Recommended method for collecting samples in dynamic update of sample library

Abstract

The present invention relates to the field of artificial intelligence, particularly intelligent recognition technology, and more specifically to a method for recommending samples collected during dynamic updates of a sample library. By screening and analyzing recognition results, the method reversely generates recommendations for samples in the sample library that require updating, and generates different recommendation levels based on different situations. This method can then guide operations and maintenance personnel in rationally arranging sample library updates, ensuring the effective operation of the recognition system, reducing operational intensity, and lowering costs.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, in particular to the field of intelligent recognition technology, and more specifically to a method for recommending collected samples in the dynamic update of a sample library. Background Art

[0002] Intelligent recognition is a key functional category in the field of artificial intelligence. The sample library is a crucial component of intelligent recognition and plays a decisive role in the accuracy of recognition results. If there are problems with the samples used for comparison in the sample library, the accuracy of the recognition results will inevitably be affected.

[0003] In some fields, the objects being identified can change due to various factors. If the samples in the sample library are not updated in a timely manner, errors may occur when identifying the corresponding objects. Currently, this type of operation is mainly initiated through manual intervention. However, when faced with large amounts of sample data, this proactive manual intervention method is not always effective.

[0004] For example, with the continued advancement of smart healthcare, automated medication dispensing in hospital pharmacies is an inevitable trend. This automated dispensing process requires verification of the medications to be dispensed, which must be identified by comparing them with samples in a library. However, in the pharmaceutical sector, due to the large number of drug manufacturers and the irregular schedule for new product specifications and packaging changes, it is impossible to obtain this information one by one.

[0005] Therefore, obtaining updated sample information for a specific object to be identified, and then resampling and updating the object's sample library information, is a key technical node for improving recognition accuracy and reducing operational maintenance efforts. This has also become a hot topic for researchers in this field and a technical difficulty that needs to be overcome. Summary of the Invention

[0006] The technical problem to be solved by the present invention is how to obtain the sample information that needs to be updated and collected in the sample library when the packaging change of the object to be identified cannot be known in advance, and dynamically collect and improve the information of the samples in the sample library accordingly.

[0007] In order to solve the above technical problems, the present invention discloses a method for recommending collected samples in a dynamic update of a sample library, comprising the following steps:

[0008] S1: Create a time series queue based on the recognition results of all recognition tasks within the set time interval according to the task completion time;

[0009] S2: Recommend the newly entered identification objects as a new update, and remove the identification results corresponding to these newly entered identification objects from the time series queue established in S1;

[0010] S3: Remove the known identification objects that do not rely on the sample library from the queue of S1;

[0011] S4: using the trimmed time series queue of the remaining recognition results as a candidate update queue, and creating an empty queue for each sample of all candidate update recognition objects in the candidate update queue;

[0012] S5: Take a candidate update recognition object as the target, traverse the candidate update queue in S4, and when the target appears, record the recognition time, the target image and the recognition result of the corresponding sample, and record the recognition result k 标 To identify time t 标 The target is recorded in the empty queue established in S4 in order, where the recognition result k 标 It includes two values, m and g. When the recognition result is a hit, it is marked as g, and when the recognition result is a miss, it is marked as m.

[0013] S6: After processing in S5, the empty queues of each candidate update recognition object established in S4 are replaced by the recognition result k 标 Time t 标 Fill in the order to form the identification statistics queue of each candidate update identification object;

[0014] S7: Select a relatively stable window period in the identification statistics queue as the starting segment of the sliding window, and slide the sliding window backward until the current period;

[0015] S8: Calculate k in each sliding window separately 标 The value is m, and the recommendation results are generated as follows:

[0016] M1: If the proportion of m is on an upward trend and has stabilized, output the second-level update recommendation;

[0017] M2: If the proportion of m is on an upward trend and is still in the rising stage, output the third-level update recommendation;

[0018] M3: If the proportion of m does not show a clear upward trend, no update recommendation is output;

[0019] S9: The system maintenance personnel updates the recommended candidate update identification objects according to the update recommendation data in the update priority recommendation order of first-tier update recommendation, second-tier update recommendation, and third-tier update recommendation;

[0020] The order of S2 and S3 can be swapped. That is, there is no order in which the newly entered objects are removed after recommendation and the objects that are not identified by the sample library are removed.

[0021] Further preferably, the present invention discloses three different methods of using the newly stored identification object as the update first-level recommendation in step S2.

