A blackboard conductive film touch calibration method and system based on deep learning

Abstract

The present application relates to touch calibration technical field, specifically be a kind of blackboard conductive film touch calibration method and system based on deep learning.The present application is based on deep learning to handle blackboard conductive film touch process, by extracting touch area and touch feature and dividing touch event fragment, using pre-trained long short-term memory network to identify candidate writing fragment and non-writing fragment, then combining touch calibration point coordinates and display calibration point coordinates to filter actual writing fragment, and using actual writing fragment to form incremental learning sample, while establishing non-writing event decision boundary according to non-writing fragment and isolation fragment, then combining convolutional neural network to learn the offset relationship of touch trajectory coordinates, so as to reduce palm false touch and wipe touch interference on writing trajectory calibration data, make the data involved in calibration more consistent with actual writing touch process.

Description

Technical Field

[0001] This invention relates to the field of touch calibration technology, and in particular to a blackboard conductive film touch calibration method and system based on deep learning. Background Technology

[0002] The field of touch calibration technology mainly involves detecting and adjusting the correspondence between the actual touch position in a touch device and the position identified by the system. This field of technology is usually based on the sensing information generated by the touch device. According to the touch area and position reference determined in advance by the device, the position data generated when the user touches is identified, and the data content required for calibration is determined by combining the spatial correspondence of the touch area.

[0003] Traditional blackboard conductive film touch calibration typically involves first collecting touch coordinates at several calibrated locations, then establishing a mapping relationship between touch coordinates and display coordinates, and correcting the coordinates of subsequent touch positions based on preset calibration parameters. While this can complete the basic calibration between the touch position and the display position of the blackboard conductive film, in actual writing processes, when writing touch, accidental palm touch, and wiping touch occur alternately, non-writing touch data can be mixed into the calibration data, interfering with the touch trajectory offset information on which the calibration is based. This leads to a deviation between the calibration relationship and the actual writing touch trajectory. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a blackboard conductive film touch calibration method and system based on deep learning.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a blackboard conductive film touch calibration method based on deep learning, comprising the following steps:

[0006] S1: Determine the touch area of ​​the blackboard conductive film within the scanning cycle and obtain the touch feature sequence; wherein, the touch features include the area of ​​the touch area and the peak value of capacitance change, the movement speed of the touch trajectory, and the duration of the touch event;

[0007] S2: Divide the touch event segments according to the touch area and touch feature sequence, and determine the event time range and event coordinate range of the touch event segments; then select candidate writing segments and non-writing segments from the touch event segments according to the pre-trained long short-term memory network;

[0008] S3: Obtain the coordinates of the touch calibration point in the blackboard conductive film touch coordinate system, which represents the touch position of the touch calibration point, and the coordinates of the display calibration point in the blackboard display area display coordinate system, which represents the display position of the display calibration point. Refer to the touch calibration point coordinates and the display calibration point coordinates, select the actual writing segment from the candidate writing segments, and use the remaining candidate writing segments as isolated segments.

[0009] S4: Construct incremental learning samples based on actual written fragments; calculate the decision boundary for non-written events based on non-written fragments and isolated fragments;

[0010] S5: Build a convolutional neural network, then update the convolutional neural network using incremental learning samples, and combine the non-writing event decision boundary and the updated convolutional neural network to calibrate the touch trajectory and obtain the touch trajectory calibration result.

[0011] A deep learning-based blackboard conductive thin film touch calibration system, the system comprising:

[0012] The touch feature extraction module determines the touch area of ​​the blackboard conductive film within the scanning cycle and acquires the touch feature sequence; wherein, the touch features include the area of ​​the touch area and the peak value of capacitance change, the movement speed of the touch trajectory, and the duration of the touch event;

[0013] The event segmentation module divides touch event segments according to the touch area and touch feature sequence, and determines the event time range and event coordinate range of the touch event segments; then, it selects candidate writing segments and non-writing segments from the touch event segments based on the pre-trained long short-term memory network.

[0014] The writing segment confirmation module obtains the coordinates of the touch calibration point in the blackboard conductive film touch coordinate system, which represents the touch position of the touch calibration point, and the coordinates of the display calibration point in the blackboard display area display coordinate system, which represents the display position of the display calibration point. Referring to the touch calibration point coordinates and the display calibration point coordinates, the actual writing segment is selected from the candidate writing segments, and the remaining candidate writing segments are used as isolated segments.

[0015] The boundary construction module builds incremental learning samples based on actual written fragments; it calculates the decision boundary for non-written events based on non-written fragments and isolated fragments.

[0016] The trajectory calibration module establishes a convolutional neural network, then updates the convolutional neural network using incremental learning samples, and calibrates the touch trajectory by combining the non-writing event decision boundary and the updated convolutional neural network to obtain the touch trajectory calibration result.

[0017] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0018] This invention processes the touch control process of a conductive film on a blackboard using deep learning. It extracts the touch area and touch features, segments touch events, and uses a pre-trained long short-term memory network to identify candidate writing and non-writing segments. Then, it combines the coordinates of touch calibration points with those of display calibration points to filter actual writing segments and uses these segments to form incremental learning samples. Simultaneously, it establishes a non-writing event decision boundary based on non-writing and isolated segments, and combines this with a convolutional neural network to learn the offset relationship of touch trajectory coordinates. This reduces the interference of accidental hand touches and wiping touches on the writing trajectory calibration data, making the calibration data more closely match the actual writing touch process. Furthermore, the touch event recognition and trajectory offset learning process based on deep learning can update the calibration relationship using actual writing data, making the positional correspondence between the touch trajectory and the displayed handwriting more accurate. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the steps of the method of the present invention;

[0020] Figure 2 This is a flowchart illustrating the steps and principles of the method of the present invention.

