A system and method for adjusting a recognition algorithm after recognizing a scene through intelligent detection

Abstract

The application relates to the technical field of intelligent detection, and discloses a system and a method for adjusting a recognition algorithm after intelligently detecting and recognizing a scene, wherein the method comprises the following steps: acquiring current frame detection data and extracting multi-dimensional scene feature descriptors, and maintaining a scene descriptor time sequence cache queue; performing time sequence difference analysis on the queue to generate a scene mutation prediction signal; based on the signal state, pre-fetching a target scene recognition algorithm parameter set to a pipeline preparation register; when a new frame arrives, promoting the parameter set in the preparation register to an effective parameter and performing feature separation operation; performing multi-dimensional target recognition on each feature response separation graph and calculating a weighted fusion confidence score; screening a candidate target region and outputting a detection and recognition result.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection technology, and more specifically to a system and method for adjusting the recognition algorithm after identifying a scene through intelligent detection. Background Technology

[0002] In intelligent detection applications, detection systems need to handle various types of scenarios, each corresponding to different object categories, environmental conditions, or material characteristics. Current detection processes extract scene feature descriptors from each frame of detection data, obtain scene category labels through scene classification, and then retrieve the corresponding recognition algorithm parameter configurations from the algorithm configuration library based on the scene category labels.

[0003] The above approach has the following drawbacks. First, scene classification calculations require the complete extraction and analysis of multi-dimensional scene features in the current frame, resulting in a computational delay of several frames. When a scene change occurs between consecutive frames, the scene classification results have not yet been updated, and the recognition algorithm still uses the parameter configuration of the previous scene to process the detection data of the new scene, leading to parameter mismatch in the transition frames. Effective suppression strategies in the previous scene may incorrectly filter real target signals in the new scene, and background features preserved in the previous scene may be misclassified as target candidates in the new scene. Second, the exact frame position of the scene change is unpredictable due to fluctuations in actual detection conditions, and a switch point cannot be preset using a fixed frame number. Third, performing complete scene feature extraction frame by frame consumes excessive computational resources under high frame rate detection conditions, compressing the available time window for recognition and inference. Summary of the Invention

[0004] This invention provides a system and method for adjusting the recognition algorithm after intelligent scene detection, solving the technical problem in related technologies where the recognition algorithm cannot adaptively switch parameters according to scene changes, leading to a decrease in cross-scene detection accuracy.

[0005] This invention discloses a method for adjusting the recognition algorithm after intelligent scene detection and recognition, comprising: extracting multi-dimensional scene feature descriptors from the current frame detection data and appending them to the scene descriptor temporal cache queue; Calculate the first-order difference sequence for each dimension descriptor sequence in the time-series cache queue, compare the absolute value of the first-order difference of each dimension with the scene change threshold of the corresponding dimension, and generate a scene change prediction signal. When the scene mutation prediction signal is in an active state, the target scene's recognition algorithm parameter set is retrieved from the recognition algorithm configuration library and written into the pipeline preparation register; when it is in an inactive state, the current recognition algorithm parameters are maintained. When a new frame of detection data enters the recognition processing pipeline, the set of recognition algorithm parameters in the pipeline preparation register is promoted to the current valid parameters. Based on the valid parameters, feature separation operation is performed on the detection data to generate a feature response separation map. Target recognition processing is performed on each feature response separation map, and the weighted fusion confidence score of each candidate target region is calculated. Candidate target regions whose weighted fusion confidence scores exceed the scene adaptive judgment threshold are selected, and the detection and recognition results are output.

[0006] Furthermore, the multidimensional scene feature descriptor includes a texture complexity index, an intensity distribution statistic, and a structural pattern consistency index. The texture complexity index characterizes the degree of intensity of texture changes within the detection area of ​​the current frame, the intensity distribution statistic characterizes the concentrated or discrete distribution of signal intensity in the detection data, and the structural pattern consistency index characterizes the regularity of repetitive structural patterns in the current frame. After extracting the multidimensional scene feature descriptors, Z-score normalization is applied to each dimension descriptor to eliminate the dimensional differences between the dimensions.

[0007] Furthermore, the generation of the scene mutation prediction signal includes: when the absolute value of the first-order difference of two consecutive frames in any dimension exceeds the scene mutation threshold of the corresponding dimension, the scene mutation prediction signal is set to an active state. When the absolute value of the first difference in each dimension does not meet the above conditions, the scene mutation prediction signal is set to an inactive state.