[0022] The first method includes the following steps:

[0023] K1: Create a time series queue based on the task completion time for all recognition results of different recognition objects in the set time interval;

[0024] K2: Extract all the identified objects that appear in the time series queue and create an empty queue for each identified object;

[0025] K3: Take each recognition object as the target, traverse the time series queue established in K1, and when the target appears, record the recognition time and the corresponding recognition result, and record the recognition result D 标 To identify time T 标 In order, record in the empty queue established by K2, where D 标 The value of includes two, 1 and -1. When the recognition result is a hit, it is marked as 1, and when the recognition result is a miss, it is marked as -1;

[0026] K4: After all recognition results are collected, a recognition result queue of each recognition object is formed in order of recognition time;

[0027] K5: Traverse the recognition result queue in K4. When the recognition results of a certain recognition object are all -1 from the starting end, the recognition object is selected and recommended as the update level.

[0028] When the order of S2 and S3 is swapped, the newly stored recognition objects in S2 can be combined with other sample recommendation steps as the step of updating the first level recommendation. At this time, for the convenience of labeling, the labeling method is unified. For example, k 标 Instead of D 标 , with t 标 Instead of T 标 , the recognition results originally marked as 1 representing hits are also marked as g, and the recognition results originally marked as -1 representing misses are also marked as m.

[0029] Since this method still adopts the reverse recommendation method, in order to save computing time and computing resources, it can be combined with other sample recommendation method steps, and only processed separately when judging the results and outputting the recommendations.

[0030] The second method is to manually recommend the known newly entered recognition objects as the update first-level recommendations.

[0031] The third method is to compare all the recognition objects in the recognition task with the names of the recognition objects established in the sample library, and recommend the recognition objects that do not match as the first-level update.

[0032] These three methods can be used individually or in combination.

[0033] Further preferably, when the identification object is a pharmaceutical product, the known identification object that does not rely on sample library identification is a pharmaceutical product that requires manual review. For example, in the field of pharmaceutical identification, transparent pharmaceutical products such as infusion bags cannot be automatically identified and therefore do not conform to the sample library identification method and must be manually reviewed.

[0034] Furthermore, the present invention also discloses that the width of the sliding window is obtained through machine learning in an adaptive manner based on the characteristics of the recognition statistical queue of each candidate updated recognition object.

[0035] Of course, the width of the sliding window may also be manually set by a technician based on experience or the like.

[0036] Furthermore, the present invention also discloses that the method for matching the target image identified this time with its corresponding sample in step S5 is to match through similarity comparison. When the similarity reaches a set threshold, it is considered a match; when the similarity is lower than the set threshold, it is considered that a certain sample cannot be matched. At this time, the recognition result of the target image identified this time is marked to all samples in the sample library of the corresponding recognition object.

[0037] This method can effectively solve the system problem caused by the inability to mark the recognition results when the target image cannot be matched with any sample in some recognition scenarios.

[0038] After adopting the technical solution disclosed in the present invention, it is possible to reversely generate sample recommendations that need to be updated in the sample library based on the recognition results, and generate different recommendation levels according to different situations, thereby guiding operation and maintenance personnel to reasonably arrange the update of the sample library, ensure the effective operation of the recognition system, reduce operation and maintenance intensity, and reduce operation and maintenance costs. DETAILED DESCRIPTION

[0039] In order to better understand the present invention, the present invention is further described below with reference to specific embodiments.

[0040] Example 1

[0041] Taking drug identification and verification as an example, the recommended method for collecting samples in the dynamic update of the sample library disclosed in this embodiment includes the following steps:

[0042] S1: Create a time series queue based on the recognition results of all recognition tasks within the set time interval according to the task completion time;

[0043] S2: Recommend the newly entered identification objects as a new update, and remove the identification results corresponding to these newly entered identification objects from the time series queue established in S1;

[0044] S3: Remove the known identification objects that do not rely on the sample library from the queue of S1;

[0045] S4: using the trimmed time series queue of the remaining recognition results as a candidate update queue, and creating an empty queue for each sample of all candidate update recognition objects in the candidate update queue;

[0046] S5: Take a candidate update recognition object as the target, traverse the candidate update queue in S4, and when the target appears, record the recognition time, the target image and the recognition result of the corresponding sample, and record the recognition result k 标 To identify time t 标 The target is recorded in the empty queue established in S4 in order, where the recognition result k 标 It includes two values, m and g. When the recognition result is a hit, it is marked as g, and when the recognition result is a miss, it is marked as m.