[0021] Figure 3 This is a block diagram of the system of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0023] Please see Figure 1-2 This invention provides a technical solution: a deep learning-based method for calibrating touch controls on a blackboard conductive film, comprising the following steps:

[0024] S1: Determine the touch area of ​​the blackboard conductive film within the scanning cycle and acquire the touch feature sequence; wherein, the touch features include the area of ​​the touch area and the peak value of capacitance change, the movement speed of the touch trajectory, and the duration of the touch event, including:

[0025] First, a capacitance acquisition circuit electrically connected to each capacitance sensing node in the conductive film of the blackboard sequentially selects the capacitance sensing nodes according to the scan cycle. Within one scan cycle, capacitance readings of all capacitance sensing nodes are acquired. Simultaneously with reading the capacitance change, the sampling timestamp generated by the scan clock is read. Based on the row and column arrangement of the capacitance sensing nodes, the node coordinates of the same capacitance sensing node in the touch coordinate system are read from the node position configuration data, ensuring that each capacitance change is associated with the sampling timestamp of the same scan cycle and the node coordinates of the same capacitance sensing node. The touch detection threshold is set using the capacitance changes acquired from calibration sample touch events. The capacitance changes generated by each calibration sample touch event in each scan cycle are summarized, and the different capacitance changes are arranged in ascending order. The difference between two adjacent capacitance changes is calculated sequentially, and the largest difference is selected from all adjacent differences. Then, the average of the two capacitance changes that form the largest difference is taken, and this average value is used as the median value between them. This median value is saved as the touch detection threshold. After setting the touch detection threshold, the capacitance change of each capacitive sensing node is compared with the touch detection threshold. If the capacitance change is greater than or equal to the touch detection threshold, the capacitive sensing node is registered as a touch node; otherwise, it is not registered. Then, the spatial adjacency relationship between touch nodes is checked according to their coordinates. Two touch nodes that are horizontally adjacent and vertically identical are established as horizontally adjacent; two touch nodes that are vertically adjacent and horizontally identical are established as vertically adjacent. Starting from any ungrouped touch node, the search continues to find adjacent touch nodes, and continues to search for ungrouped neighboring touch nodes around the newly added touch node, until no more touch nodes can be added to the current touch node set. Then, the same process is repeated from the next ungrouped touch node, ultimately obtaining the touch area defined by consecutive adjacent touch nodes in each scanning cycle.

[0026] Next, the capacitance changes and node coordinates of all touch nodes within a single touch area are read, and the capacitance changes of all touch nodes are accumulated. Then, the ratio of the capacitance change of each touch node to the sum of the capacitance changes of all touch nodes is calculated, and this ratio is used as the normalized weight of the corresponding touch node. After obtaining all normalized weights, a weighted average is performed on the horizontal and vertical coordinates of the node coordinates. , ;in, Indicates the x-coordinate of the center of the touch area. Represents the center vertical coordinate of the touch area. Indicates the first Normalized weights of each touch node, Indicates the first The x-coordinate of each touch node, Indicates the first The ordinate of each touch node, Indicates the index of the touch node participating in the weighted calculation. This indicates the total number of touch nodes within the current touch area. The area of ​​a single node is read from the node area configuration data of the blackboard conductive film. The number of touch nodes in the current touch area is counted, and then the product of the number of touch nodes and the area of ​​a single node is calculated. For the peak capacitance change, the capacitance change of the first touch node is written to the current maximum value record. Subsequently, the capacitance changes of the remaining touch nodes are read sequentially. If the subsequent capacitance change is greater than the current maximum value record, the subsequent capacitance change record is used to replace the current maximum value record. If the subsequent capacitance change is less than or equal to the current maximum value record, the current maximum value record remains unchanged. This process continues until all touch nodes are compared, ultimately obtaining the center coordinates of the touch area, the area of ​​the touch area, and the peak capacitance change of the touch area.

[0027] Then, the center coordinates of the touch areas formed by the same touch event within consecutive scan cycles are arranged from earliest to latest according to the sampling timestamp, and two scan cycles with adjacent time positions are read sequentially. The distance between the center coordinates of the touch areas in adjacent scan cycles is calculated according to the following formula: ;in, Indicates the first The scan cycle and the first The distance between the center coordinates of the touch area between scan cycles Indicates the first The horizontal coordinate of the center of the touch area in each scan cycle. Indicates the first The horizontal coordinate of the center of the touch area in each scan cycle. Indicates the first The vertical coordinate of the center of the touch area per scan cycle. Indicates the first The vertical coordinate of the center of the touch area per scan cycle. This represents the scan cycle index arranged according to the sampling timestamps. Simultaneously, the sampling timestamps of two adjacent scan cycles are read, the time difference between the later and earlier sampling timestamps is calculated, and then the ratio of the distance between the center coordinates of the touch area in adjacent scan cycles to this time difference is calculated as the touch trajectory movement speed between adjacent scan cycles. Subsequently, the first and last sampling timestamps of the same touch event are read, and the time difference between the last and first sampling timestamps is calculated as the touch event duration. After completing the data processing for each scan cycle, using the scan cycle number as the time index, the touch area, peak capacitance change, touch trajectory movement speed, and touch event duration of the same scan cycle are written to the same time sequence location, and then arranged continuously according to the scan cycle order to finally obtain the touch feature sequence.