[0008] Furthermore, it also includes: calculating a second-order difference sequence for each dimension descriptor sequence, wherein the second-order difference sequence is the difference between the first-order differences of two adjacent frames; When the second difference of a certain dimension is continuously positive, the mutation direction identifier of that dimension is set to the upward direction; when the second difference of a certain dimension is continuously negative, the mutation direction identifier of that dimension is set to the downward direction; the mutation direction identifier is used to determine the parameter retrieval direction of the target scene when retrieving the recognition algorithm parameter set from the recognition algorithm configuration library.

[0009] Furthermore, the recognition algorithm parameter set includes feature extraction range configuration, multi-algorithm weight configuration, and sensitivity threshold configuration. The feature extraction range configuration defines the detection dimensions and spatial range covered when performing feature separation operations on the target scene. The multi-algorithm weight configuration defines the fusion weight allocation of the target recognition algorithm output results on each detection dimension. The sensitivity threshold configuration defines the scene adaptive judgment threshold under the target scene conditions.

[0010] Furthermore, the pipeline preparation register is a buffer area independent of the current valid parameter register. The prefetching process of the recognition algorithm parameter set is executed in parallel with the recognition processing process of the current frame. When a new frame enters the recognition processing pipeline, the recognition algorithm parameter set in the pipeline preparation register is directly promoted to valid parameters, thereby realizing the synchronization of parameter switching and frame data arrival.

[0011] Furthermore, when the scene mutation prediction signal is in an inactive state, the multi-channel detection data of the current frame is spatially downsampled according to a preset downsampling ratio, a simplified scene descriptor is extracted from the downsampled multi-channel detection data, and the simplified scene descriptor is added to the scene descriptor temporal cache queue to replace the multi-dimensional scene feature descriptor.

[0012] Furthermore, the target recognition process includes morphological connected component analysis, phase-consistent edge detection, and local contrast anomaly detection; The calculation of the weighted fusion confidence score of each candidate target region includes: normalizing the confidence scores of each candidate target region under each recognition algorithm using mean normalization based on the range to uniformly map them to the interval between zero and one; and weighting and summing the normalized confidence scores according to the fusion weights corresponding to each recognition algorithm in the multi-algorithm weight configuration to obtain the weighted fusion confidence score of each candidate target region. The sum of the fusion weights of each recognition algorithm is one, and each fusion weight is greater than or equal to zero.

[0013] Furthermore, when the current frame is within a preset transition period frame number after the scene mutation prediction signal is activated, the scene adaptive judgment threshold in the sensitivity threshold configuration is multiplied by a relaxation coefficient, the relaxation coefficient being a preset value greater than zero and less than one, so that the judgment threshold actually used during the transition period is lower than the standard scene adaptive judgment threshold. After the transition period ends, the scene adaptive judgment threshold will be restored to the standard configuration value of the target scene.

[0014] This invention provides a system for adjusting the recognition algorithm after intelligently detecting and recognizing a scene, comprising: The scene feature extraction module is used to extract multi-dimensional scene feature descriptors from the detection data of the current frame and append the multi-dimensional scene feature descriptors to the scene descriptor temporal cache queue. The temporal difference analysis module is used to calculate the first-order difference sequence of each dimension descriptor sequence in the temporal cache queue, compare the absolute value of the first-order difference of each dimension with the scene change threshold of the corresponding dimension, and generate a scene change prediction signal. The parameter prefetching module is used to retrieve the target scene recognition algorithm parameter set from the recognition algorithm configuration library and write it into the pipeline pre-register when the scene mutation prediction signal is in an active state; when it is in an inactive state, it maintains the current recognition algorithm parameters. The feature separation module is used to upgrade the set of recognition algorithm parameters in the pipeline preparation register to the current valid parameters when a new frame of detection data enters the recognition processing pipeline, and to perform feature separation operation on the detection data according to the valid parameters to generate a feature response separation map. The target recognition module is used to perform target recognition processing on each feature response separation map and calculate the weighted fusion confidence score of each candidate target region; The result filtering module is used to filter candidate target regions whose weighted fusion confidence scores exceed the scene adaptive judgment threshold and output the detection and recognition results.

[0015] This invention performs first-order difference analysis on the historical frame descriptor sequence in the scene descriptor temporal cache queue to identify the mutation trend and generate a scene mutation prediction signal before the actual occurrence of scene mutations. It also preloads the target scene recognition algorithm parameter set into the pipeline pre-register. Since the parameter set prefetching process is executed in parallel with the current frame's recognition processing, the matching algorithm configuration can be obtained when a new scene frame arrives without waiting for scene classification calculations to complete. This eliminates the problem of parameter mismatch in transition frames caused by scene classification calculation delays, thereby avoiding misfiltering and misjudgment of new scene detection data caused by previous scene parameters.