[0047] S6: After processing in S5, the empty queues of each candidate update identification object established in S4 are filled with multiple identification time t 标 And the recognition result k 标 Filling to form the identification statistics queue of each candidate updated identification object;

[0048] S7: Select a relatively stable window period in the identification statistics queue as the starting segment of the sliding window, and slide the sliding window backward until the current period;

[0049] S8: Calculate k in each sliding window separately 标 The value is m, and the recommendation results are generated as follows:

[0050] M1: If the proportion of m is on an upward trend and has stabilized, output the second-level update recommendation;

[0051] M2: If the proportion of m is on an upward trend and is still in the rising stage, output the third-level update recommendation;

[0052] M3: If the proportion of m does not show a clear upward trend, no update recommendation is output;

[0053] S9: The system maintenance personnel updates the recommended candidate update identification objects according to the update recommendation data in the update priority recommendation order of first-tier update recommendation - second-tier update recommendation - third-tier update recommendation.

[0054] Since there is no order between S2 and S3, in order to facilitate operation, in this embodiment, S3 is performed first, that is, the objects that are not identified by the sample library are eliminated, and then the newly entered objects are recommended and eliminated.

[0055] In this embodiment, preferably, in step S2, a reverse automatic recommendation method is used to recommend the newly stored identification object as an update file. The specific method includes the following steps:

[0056] K1: Create a time series queue based on the task completion time for all recognition results of different recognition objects in the set time interval;

[0057] K2: Extract all the identified objects that appear in the time series queue and create an empty queue for each identified object;

[0058] K3: Take each recognition object as the target, traverse the time series queue established in K1, and when the target appears, record the recognition time and the corresponding recognition result, and record the recognition result D 标 To identify time T 标 In order, record in the empty queue established by K2, where D 标 The value of includes two, 1 and -1. When the recognition result is a hit, it is marked as 1, and when the recognition result is a miss, it is marked as -1;

[0059] K4: After all recognition results are collected, a recognition result queue of each recognition object is formed in order of recognition time;

[0060] K5: Traverse the recognition result queue in K4. When the recognition results of a certain recognition object are all -1 from the starting end, the recognition object is selected and recommended as the update level.

[0061] Since the above-mentioned method of recommending new objects for identification has the basis for merging processing after the order is changed, in this embodiment, the method steps for recommending other samples are merged with the method steps for recommending new objects for identification, and are only processed separately when judging the results and outputting the recommendations.

[0062] For example, in this embodiment, the identification results of all drug identification tasks within an update monitoring cycle are firstly used to establish a time series queue according to the task completion time; then, the drugs that are determined to be required to be manually reviewed are removed from this time series queue.

[0063] Extract all the identified objects that appear in the time series queue after the previous step of elimination, and create an empty queue for each identified object;

[0064] Take a certain recognition object as the target, traverse the queue, and when the target appears, record the recognition time, the target image and the recognition result of the corresponding sample, and record the recognition result k 标 To identify time t 标 The recognition result k is recorded in the empty queue of the target. 标 It includes two values, m and g. When the recognition result is a hit, it is marked as g, and when the recognition result is a miss, it is marked as m. Here, due to the merging of new recognition objects, k is also used. 标 Instead of D 标 , with t 标 Instead of T 标 , mark the hit recognition result originally marked as 1 as g, and mark the miss recognition result originally marked as -1 as m;

[0065] When all recognition results are collected, a recognition result queue of each recognition object is formed in order of recognition time;

[0066] The recognition result queue is traversed. When the recognition results of a certain recognition object are all m from the starting end, the recognition object is selected and recommended as the next level of update.

[0067] Then traverse the recognition result queue of the remaining recognition objects and calculate the k 标 The value is m, and the recommendation results are generated as follows:

[0068] M1: If the proportion of m is on an upward trend and has stabilized, output the second-level update recommendation;

[0069] M2: If the proportion of m is on an upward trend and is still in the rising stage, output the third-level update recommendation;

[0070] M3: If the proportion of m does not show a clear upward trend, no update recommendation is output;

[0071] Finally, the system maintenance personnel updates the recommended candidate update identification objects according to the update recommendation data in the update priority recommendation order of first-tier update recommendation - second-tier update recommendation - third-tier update recommendation.

[0072] Because the candidate update objects already contain the object name and sample name used for identification, such as "XX Capsules - ** Company - 12 capsules x 3 plates" + "front," system maintenance personnel can quickly find samples to update based on the recommended information. Furthermore, new sample information that has never been entered into the sample database is displayed as unidentified images.