[0028] S2: Based on the touch area and touch feature sequence, segment the touch event into segments, and determine the event time range and event coordinate range of each segment; then, based on a pre-trained Long Short-Term Memory network, select candidate written segments and non-written segments from the touch event segments, including:

[0029] The touch area, peak capacitance change, touch trajectory movement speed, and touch event duration are extracted from the touch feature sequence. For each touch feature, the average and standard deviation are calculated. Then, the standard score of the current touch feature value relative to the average and standard deviation of the same touch feature is calculated. If the standard deviation of a touch feature is greater than zero, the corresponding standard score is calculated. If the standard deviation of a touch feature is equal to zero, the current touch feature values ​​are the same, and the standard score of that touch feature is uniformly recorded as zero. The time threshold is set using the time difference formed by adjacent sampling timestamps in the touch feature sequence. First, the time difference between all adjacent sampling timestamps is calculated, and the different time differences are arranged in ascending order of value. Then, the interval between two adjacent time differences is calculated, and the two time differences on either side of the largest time interval are selected. The midpoint between the two time differences is taken as the time threshold. The distance threshold is calculated by treating the center coordinate distance between the center coordinates of the touch area in adjacent scan cycles in the same way. First, the center coordinate distances of all adjacent scan cycles are calculated, and the different center coordinate distances are arranged in ascending order of value. Then, the interval between two adjacent center coordinate distances is calculated, and the center coordinate distances on both sides of the largest distance interval are selected, with the median value of the two center coordinate distances being taken as the distance threshold. During the formal segmentation, the time difference between every two adjacent sampling timestamps and the center coordinate distance between the center coordinates of the touch area in every two adjacent scan cycles are calculated. If the time difference is greater than the time threshold, the scan cycle number of the next sampling timestamp is determined as the segmentation position. If the center coordinate distance is greater than the distance threshold, the scan cycle number of the next touch area center coordinate is determined as the segmentation position. When both conditions are met, only one segmentation position formed by the same scan cycle number is retained. If neither condition is met, no segmentation position is set between two scan cycles. Subsequently, the touch feature sequence is segmented according to the continuous segmentation positions, and the standard score of the same scan cycle is extracted based on the scan cycle number retained after segmentation. Finally, the touch event segments and the standard scores corresponding to each touch event segment are obtained.

[0030] All scan records in a single touch event segment are read from earliest to latest according to the sampling timestamp. The earliest sampling timestamp is recorded as the starting boundary of the event time range, and the last sampling timestamp is recorded as the ending boundary of the event time range. For the event coordinate range, the horizontal coordinate of the center of the first touch area is read first, and the current minimum horizontal coordinate and the current maximum horizontal coordinate are written at the same time. Then, the horizontal coordinates of the centers of the remaining touch areas are read in sequence. If the current horizontal coordinate of the center of the touch area is less than the current minimum horizontal coordinate, the current minimum horizontal coordinate is replaced. If the current horizontal coordinate of the center of the touch area is greater than the current maximum horizontal coordinate, the current maximum horizontal coordinate is replaced. When the current horizontal coordinate of the center of the touch area is between the current minimum horizontal coordinate and the current maximum horizontal coordinate, the two horizontal boundaries remain unchanged. The vertical coordinate of the center of the touch area is obtained by the same value-by-value comparison process. The minimum vertical coordinate and the maximum vertical coordinate are obtained by the minimum horizontal coordinate, the maximum horizontal coordinate, the minimum vertical coordinate, and the maximum vertical coordinate together define the event coordinate range. Then, according to the scan cycle number within the touch event segment, the standard scores of the touch area, peak capacitance change, touch trajectory movement speed, and touch event duration for the same scan cycle are read sequentially. This ensures that the four standard scores for each scan cycle are arranged in a fixed order and continuously according to the sampling timestamp, ultimately yielding the event time range, event coordinate range, and standard score sequence for the touch event segment.