[0016] Using the absolute value of the first-order difference of two consecutive frames exceeding the scene abrupt change threshold as the activation condition, instantaneous noise disturbances are distinguished from persistent scene shifts, reducing the probability of false responses to random fluctuations. The abrupt change direction identifier determined based on the second-order difference sequence provides directional constraints for parameter retrieval, narrowing the search range of candidate scenes.

[0017] In the steady-state stage where the scene mutation prediction signal is inactive, downsampling is used to simplify the scene descriptor calculation instead of extracting the complete multi-dimensional scene feature descriptor. This reduces the computational load of the scene evaluation stage and allocates the freed computational resources to the recognition and reasoning stage, thus alleviating the problem of scene evaluation and recognition and reasoning competing for computational time windows under high frame rate detection conditions.

[0018] Before the sudden change of scene During the frame transition period, the scene adaptively determines the threshold. Apply relaxation factor This makes the actual judgment threshold Below the standard threshold, of which This allows for a moderate reduction in the filtering threshold within the boundary range of switching between old and new scene parameters, thus avoiding the incorrect filtering of real targets due to boundary effects. Attached Figure Description

[0019] Figure 1 This is a flowchart of a method for adjusting the recognition algorithm after intelligently detecting and recognizing a scene, as provided in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the changing trends of various dimensions of the scene descriptor time-series cache queue provided in an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the absolute value of the first-order difference in each dimension with the mutation threshold provided in the embodiments of the present invention; Figure 4 This is a schematic diagram comparing the first-order difference and the second-order difference of texture complexity dimension provided in an embodiment of the present invention; Figure 5 This is a schematic diagram comparing the weight configurations of the scene recognition algorithms for ceramic parts and metal parts provided in this embodiment of the invention; Figure 6 This is a schematic diagram of the confidence scores of each algorithm and the weighted fusion confidence score of the candidate target region provided in the embodiments of the present invention; Figure 7 This is a schematic diagram of the transition period determination threshold and candidate region confidence screening provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of the scene descriptor time-series cache queue heatmap provided in an embodiment of the present invention; Figure 9 This is a schematic diagram comparing the allocation of computational resources between the non-mutation steady-state stage and the mutation activation stage provided in an embodiment of the present invention. Detailed Implementation

[0020] Example 1 In intelligent detection applications, detection systems need to handle various types of scenarios, each corresponding to different object categories, environmental conditions, or material characteristics. The current detection process extracts scene feature descriptors from each frame of detection data, obtains scene category labels through scene classification, and then retrieves the corresponding recognition algorithm parameter configuration from the algorithm configuration library based on the scene category labels. However, scene classification calculation requires the complete extraction and analysis of multi-dimensional scene features in the current frame, resulting in a computational delay of several frames. When a scene abruptly changes between consecutive frames, the scene classification result has not yet been updated, and the recognition algorithm still uses the parameter configuration of the previous scene to process the detection data of the new scene. Effective suppression strategies in the previous scene may incorrectly filter true target signals in the new scene, and background features preserved in the previous scene may be misclassified as target candidates in the new scene. Furthermore, the exact frame position of the scene abrupt change is unpredictable due to fluctuations in actual detection conditions, making it impossible to preset a switching point using a fixed frame number. Performing complete scene feature extraction frame by frame consumes excessive computational resources under high frame rate detection conditions, compressing the available time window for recognition inference.

[0021] A method for adjusting the recognition algorithm after intelligently detecting and recognizing a scene, according to an embodiment of the present invention, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain the detection data of the current frame and extract the multi-dimensional scene feature descriptor.

[0022] The system acquires multi-channel detection data for the current frame, extracts multi-dimensional scene feature descriptors from the multi-channel detection data, and appends the multi-dimensional scene feature descriptors to the scene descriptor temporal cache queue. The scene descriptor temporal cache queue is a first-in-first-out (FIFO) storage queue used to maintain the most recently accessed descriptor. The multidimensional scene feature descriptor of the frame, where This is the length of the cache window. When the number of descriptor frames in the queue exceeds... At that time, the descriptor of the earliest frame is removed from the queue.

[0023] It should be noted that the multidimensional scene feature descriptor includes a texture complexity index, an intensity distribution statistic, and a structural pattern consistency index. The texture complexity index characterizes the drasticness of texture changes within the detection region of the current frame. The intensity distribution statistic characterizes the concentrated or discrete distribution of signal intensity in the detection data. The structural pattern consistency index characterizes the regularity of repetitive structural patterns in the current frame. These three elements—texture complexity index, intensity distribution statistic, and structural pattern consistency index—describe the scene attributes of the current frame from different dimensions, providing multidimensional input for subsequent temporal difference analysis. After extracting the multidimensional scene feature descriptor, Z-score standardization is applied to each dimension to eliminate the dimensional differences between the texture complexity index, intensity distribution statistic, and structural pattern consistency index, ensuring comparability of the descriptors in subsequent temporal difference analysis.

