channel binary mask, and if so, outputting the orientation features of the contact point between the glass rod detection frame and the three-layer filter paper, and encapsulating the orientation features of the three-layer filter paper and the orientation features of the contact point into an interactive analysis result; The functional posture assessment includes: extracting the latest hand detection frame, beaker detection frame, and glass rod detection frame from the key information; calculating the intersection-union ratio (IUU) of the two hand detection frames with the beaker detection frame and the glass rod detection frame, respectively; determining the hand orientation attributes of the beaker and the glass rod; invoking preset opposite-side traffic diversion rules; using the three-layer filter paper orientation features and the hand orientation attributes of the beaker in the interaction analysis results as the criteria to determine whether the user's current traffic diversion operation is compliant; and outputting the determination result as the final assessment result.

2. The visual inspection method for evaluating the standardization of drainage operations as described in claim 1, characterized in that, The key information includes at least the funnel detection frame, the glass rod detection frame, the beaker detection frame, and the hand detection frame.

3. The visual inspection method for evaluating the standardization of drainage operations as described in claim 1, characterized in that, The optimal observation window conditions for the traffic diversion task include continuously detecting whether the current frame image simultaneously satisfies a first condition and a second condition. The first condition includes that the funnel detection box is stably present in the image, and the second condition includes that the glass rod detection box is not present, or that the intersection-union ratio of the glass rod detection box and the funnel detection box is less than a preset threshold.

4. The visual inspection method for evaluating the standardization of drainage operations as described in claim 1, characterized in that, The interactive event conditions include that the intersection-union ratio of the glass rod detection frame and the funnel detection frame is greater than a preset threshold.

5. The visual inspection method for evaluating the standardization of drainage operations as described in claim 1, characterized in that, The conditions for the tilting action include detecting that the distance between the beaker detection frame and the funnel detection frame is less than a preset distance threshold, and that the change in the aspect ratio of the beaker detection frame is greater than a preset change threshold.

6. The visual inspection method for evaluating the standardization of drainage operations as described in claim 1, characterized in that, In the timed segmentation, if the confidence score of the single-channel binary mask obtained from other image segmentation is higher than the cached confidence score, then all variables in the optimal segmentation result are updated.

7. The visual inspection method for evaluating the standardization of drainage operations as described in claim 1, characterized in that, The opposite-side drainage rules include: if the orientation characteristics of the three-layer filter paper are consistent with the hand orientation of the beaker, the current drainage operation is determined to be in compliance with the specifications; if the orientation characteristics of the three-layer filter paper are opposite to the hand orientation of the beaker, the current drainage operation is determined to have a same-direction pouring error.

Images