[0031] Before pre-training the Long Short-Term Memory (LSTM) network, continuous touch data generated by writing, palm touch, and wiping operations are collected by a blackboard conductive film using the same scanning method as the calibration sampling. Each set of continuous touch data is processed by touch area extraction and touch feature calculation, and then tagged with either a writing category, palm touch category, or wiping category according to the actual touch operation performed. The resulting standard score sequence is then formed using the aforementioned standard score processing method. The LSM network receives four input data items in each scan cycle: standard score for touch area, standard score for peak capacitance change, standard score for touch trajectory movement speed, and standard score for touch event duration. The LSM network has 16 memory nodes, which continuously receive the four standard scores of the current scan cycle and the hidden state retained from the previous scan cycle in each scan cycle according to the same connection relationship. The output has three category positions representing the writing category, palm touch category, and wiping category. When processing a single scan cycle, the Long Short-Term Memory (LSTM) network first performs a weighted combination of the current four standard scores and the hidden state from the previous scan cycle, and uses a logistic function to form a forgetting ratio between zero and one. This forgetting ratio controls the proportion of the memory state from the previous scan cycle that is retained in the current scan cycle. Next, another set of weighted processing is performed on the current four standard scores and the hidden state from the previous scan cycle, and a write ratio between zero and one is formed using a logistic function. Simultaneously, a hyperbolic tangent function is used to form candidate memory contents with values ​​between -1 and 1, and these candidate contents are then written into the already retained memory state according to the write ratio. Then, the current four standard scores and the hidden state from the previous scan cycle are weighted again, and an output ratio is formed using a logistic function. The current memory state is then processed by the hyperbolic tangent function and used to form the hidden state for the current scan cycle according to the output ratio. The current memory state and the current hidden state are then passed to the next scan cycle. Before the initial training, the connection weights are initialized with a symmetrical range based on the number of inputs and outputs at the current connection position. The larger the number of inputs and outputs, the narrower the initial range, thus constraining the absolute value of the initial weight of a single connection to the connection size. Each connection weight acquires an initial value within this symmetrical range. The larger the absolute value of the connection weight, the greater the contribution of the corresponding input value to the weighted processing. The initial bias is recorded as zero. After completing the state transfer for all scanning cycles of a touch sample, the 16 hidden states formed in the last scanning cycle are input to the three category positions. First, weighted output values ​​for the writing category, the palm touch category, and the wiping category are generated respectively. Then, the three weighted output values ​​are normalized so that the sum of the probabilities of the three categories is 1.During training, the system reads the saved category labels from the touch samples, compares the classification differences between the three category probabilities and the category labels, and calculates the adjustment direction of the connection weights layer by layer from the category output position backward. After each round of weight adjustment, the training data is re-inputted to calculate the classification differences. If the classification differences decrease, the currently adjusted connection weights are retained; if the classification differences do not decrease, the adjustment range of the connection weights in this round is reduced and recalculated. The training samples are processed continuously, and the connection weights after training are saved. When formally processing the standard score sequence, if the writing category probability is the maximum of the three category probabilities, the current touch event segment is registered as a candidate writing segment; if the palm touch category probability or wiping category probability is the maximum of the three category probabilities, the current touch event segment is registered as a non-writing segment. Finally, candidate writing segments and non-writing segments are obtained.

[0032] S3: Obtain the coordinates of the touch calibration point representing the touch position in the blackboard conductive film touch coordinate system and the coordinates of the display calibration point representing the display position in the blackboard display area display coordinate system. Referring to the touch calibration point coordinates and the display calibration point coordinates, select the actual writing segment from the candidate writing segments. The remaining candidate writing segments are used as isolated segments, including:

[0033] Touch calibration positions are displayed sequentially on the blackboard display area. When an operator touches a touch calibration position, the horizontal and vertical coordinates of the corresponding touch calibration point are collected and recorded through the conductive film on the blackboard. Simultaneously, the horizontal and vertical coordinates of the display calibration point for the same touch calibration position in the blackboard display area's coordinate system are read, and the touch calibration point coordinates are paired with the display calibration point coordinates using the same calibration sequence number. Planar projection constraints are established using at least four sets of coordinate pairs where the touch calibration points are not located on the same straight line. The coordinate relationship of the homography matrix is ​​as follows: , ;in, Indicates the x-coordinate of the touch calibration point. Indicates the vertical coordinate of the touch calibration point. This represents the x-coordinate of the display calibration point that has the same calibration sequence number as the current touch calibration point. This represents the ordinate of the display calibration point that has the same calibration number as the current touch calibration point. , , , , , , , and These represent the nine projection parameters at the three rows and three columns of the homography matrix. Each set of coordinates is paired to form two projection constraints: horizontal and vertical. After substituting the touch calibration point coordinates and display calibration point coordinates into the projection relationship, the terms are rearranged to form a system of linear equations for each projection parameter. Then, non-zero principal components are selected column by column according to the coefficients of the linear equations. The current principal component is used to eliminate the coefficients of other equations at the same projection parameter position. This process is repeated for each projection parameter. Since changing all nine projection parameters simultaneously according to the same proportion will not change the projected coordinate positions, [the following steps are taken]. The coordinates are fixed at 1, and the remaining 8 projection parameters are obtained from the remaining coordinate constraints. Finally, the homography matrix is ​​obtained by arranging the coordinates in a three-row, three-column configuration.

[0034] Next, the x-coordinate and y-coordinate of the touch area center corresponding to each sampling timestamp within the candidate writing fragment are read, and the nine projection parameters in the homography matrix are read to perform coordinate transformation on the center coordinates of each touch area: , ;in, Indicates the first candidate written fragment The center horizontal coordinate of each touch area Indicates the first The center ordinate of each touch area This indicates the first coordinate system after switching to the blackboard display area. The horizontal coordinate of the touch trajectory Indicates the transformed first The vertical coordinate of the touch trajectory This indicates the position index within the candidate written fragment, arranged according to the sampling timestamp. to This represents the homography matrix projection parameters already obtained for segment 7. When displaying handwriting, the blackboard display area simultaneously saves the horizontal and vertical coordinates of the displayed handwriting and the sampling timestamp. It reads the coordinates of displayed handwriting within the same time range as the candidate writing segments, and then uses the same sampling timestamp as a matching condition to pair the touch trajectory coordinates with the displayed handwriting coordinates under the same sampling timestamp. The Euclidean distance of each coordinate pair is calculated using the following formula: ;in, Indicates the first The Euclidean distance between matching coordinate pairs Indicates the first The horizontal coordinate of the touch trajectory Indicates the first The vertical coordinate of the touch trajectory Indicates the relationship with the first Each touch trajectory coordinate has the same sampling timestamp on the horizontal axis of the displayed handwriting. Indicates the relationship with the first Each touch trajectory coordinate has the same sampling timestamp on the display handwriting vertical coordinate. This represents the index of the matching coordinate pair. After calculating all Euclidean distances for the current candidate written segment, the total Euclidean distances are accumulated and divided by the number of matching coordinate pairs to obtain the average trajectory distance of the current candidate written segment.