[0073] The above is a specific embodiment of the present invention. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A recommended method for collecting samples in a dynamic update of a sample library, characterized in that: The following steps are involved: S1: Create a time series queue based on the recognition results of all recognition tasks within the set time interval according to the task completion time; S2: Recommend the newly entered identification objects as a new update, and remove the identification results corresponding to these newly entered identification objects from the time series queue established in S1; S3: Remove the known identification objects that do not rely on the sample library from the queue of S1; S4: using the trimmed time series queue of the remaining recognition results as a candidate update queue, and creating an empty queue for each sample of all candidate update recognition objects in the candidate update queue; S5: Take a candidate update recognition object as the target, traverse the candidate update queue in S4, and when the target appears, record the recognition time and the recognized target image and the recognition result of the corresponding sample, and then send the recognition result k 标 To identify time t 标 The target is recorded in the empty queue established in S4 in order, where the recognition result k 标 It includes two values, m and g. When the recognition result is a hit, it is marked as g, and when the recognition result is a miss, it is marked as m. S6: After processing in S5, the empty queues of each candidate update recognition object established in S4 are replaced by the recognition result k 标 Time t 标 Fill in the order to form the identification statistics queue of each candidate update identification object; S7: Select a stable window period in the identification statistics queue as the starting segment of the sliding window, and slide the sliding window backward until the current period; S8: Calculate k in each sliding window separately 标 The value is m, and the recommendation results are generated as follows: M1: If the proportion of m is on an upward trend and has stabilized, output the second-level update recommendation; M2: If the proportion of m is on an upward trend and is still in the rising stage, output the third-level update recommendation; M3: If the proportion of m does not show a clear upward trend, no update recommendation is output; S9: The system maintenance personnel updates the recommended candidate update identification objects according to the update recommendation data in the order of first-tier update recommendation, second-tier update recommendation, and third-tier update recommendation. The order of S2 and S3 can be swapped.

2. The method for recommending collected samples in dynamic update of a sample library according to claim 1, characterized in that: The method for recommending the newly stored identification object as a first-level update in step S2 includes the following steps: K1: Create a time series queue based on the task completion time for all recognition results of different recognition objects in the set time interval; K2: Extract all the identified objects that appear in the time series queue and create an empty queue for each identified object; K3: Take each recognition object as the target, traverse the time series queue established in K1, and when the target appears, record the recognition time and the corresponding recognition result, and record the recognition result D 标 To identify time T 标 In order, record in the empty queue established by K2, where D 标 The value of includes two, 1 and -1. When the recognition result is a hit, it is marked as 1, and when the recognition result is a miss, it is marked as -1; K4: After all recognition results are collected, a recognition result queue of each recognition object is formed in order of recognition time; K5: Traverse the recognition result queue in K4. When the recognition results of a certain recognition object are all -1 from the beginning, the recognition object is selected and recommended as the first-level update; When the order of S2 and S3 is swapped, the newly stored identification object in S2 is merged with the method steps of other sample recommendations as a first-level update recommendation step. At this time, for the convenience of marking, the marking method is unified.

3. The method for recommending collected samples in dynamic update of a sample library according to claim 1, characterized in that: In step S2, the method of recommending the newly stored identification objects as the first-level update is to manually recommend the known newly stored identification objects as the first-level update.

4. The method for recommending collected samples in dynamic update of a sample library according to claim 1, characterized in that: In step S2, the method of recommending the newly entered recognition objects as the first-level update is to compare all the recognition objects in the recognition task with the names of the recognition objects established in the sample library, and recommending the recognition objects that do not match as the first-level update.

5. The method for recommending collected samples in dynamic update of a sample library according to claim 1, characterized in that: When the identification object is a drug, the known identification object that does not rely on sample library identification is a drug that must be manually reviewed.

6. The method for recommending collected samples in dynamic update of a sample library according to claim 1, characterized in that: The width of the sliding window is adaptively obtained by machine learning or set by manual experience according to the characteristics of the identification statistical queue of each candidate updated identification object.

7. The method for recommending collected samples in dynamic update of a sample library according to claim 1, characterized in that: In step S5, the method of matching the identified target image with its corresponding sample is to match them through similarity comparison. When the similarity reaches the set threshold, it is considered a match; when the similarity is lower than the set threshold, it is considered that a certain sample cannot be matched. At this time, the recognition result of the identified target image is marked to all samples in the sample library of the corresponding recognition object.