[0024] Step 2: Perform temporal differential analysis on the scene descriptor temporal cache queue to generate scene mutation prediction signals.

[0025] The most recent scene descriptor in the time-series cache queue The first-order difference sequence is calculated for each dimension of the frame's descriptor sequence, where The buffer window length is used. The absolute values ​​of the first-order differences in each dimension are compared with preset scene mutation thresholds. When the absolute values ​​of the first-order differences for two consecutive frames in any dimension both exceed the scene mutation threshold for that dimension, a scene mutation prediction signal is generated and set to the active state. When the absolute values ​​of the first-order differences in each dimension do not meet the above condition, the scene mutation prediction signal is set to the inactive state.

[0026] It should be noted that, let the first... The dimension in the first The frame descriptor value is Then this dimension is in the th dimension. The first difference of a frame is divided into ,in For dimensional indexing, For frame index, For this dimension in the th The frame descriptor value. The activation condition is: the existence of a certain dimension. , making and Established at the same time, among which For the first Each dimension corresponds to a scene mutation threshold. Using the absolute value of the first-order difference of two consecutive frames exceeding the scene mutation threshold as the activation condition is to distinguish between instantaneous noise disturbances and persistent scene mutations. Fluctuations in the descriptor of a single frame may be caused by random interference in the detection data, while significant changes in two consecutive frames indicate a directional and continuous shift in scene features, thereby reducing the probability of false activations.

[0027] It should be noted that a second-order difference sequence is also calculated for each dimension's descriptor sequence, and the mutation direction identifier is determined based on the sign of the second-order difference sequence. The dimension in the first The second difference of a frame is divided into ,in For the first The dimension in the first First-order difference of frames, For the first The dimension in the first The first-order difference of a frame. When a certain dimension... When the value remains positive, it indicates that the rate of change of that dimension descriptor is still increasing, the complexity of the scene features is rising, and the direction of mutation is indicated by the upward direction. A consistently negative value indicates that the rate of change is converging, the complexity of scene features is decreasing, and the mutation direction is identified as the decreasing direction. The mutation direction identifier is used to determine the parameter retrieval direction of the target scene in subsequent steps.

[0028] Step 3: Based on the state of the scene change prediction signal, perform prefetching or maintenance processing of the recognition algorithm parameters.

[0029] When the scene mutation prediction signal is in an active state, based on the mutation direction identifier and the changing trend of the current multi-dimensional scene feature descriptor, the recognition algorithm parameter set corresponding to the target scene is retrieved from the recognition algorithm configuration library and written into the pipeline preparation register. When the scene mutation prediction signal is in an inactive state, the current recognition algorithm parameters remain unchanged.

[0030] It should be noted that the recognition algorithm parameter set includes feature extraction range configuration, multi-algorithm weight configuration, and sensitivity threshold configuration. The feature extraction range configuration defines the detection dimensions and spatial range covered when performing feature separation operations on the target scene. The multi-algorithm weight configuration defines the fusion weight allocation of the target recognition algorithm outputs across each detection dimension. The sensitivity threshold configuration defines the scene-adaptive decision threshold for determining whether a candidate target region is a real target under the target scene conditions.

[0031] It should be noted that the pipeline pre-register is a buffer area independent of the currently valid parameter register. After the scene change prediction signal is activated, the target scene recognition algorithm parameter set is pre-written into the pipeline pre-register, but it does not immediately replace the currently valid parameters. This separate storage method allows the pre-fetching process of the recognition algorithm parameter set to be executed in parallel with the recognition processing of the current frame. When a new frame enters the processing pipeline, the recognition algorithm parameter set in the pipeline pre-register can be directly promoted to valid parameters, achieving synchronization between parameter switching and frame data arrival.

[0032] In this embodiment, to reduce the computational overhead of scene feature extraction during the non-mutation steady-state phase, when the scene mutation prediction signal is in an inactive state, only simplified scene descriptors of downsampled detection data are calculated for the current frame. Specifically, the multi-channel detection data of the current frame is spatially downsampled according to a preset downsampling ratio, simplified scene descriptors are extracted from the downsampled multi-channel detection data, and the simplified scene descriptors are appended to the scene descriptor temporal cache queue. The simplified scene descriptors retain the change trend information of descriptors in each dimension, which can support the temporal difference analysis of subsequent frames, while reserving the computational resources released in the scene evaluation stage for the recognition and inference stage.

[0033] Step 4: Promote the set of recognition algorithm parameters in the pipeline preparation register to valid parameters, and perform feature separation operation on the detection data.