[0035] Subsequently, a trajectory distance threshold is set using the average trajectory distance calculated for all candidate writing fragments. Different average trajectory distances are arranged in ascending order of value, and the numerical interval between adjacent average trajectory distances is calculated sequentially. The largest interval is selected from all numerical intervals, and the median value between the two average trajectory distances forming the largest interval is obtained and recorded as the trajectory distance threshold. During formal screening, the average trajectory distance of each candidate writing fragment is compared with the trajectory distance threshold. If the average trajectory distance is less than or equal to the trajectory distance threshold, the current candidate writing fragment is registered as an actual writing fragment. If the average trajectory distance is greater than the trajectory distance threshold, the current candidate writing fragment is registered as an isolated fragment. The established event time range, event coordinate range, and sampling timestamp index of the current candidate writing fragment remain unchanged, ultimately resulting in actual writing fragments and isolated fragments.

[0036] S4: Construct incremental learning samples based on actual written fragments; calculate the decision boundary for non-written events based on non-written fragments and isolated fragments, including:

[0037] The sampling timestamps are filtered according to the event time range of the actual writing fragment, and the node coordinates are filtered according to the event coordinate range of the actual writing fragment. Capacitance changes that are simultaneously within the event time range and event coordinate range are read from the already saved capacitance sensing node records. The row position of the two-dimensional capacitance data is determined by the vertical position in the node coordinates, and the column position is determined by the horizontal position in the node coordinates. Each capacitance change within the same scan cycle is written to the corresponding row and column positions. Then, the capacitance changes in the two-dimensional capacitance data are compared position by position with the touch detection threshold. Positions with capacitance changes greater than or equal to the touch detection threshold participate in the connectivity search, while positions with capacitance changes less than the touch detection threshold do not participate. Starting from the center coordinate position of the touch area in the actual writing fragment, the horizontal and vertical adjacent positions are checked sequentially. As long as an adjacent position participates in the connectivity search, the search continues outward from the newly added position until there are no more unchecked consecutive adjacent positions. The node positions that can be continuously connected to the center coordinate position of the touch area form the handwriting connected region. For capacitance sensing nodes outside the handwriting connected region, those with capacitance changes greater than or equal to the touch detection threshold are registered as shielded nodes, while those with capacitance changes less than the touch detection threshold are not registered as shielded nodes. For the The shielded node, the first in the handwriting connected region The reference candidate node and the first The distance between shielded nodes is calculated using the following formula: ;in, Indicates the first The reference candidate node and the first The distance between shielded nodes Indicates the first The x-coordinates of the reference candidate nodes, Indicates the first The y-coordinates of the reference candidate nodes, Indicates the first The x-coordinate of each shielded node, Indicates the first The ordinate of each shielded node, This represents the index of a reference candidate node within the connected region of the handwriting. This represents the index of the masked node. All node distances are sorted in ascending order, and the reference node with the minimum distance is selected. When multiple reference candidate nodes have the same minimum distance, they are all included in the inverse distance weighting. The inverse distance weight and the change in capacitance are calculated using the following formula: , ;in, Indicates the first The inverse distance weight of each reference node, Indicates the first The node distance between each reference node and the currently masked node Indicates the first The node distance between each reference node and the currently masked node This represents the index of the reference node that participates in the reciprocal distance accumulation. This indicates the number of reference nodes currently participating in the inverse distance weighted average. Indicates the first Change in the replacement capacitance of each shielded node. Indicates the first The change in capacitance at each reference node Indicates the index of the current reference node. This represents the index of the currently masked node. After replacing the capacitance change with the original capacitance change at the location of the masked node in the two-dimensional capacitance data, normalization is performed on the replaced two-dimensional capacitance data to convert the capacitance change at each node location to a uniform numerical scale, ultimately obtaining the normalized capacitance data of the actual written fragment.

[0038] Then, the obtained display handwriting coordinates and touch trajectory coordinates for each actual writing segment are arranged according to the sampling timestamp. Using identical sampling timestamp values ​​as a matching condition, the display handwriting coordinates and touch trajectory coordinates under the same sampling timestamp are written into the same matching record. For each matching coordinate pair, the horizontal actual offset is obtained by subtracting the horizontal coordinate of the touch trajectory from the horizontal coordinate of the display handwriting, and the vertical actual offset is obtained by subtracting the vertical coordinate of the touch trajectory from the vertical coordinate of the display handwriting. A positive horizontal offset is recorded when the horizontal actual offset is greater than zero, no horizontal offset is recorded when the horizontal actual offset is equal to zero, and a negative horizontal offset is recorded when the horizontal actual offset is less than zero. The vertical actual offset is recorded according to the same sign rule. Subsequently, normalized capacitance data at the same scan time is read using the sampling timestamp of the matching coordinate pair. One set of normalized capacitance data, along with the horizontal and vertical actual offsets under the same sampling timestamp, is written into the same training record, maintaining the original two-dimensional node arrangement of the normalized capacitance data. Finally, an incremental learning sample composed of normalized capacitance data and actual offsets is obtained.

[0039] Next, the standard scores of each non-writing segment and each isolated segment, saved according to the scan cycle, are read. First, the maximum number of scan cycles in all non-writing segments and isolated segments is counted. For non-writing segments or isolated segments with a scan cycle count less than the maximum number of scan cycles, zero values ​​are padded to the end of the standard score sequence to ensure that each non-writing segment and isolated segment has the same number of time positions. Then, the standard scores of touch area, peak capacitance change, touch trajectory movement speed, and touch event duration are fixedly arranged according to each time position and expanded into non-writing event feature vectors according to time position. The single-class support vector machine uses a Gaussian radial basis function kernel to process the non-writing event feature vectors. The kernel function parameters are set by the pairwise Euclidean distances between all non-writing event feature vectors. First, the squared Euclidean distance between any two different non-writing event feature vectors is calculated. All squared Euclidean distances greater than zero are arranged in ascending order of value, and the median is obtained. Then, the reciprocal of the median is set as the kernel function parameter. The Gaussian radial basis function kernel is calculated according to the following formula. ;in, Indicates the current non-writing event feature vector and the first... Gaussian radial basis kernel function values ​​between the feature vectors of the training non-written events. This represents the feature vector of the non-written event of the current input. Indicates the first A training non-written event feature vector, This represents the index of the training non-written event feature vector. The kernel function parameters represent the Gaussian radial basis kernel. Represents the base of the natural exponential function. This represents the squared Euclidean distance between two non-written event feature vectors. The larger the kernel function parameter value, the faster the Gaussian radial basis function value decreases for the same feature distance. Support vector coefficients are set to the same non-negative number at the beginning of training, and the sum of all support vector coefficients is made 1. Subsequently, the support vector coefficients are iteratively adjusted, and the training target value is recalculated after each adjustment. Only the support vector coefficient combinations that continue to change the training target in the convergence direction are retained. Training non-written event feature vectors with support vector coefficients adjusted to zero are no longer used as support vectors; training non-written event feature vectors with support vector coefficients remaining greater than zero are retained as support vectors. The larger the support vector coefficient, the greater the contribution of the corresponding support vector's kernel function value to the discriminant function value. The discriminant function is established according to the following formula: ;in, The discriminant function value represents the feature vector of the input non-written event. This represents the feature vector of the input non-written event. This indicates the number of support vectors saved after training. Represents the support vector index. Indicates the first The support vector coefficients of each support vector. This indicates that the input non-written event feature vector is related to the first... Gaussian radial basis kernel function values ​​between the support vectors Indicates the first Support vectors, This represents the bias. The bias is adjusted along with the support vector coefficients during training, so that when the support vectors located at the training data boundary are substituted into the discriminant function, the function value is equal to zero. Finally, the non-written event decision boundary is obtained, which is defined by the support vectors, support vector coefficients, kernel function parameters, and bias.

[0040] S5: Build a convolutional neural network, then update the convolutional neural network using incremental learning samples. Combine the non-writing event decision boundary and the updated convolutional neural network to calibrate the touch trajectory, obtaining the touch trajectory calibration results, including:

[0041] During the initial training of the convolutional neural network (CNN), it reads the already formed normalized capacitance data and the actual offsets with the same sampling timestamp as the normalized capacitance data. The normalized capacitance data is input into the CNN in a two-dimensional arrangement of the blackboard conductive film capacitance sensing nodes. The actual horizontal and vertical offsets serve as training labels for the two output positions, respectively. The CNN is configured with an input layer, a first convolutional layer, a first max-pooling layer, a second convolutional layer, a second max-pooling layer, a fully connected layer, and a dual-output layer. The first convolutional layer has 16 three-row, three-column convolutional kernels, the second convolutional layer has 32 three-row, three-column convolutional kernels, the fully connected layer has 64 neurons, and the dual-output layer has 2 output nodes, corresponding to the horizontal and vertical predicted offsets, respectively. The three-row, three-column convolutional kernels cover the current capacitance sensing node and its surrounding adjacent positions position by position, with a convolution stride of 1. The max-pooling range is set to two rows and two columns, with a pooling stride of 2. Before the first training, the convolutional kernel weights and fully connected weights are initialized with a symmetrical range based on the number of inputs and outputs at the current connection position. As the number of inputs and outputs increases, the initial range decreases accordingly. Each network weight obtains its initial value from this symmetrical range. The larger the absolute value of the network weight, the greater the contribution of the input value multiplied by the network weight to the current weighted output. The convolutional bias and fully connected bias are initially set to zero. The first convolutional layer performs weighted accumulation of the normalized capacitance change in the three-row, three-column range with the corresponding convolutional kernel weights position by position and combines it with the convolutional bias. When the convolutional output is less than zero, it is recorded as zero after linear rectification. When the convolutional output is greater than or equal to zero, it retains its original value. The first max pooling layer selects the maximum convolutional output from each two-row, two-column range and enters the second convolutional layer. The second convolutional layer and the second max pooling layer are processed according to the same convolution, linear rectification, and maximum value selection process. Then, the output of the second max pooling layer is unfolded according to the fixed row, column, and convolutional channel order and input into 64 fully connected neural nodes. Finally, the horizontal prediction offset and the vertical prediction offset are formed by two linear output nodes. After initial training, the network weight parameters are saved. Before proceeding to the current incremental update, the currently saved kernel weights, convolutional biases, fully connected weights, and fully connected biases of the convolutional neural network are copied item by item to form backup network weight parameters. The mean squared error of a single incremental learning sample is calculated using the following formula: ;in, This represents the mean squared error of the current incremental learning sample. The horizontal coordinates of the output of the convolutional neural network The predicted offset, Represents the horizontal coordinates in the same incremental learning sample The actual offset. Represents the vertical coordinate of the output of the convolutional neural network The predicted offset, Represents the vertical coordinates in the same incremental learning sample The actual offset. Calculate the average mean squared error of all incremental learning samples, and then calculate the adjustment direction of the average mean squared error relative to each network weight parameter layer by layer from the dual output layer forward. The network weight update step size starts from 1 and attempts to complete one candidate update using the current network weight update step size. After that, recalculate the average mean squared error. If the updated average mean squared error is less than the original average mean squared error, retain the current network weight update. If the updated average mean squared error is greater than or equal to the original average mean squared error, halve the network weight update step size and re-execute the candidate update until a network weight update that reduces the average mean squared error is obtained. Finally, the updated network weight parameters are obtained, and a backup of the network weight parameters is retained.