[0034] When a new frame of detection data enters the recognition processing pipeline, the recognition algorithm parameter set in the pipeline's pre-register is simultaneously promoted to the currently valid parameters. If there is no recognition algorithm parameter set to be promoted in the pipeline's pre-register, the existing currently valid parameters are used. Based on the feature extraction range configuration in the valid parameters, the corresponding feature separation operation is performed on the current frame of detection data to generate feature response separation maps for different detection dimensions.

[0035] It should be noted that feature separation operation refers to performing independent feature extraction processing on the detection data in each dimension according to the detection dimension division method defined in the feature extraction range configuration, generating feature response separation maps for each dimension. Each feature response separation map corresponds to a detection dimension and represents the target response intensity distribution of the detection data in that detection dimension.

[0036] Step 5: Perform multi-dimensional target recognition processing on each feature response separation map to generate a set of candidate target regions for each dimension and calculate the weighted fusion confidence score.

[0037] Each feature response separation map is processed using a target recognition algorithm matched to its detection dimension to obtain a set of candidate target regions for each dimension. The target recognition algorithm processing includes morphological connected component analysis, phase consistency edge detection, and local contrast anomaly detection. Morphological connected component analysis takes a single feature response separation map as input and outputs the regions in the feature response separation map that satisfy the connectivity condition and their morphological attributes. Phase consistency edge detection takes the feature response separation map as input and outputs candidate regions with significant edge responses and their edge strength. Local contrast anomaly detection takes the feature response separation map as input and outputs anomalous regions where the local contrast deviates from the statistical background and the degree of deviation.

[0038] Based on the multi-algorithm weight configuration in the effective parameters, the confidence scores of each candidate target region obtained under different recognition algorithms are assigned corresponding weights, and the weighted fusion confidence score of each candidate target region is calculated. Let a certain candidate target region be in the... The confidence level under the recognition algorithm is The corresponding fusion weight is The weighted fusion confidence score of the candidate target region is then... for: in The total number of recognition algorithms participating in the fusion. For the index of the identification algorithm, For the candidate target region in the th Confidence level under various recognition algorithms For the first The fusion weights corresponding to the recognition algorithms satisfy and Confidence levels output by each recognition algorithm Before participating in weighted fusion, mean normalization based on range is used to uniformly map to... The interval eliminates the impact of inconsistent output dimensions of three types of algorithms—morphological connected component analysis, phase consistency edge detection, and local contrast anomaly detection—on the fusion operation.

[0039] Step 6: Filter candidate target regions and output the final detection and recognition results.

[0040] Candidate target regions whose weighted fusion confidence scores exceed the scene adaptive judgment threshold in the sensitivity threshold configuration are selected. For the selected candidate target regions, their scene category and detection dimension information are labeled, and the final detection and recognition results are output. At the same time, the multi-dimensional scene feature descriptor of the current frame is retained in the scene descriptor temporal cache queue for use in scene change prediction of subsequent frames.

[0041] In this embodiment of the application, in order to improve the robustness of the screening during the scene change transition phase, the following processing is also included in step 6: when the current frame is in the period after the scene change prediction signal is activated. During the frame transition period, a relaxation factor is applied to the scene adaptive judgment threshold in the sensitivity threshold configuration, where The transition period frame count is defined by a relaxation factor that is greater than zero and less than one, set as a preset value. The standard scene adaptive judgment threshold is set to... The relaxation factor is The threshold for determining actual use during the transition period is: ,in This is the threshold for determining actual use during the transition period. As a relaxation factor, For standard scenarios, adaptive threshold determination is performed. , making This means that during the transition period, the scene adaptive judgment threshold is appropriately reduced to avoid incorrect filtering of real targets due to the boundary effects of switching between old and new scene parameters. After the transition period, the scene adaptive judgment threshold is restored to the standard configuration value of the target scene. .

[0042] This implementation identifies the mutation trend before the actual occurrence of a scene mutation by performing first-order difference analysis on the historical frame descriptor sequence in the scene descriptor temporal cache queue, generating a scene mutation prediction signal. Based on the scene mutation prediction signal, the target scene's recognition algorithm parameter set is preloaded into the pipeline pre-register, so that when a mutated frame arrives at the recognition processing pipeline, the recognition algorithm parameter set in the pipeline pre-register can be synchronously promoted to valid parameters. Since the pre-fetching process of the recognition algorithm parameter set is executed in parallel with the recognition processing of the current frame, the matching algorithm configuration can be obtained when a new scene frame arrives without waiting for the scene classification calculation to be completed, thereby eliminating parameter mismatch in transition frames caused by scene classification calculation delay. The activation condition that the first-order difference of two consecutive frames exceeds the scene mutation threshold distinguishes instantaneous noise disturbances from persistent scene offsets, avoiding misresponding to random fluctuations. The mutation direction identifier determined by the second-order difference sequence provides directional constraints for parameter retrieval, narrowing the range of candidate scenes. Meanwhile, in the non-mutation steady-state stage, a downsampling simplified scene descriptor calculation method is adopted to replace the complete multi-dimensional scene feature descriptor extraction, which reduces the computational load of the scene evaluation stage and allocates the freed computational resources to the recognition and reasoning stage, thus alleviating the problem of scene evaluation and recognition and reasoning competing for computational time window under high frame rate detection conditions.

[0043] The following is an example of an application of the present invention, such as... Figure 2-9 As shown, the implementation process is as follows: An online inspection system on an industrial production line performs real-time foreign object detection on products on a conveyor belt. The system acquires multi-channel detection data at a frame rate of 60 frames per second, with a buffer window length of [missing information]. The number of frames during the transition period is 5. The sequence consists of 3 frames. In a batch inspection task in 20XX, the conveyor belt switched from transporting light-colored ceramic parts to transporting dark-colored metal parts around frame 41. The texture, strength, and structural characteristics of the two types of products differed significantly, triggering a scene change prediction signal. The data flow process is explained below step by step.

[0044] Step 1: Obtain the detection data of the current frame and extract the multi-dimensional scene feature descriptor.

[0045] The system extracts texture complexity, intensity distribution statistics, and structural pattern consistency index from the multi-channel detection data of frames 39 to 43, and performs Z-score normalization on each dimension. The normalized descriptors are appended to the scene descriptor temporal cache queue, which always maintains the data of the most recent 5 frames. Descriptors of frames 38 and earlier are removed sequentially when the queue is updated.

[0046] Table 1. Scene Descriptor Time-Sequence Cache Queue (after Z-score Normalization) Frames 39 and 40 correspond to the light-colored ceramic part scene, with low texture complexity, concentrated intensity distribution, and high structural consistency; starting from frame 41, the scene transition area begins, and the descriptors of each dimension show directional shifts.

[0047] Step 2: Perform temporal differential analysis on the scene descriptor temporal cache queue to generate scene mutation prediction signals.

[0048] Calculate the first difference for each dimension descriptor sequence in the cache queue. ,in For the first Dimension in the th The frame's normalized descriptor value; and the second-order difference is calculated. Scene mutation thresholds in various dimensions In order: texture complexity Intensity distribution Structural consistency .

[0049] Table 2 First-order difference and activation judgment for each dimension Taking texture complexity as an example, the absolute values ​​of the first-order difference between frames 41 and 42 are 1.22 and 0.72 respectively, both exceeding [the required values]. If the activation condition of exceeding the threshold is met for two consecutive frames, the scene change prediction signal is set to the active state in the 42nd frame.

[0050] Calculate the second-order difference for frame 42: texture complexity dimension The values ​​remain negative, indicating that the rate of change is converging; the second-order differences of the intensity distribution and structural consistency dimensions are also negative. The direction of abrupt change is identified as follows: texture complexity and intensity distribution increase, while structural consistency decreases, corresponding to a shift from a low-texture, high-consistency ceramic scene to a high-texture, low-consistency metal scene.

[0051] Step 3: Based on the state of the scene change prediction signal, perform prefetching or maintenance processing of the recognition algorithm parameters.

[0052] After the scene change prediction signal is activated in frame 42, the system retrieves the recognition algorithm parameter set corresponding to the metal part detection scene from the recognition algorithm configuration library based on the change direction identifier and writes it into the pipeline preparation register. The recognition processing of the current frame (frame 42) continues to use the valid parameters of the ceramic part scene, and the prefetching process is executed in parallel with the recognition processing of the current frame.

[0053] Table 3 Comparison of recognition algorithm parameter sets (ceramic parts scenario and metal parts scenario) During the non-mutation steady-state phase from frame 39 to 42, the scene mutation prediction signal is in an inactive state. The system only calculates a simplified scene descriptor with a downsampling ratio of 0.25 for each frame, and allocates the saved computing resources to the recognition and reasoning stage.

[0054] Step 4: Promote the set of recognition algorithm parameters in the pipeline preparation register to valid parameters, and perform feature separation operation on the detection data.

[0055] When the 43rd frame of detection data enters the recognition processing pipeline, the system simultaneously upgrades the parameter set of the metal scene recognition algorithm in the pipeline's pre-register to the currently valid parameters. Based on the feature extraction range configuration (edge ​​contour dimension and reflection intensity dimension), independent feature extraction processing is performed on the 43rd frame of detection data, generating one edge contour feature response separation map and one reflection intensity feature response separation map.