[0042] Next, the conductive film on the blackboard continues to collect touch events to be calibrated according to the predetermined scanning cycle, reading the center coordinates of the touch area, touch feature sequence, capacitance change, and node coordinates formed in each scanning cycle of the touch events to be calibrated. The average and standard deviation of the touch area, the average and standard deviation of the peak capacitance change, the average and standard deviation of the touch trajectory movement speed, and the average and standard deviation of the touch event duration, all saved when the standard scores of the calibration sample touch events were generated, are used to calculate the standard scores for the four touch features in the touch events to be calibrated. A standard score is calculated when the standard deviation of a corresponding touch feature is greater than zero, and a standard score is recorded as zero when the standard deviation is equal to zero. Then, using the maximum number of scanning cycles and the order of identical fields determined during the training of the single-class support vector machine, the standard scores of the touch events to be calibrated are organized into feature vectors for the touch events to be calibrated. The single-class support vector machine reads the saved support vectors, support vector coefficients, kernel function parameters, and biases. First, it processes the feature vectors of the event to be calibrated and each support vector according to the Gaussian radial basis function established in section 12. Then, it weights and accumulates the kernel function values ​​according to the support vector coefficients, and combines this with the biases to form the discriminant value of the event to be calibrated. According to the original judgment criteria, if the discriminant value of the event to be calibrated is less than zero, the touch event to be calibrated is retained as a writing touch event; if the discriminant value is greater than or equal to zero, the touch event to be calibrated is registered as a non-writing touch event and discarded. For the retained writing touch events, the row position of the two-dimensional arrangement is determined according to the vertical position in the node coordinates, and the column position is determined according to the horizontal position in the node coordinates. The capacitance change in the same scan cycle is written to the corresponding node position, and then the arranged capacitance change is normalized to obtain the normalized capacitance data to be calibrated. The convolutional neural network loads the updated network weight parameters. The normalized capacitance data to be calibrated is processed sequentially through three rows and three columns of convolution, linear rectification, two rows and two columns of max pooling, second convolution, second max pooling, fully connected processing, and dual output processing. Finally, the touch trajectory coordinate offset to be calibrated is obtained, which consists of the horizontal touch trajectory coordinate offset to be calibrated and the vertical touch trajectory coordinate offset to be calibrated.

[0043] Finally, the center coordinates of the touch area of ​​the touch event to be calibrated are read, and the horizontal and vertical coordinates of the touch area center are converted into the initial touch trajectory horizontal and vertical coordinates in the blackboard display area coordinate system using the saved homography matrix and the planar projection relationship established in section 7. Then, the horizontal coordinates of the initial touch trajectory are added to the horizontal offset of the touch trajectory to be calibrated to obtain the calibrated horizontal coordinates, and the vertical coordinates of the initial touch trajectory are added to the vertical offset of the touch trajectory to be calibrated to obtain the calibrated vertical coordinates. The display handwriting coordinates with the same sampling timestamp as the calibrated touch trajectory coordinates are read according to the sampling timestamp, and the root mean square error of the current touch event to be calibrated is calculated. ;in, This represents the root mean square error of the current touch event to be calibrated. This indicates the number of matches between calibrated touch trajectory coordinates and displayed handwriting coordinates that have the same sampling timestamp. Indicates the matching coordinate index. Indicates the first The horizontal coordinate of the calibrated touch trajectory Indicates the first The vertical coordinate of the calibrated touch trajectory Indicates the relationship with the first After calibration, the display handwriting abscissa coordinates have the same sampling timestamp. Indicates the relationship with the first The calibrated touch trajectory coordinates are the display handwriting coordinates with the same sampling timestamp. The coordinate error threshold is set using a predetermined actual writing segment. The root mean square error (RMSE) between the touch trajectory coordinates of each actual writing segment and the display handwriting coordinates with the same sampling timestamp is calculated. The RMSEs of all actual writing segments are arranged in ascending order, and the largest RMSE is taken as the coordinate error threshold. According to the original judgment conditions, if the RMSE of the current touch event to be calibrated is less than or equal to the coordinate error threshold, the calibrated touch trajectory coordinates are used as the current touch trajectory calibration result. If the RMSE of the current touch event to be calibrated is greater than the coordinate error threshold, the convolutional neural network loads the backup network weight parameters and re-inputs the same normalized capacitance data to be calibrated. The horizontal and vertical offsets of the touch trajectory coordinates to be calibrated are recalculated, and then added to the initial horizontal and vertical coordinates of the touch trajectory to form the reverted touch trajectory coordinates. The reverted touch trajectory coordinates are then used as the current touch trajectory calibration result.