[0056] Step 5: Perform multi-dimensional target recognition processing on each feature response separation map to generate a set of candidate target regions for each dimension and calculate the weighted fusion confidence score.

[0057] Morphological connected component analysis, phase consistency edge detection, and local contrast anomaly detection are performed on the two feature response separation maps respectively. After normalization, the confidence scores of each candidate target region under the three algorithms are obtained. The weighted fusion confidence scores are calculated according to the multi-algorithm weight configuration of the metal part scene. ,in For the first The weights of the algorithm, This represents the normalized confidence level of the corresponding algorithm.

[0058] Table 4. Calculation of Weighted Fusion Confidence of Candidate Target Regions Weight configuration as , , The sum of the three weights is 1.00. Taking region A as an example, the weighted fusion confidence score is... .

[0059] Step 6: Filter candidate target regions and output the final detection and recognition results.

[0060] Frame 43 is the first frame after the scene change prediction signal is activated, which is in the transition period. Within a frame. Relaxation factor. Standard scenario adaptive judgment threshold The actual threshold for judgment during the transition period is .

[0061] Table 5. Candidate Region Screening Results The weighted fusion confidence scores of regions A and C both exceeded the transition period decision threshold of 0.61. The output was the final detection and recognition result, labeled with the scene category as metal part detection scene and the corresponding detection dimension information. After the transition period ended (starting from frame 46), the decision threshold returned to [value missing]. .

[0062] The entire data flow process exhibits a clear forward dependency: the Z-score normalized descriptors extracted in step 1 fill the scene descriptor temporal cache queue, and the queue data is directly input into the first-order and second-order difference calculations in step 2; the scene mutation prediction signal state and mutation direction identifier generated in step 2 control the parameter retrieval behavior in step 3, determining the parameter set content written to the pipeline pre-register; the recognition algorithm parameter set prefetched in step 3 is synchronously promoted to effective parameters in step 4, where the feature extraction range configuration determines which feature response separation maps are generated in step 4, and the multi-algorithm weight configuration and standard scene adaptive judgment threshold play roles in steps 5 and 6, respectively; the weighted fusion confidence score calculated in step 5 and the transition period relaxation threshold in step 6 jointly determine the final output result. The input and output data of each step remain consistent in the frame sequence of the same detection task, reflecting the collaborative operation logic of scene mutation prediction and parameter prefetching.

[0063] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for adjusting the recognition algorithm after intelligently detecting and recognizing a scene, characterized in that, include: Extract the multi-dimensional scene feature descriptor from the current frame detection data and append it to the scene descriptor temporal cache queue; Calculate the first-order difference sequence for each dimension descriptor sequence in the time-series cache queue, compare the absolute value of the first-order difference of each dimension with the scene change threshold of the corresponding dimension, and generate a scene change prediction signal. When the scene mutation prediction signal is in an active state, the target scene's recognition algorithm parameter set is retrieved from the recognition algorithm configuration library and written into the pipeline preparation register; when it is in an inactive state, the current recognition algorithm parameters are maintained. When a new frame of detection data enters the recognition processing pipeline, the set of recognition algorithm parameters in the pipeline preparation register is promoted to the current valid parameters. Based on the valid parameters, feature separation operation is performed on the detection data to generate a feature response separation map. Target recognition processing is performed on each feature response separation map, and the weighted fusion confidence score of each candidate target region is calculated. Candidate target regions whose weighted fusion confidence scores exceed the scene adaptive judgment threshold are selected, and the detection and recognition results are output.

2. The method for adjusting the recognition algorithm after intelligent scene detection according to claim 1, characterized in that, The multidimensional scene feature descriptor includes a texture complexity index, an intensity distribution statistic, and a structural pattern consistency index. The texture complexity index represents the degree of intensity of texture changes within the detection area of ​​the current frame. The intensity distribution statistic represents the concentrated or discrete distribution of signal intensity in the detection data. The structural pattern consistency index represents the regularity of repetitive structural patterns in the current frame. After extracting the multidimensional scene feature descriptors, Z-score normalization is applied to each dimension descriptor to eliminate the dimensional differences between the dimensions.

3. The method for adjusting the recognition algorithm after intelligently detecting and recognizing a scene according to claim 1, characterized in that, The generation of the scene mutation prediction signal includes: when the absolute value of the first-order difference of two consecutive frames in any dimension exceeds the scene mutation threshold of the corresponding dimension, the scene mutation prediction signal is set to the active state. When the absolute value of the first difference in each dimension does not meet the above conditions, the scene mutation prediction signal is set to an inactive state.

4. The method for adjusting the recognition algorithm after intelligent scene detection according to claim 3, characterized in that, Also includes: Calculate the second-order difference sequence for each dimension descriptor sequence, where the second-order difference sequence is the difference between the first-order differences of two adjacent frames; When the second difference of a certain dimension is continuously positive, the mutation direction identifier of that dimension is set to the upward direction; when the second difference of a certain dimension is continuously negative, the mutation direction identifier of that dimension is set to the downward direction; the mutation direction identifier is used to determine the parameter retrieval direction of the target scene when retrieving the recognition algorithm parameter set from the recognition algorithm configuration library.

5. The method for adjusting the recognition algorithm after intelligently detecting and recognizing a scene according to claim 1, characterized in that, The recognition algorithm parameter set includes feature extraction range configuration, multi-algorithm weight configuration, and sensitivity threshold configuration. The feature extraction range configuration defines the detection dimensions and spatial range covered when performing feature separation operations on the target scene. The multi-algorithm weight configuration defines the fusion weight allocation of the target recognition algorithm output results on each detection dimension. The sensitivity threshold configuration defines the scene adaptive judgment threshold under the target scene conditions.

6. The method for adjusting the recognition algorithm after intelligent scene detection according to claim 1, characterized in that, The pipeline preparation register is a buffer area independent of the current valid parameter register. The prefetching process of the recognition algorithm parameter set is executed in parallel with the recognition processing process of the current frame. When a new frame enters the recognition processing pipeline, the recognition algorithm parameter set in the pipeline preparation register is directly promoted to valid parameters, thereby realizing the synchronization of parameter switching and frame data arrival.

7. The method for adjusting the recognition algorithm after intelligent scene detection according to claim 1, characterized in that, When the scene mutation prediction signal is in an inactive state, the multi-channel detection data of the current frame is spatially downsampled according to a preset downsampling ratio. A simplified scene descriptor is extracted from the downsampled multi-channel detection data, and the simplified scene descriptor is added to the scene descriptor temporal cache queue to replace the multi-dimensional scene feature descriptor.

8. The method for adjusting the recognition algorithm after intelligently detecting and recognizing a scene according to claim 5, characterized in that, The target recognition process includes morphological connected component analysis, phase-consistent edge detection, and local contrast anomaly detection. The calculation of the weighted fusion confidence score of each candidate target region includes: normalizing the confidence scores of each candidate target region under each recognition algorithm using mean normalization based on the range to uniformly map them to the interval between zero and one; and weighting and summing the normalized confidence scores according to the fusion weights corresponding to each recognition algorithm in the multi-algorithm weight configuration to obtain the weighted fusion confidence score of each candidate target region. The sum of the fusion weights of each recognition algorithm is one, and each fusion weight is greater than or equal to zero.

9. The method for adjusting the recognition algorithm after intelligently detecting and recognizing a scene according to claim 5, characterized in that, When the current frame is within the preset transition period frame number after the scene mutation prediction signal is activated, the scene adaptive judgment threshold in the sensitivity threshold configuration is multiplied by a relaxation coefficient. The relaxation coefficient is a preset value that is greater than zero and less than one, so that the judgment threshold actually used during the transition period is lower than the standard scene adaptive judgment threshold. After the transition period ends, the scene adaptive judgment threshold will be restored to the standard configuration value of the target scene.

10. A system for adjusting the recognition algorithm after intelligently detecting and identifying a scene, used to execute the method for adjusting the recognition algorithm after intelligently detecting and identifying a scene as described in any one of claims 1 to 9, characterized in that, include: The scene feature extraction module is used to extract multi-dimensional scene feature descriptors from the detection data of the current frame and append the multi-dimensional scene feature descriptors to the scene descriptor temporal cache queue. The temporal difference analysis module is used to calculate the first-order difference sequence of each dimension descriptor sequence in the temporal cache queue, compare the absolute value of the first-order difference of each dimension with the scene change threshold of the corresponding dimension, and generate a scene change prediction signal. The parameter prefetching module is used to retrieve the target scene recognition algorithm parameter set from the recognition algorithm configuration library and write it into the pipeline pre-register when the scene mutation prediction signal is in an active state; when it is in an inactive state, it maintains the current recognition algorithm parameters. The feature separation module is used to upgrade the set of recognition algorithm parameters in the pipeline preparation register to the current valid parameters when a new frame of detection data enters the recognition processing pipeline, and to perform feature separation operation on the detection data according to the valid parameters to generate a feature response separation map. The target recognition module is used to perform target recognition processing on each feature response separation map and calculate the weighted fusion confidence score of each candidate target region; The result filtering module is used to filter candidate target regions whose weighted fusion confidence scores exceed the scene adaptive judgment threshold and output the detection and recognition results.

Images