[0044] Please see Figure 3 A deep learning-based blackboard conductive thin film touch calibration system, the system comprising:

[0045] The touch feature extraction module determines the touch area of ​​the blackboard conductive film within the scanning cycle and acquires the touch feature sequence; wherein, the touch features include the area of ​​the touch area and the peak value of capacitance change, the movement speed of the touch trajectory, and the duration of the touch event;

[0046] The event segmentation module divides touch event segments according to the touch area and touch feature sequence, and determines the event time range and event coordinate range of the touch event segments; then, it selects candidate writing segments and non-writing segments from the touch event segments based on the pre-trained long short-term memory network.

[0047] The writing segment confirmation module obtains the coordinates of the touch calibration point in the blackboard conductive film touch coordinate system, which represents the touch position of the touch calibration point, and the coordinates of the display calibration point in the blackboard display area display coordinate system, which represents the display position of the display calibration point. Referring to the touch calibration point coordinates and the display calibration point coordinates, the actual writing segment is selected from the candidate writing segments, and the remaining candidate writing segments are used as isolated segments.

[0048] The boundary construction module builds incremental learning samples based on actual written fragments; it calculates the decision boundary for non-written events based on non-written fragments and isolated fragments.

[0049] The trajectory calibration module establishes a convolutional neural network, then updates the convolutional neural network using incremental learning samples, and calibrates the touch trajectory by combining the non-writing event decision boundary and the updated convolutional neural network to obtain the touch trajectory calibration result.

[0050] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A deep learning-based method for calibrating touch controls on a blackboard conductive thin film, characterized in that, Includes the following steps: S1: Determine the touch area of ​​the blackboard conductive film within the scanning cycle and obtain the touch feature sequence; wherein, the touch features include the area of ​​the touch area and the peak value of capacitance change, the movement speed of the touch trajectory, and the duration of the touch event; S2: Divide the touch event segments according to the touch area and touch feature sequence, and determine the event time range and event coordinate range of the touch event segments; then select candidate writing segments and non-writing segments from the touch event segments according to the pre-trained long short-term memory network; S3: Obtain the coordinates of the touch calibration point in the blackboard conductive film touch coordinate system, which represents the touch position of the touch calibration point, and the coordinates of the display calibration point in the blackboard display area display coordinate system, which represents the display position of the display calibration point. Refer to the touch calibration point coordinates and the display calibration point coordinates, select the actual writing segment from the candidate writing segments, and use the remaining candidate writing segments as isolated segments. S4: Construct incremental learning samples based on actual written fragments; calculate the decision boundary for non-written events based on non-written fragments and isolated fragments; S5: Build a convolutional neural network, then update the convolutional neural network using incremental learning samples, and combine the non-writing event decision boundary and the updated convolutional neural network to calibrate the touch trajectory and obtain the touch trajectory calibration result.

2. The deep learning-based blackboard conductive thin film touch calibration method according to claim 1, characterized in that, Step S1 specifically includes: The capacitance change, sampling timestamp, and node coordinates in the touch coordinate system of the capacitance sensing nodes corresponding to each calibration sample touch event on the blackboard conductive film are obtained in each scanning cycle. The touch detection threshold is obtained, and the capacitance sensing nodes whose capacitance change is greater than or equal to the touch detection threshold are taken as touch nodes. Adjacent touch nodes are connected to form a touch area. Calculate the ratio of the capacitance change of each touch node to the sum of the capacitance changes of all touch nodes to obtain the normalized weight corresponding to each touch node. Use the normalized weight to perform a weighted average on the coordinates of the corresponding node to obtain the center coordinates of the touch area. Obtain the area of ​​a single node. Calculate the product of the number of touch nodes and the area of ​​the single node to obtain the area of ​​the touch area. Take the largest capacitance change in the touch area as the peak capacitance change of the touch area. The ratio of the distance between the center coordinates of the touch area in adjacent scanning cycles to the time difference between the corresponding sampling timestamps is calculated to obtain the movement speed of the touch trajectory. The time difference between the first and last sampling timestamps corresponding to the same touch event is calculated to obtain the duration of the touch event. The area and capacitance change peaks of the touch area, the movement speed of the touch trajectory, and the duration of the touch event are arranged according to the scanning cycle to obtain the touch feature sequence.

3. The deep learning-based blackboard conductive thin film touch calibration method according to claim 2, characterized in that, Step S2 specifically includes: Standard scores are calculated for each touch feature in the touch feature sequence to obtain time and distance thresholds. The time difference between adjacent sampling timestamps and the center coordinate distance between the center coordinates of the touch area in adjacent scanning cycles are calculated. The scanning cycle number corresponding to the next sampling timestamp when the time difference is greater than the time threshold, or the scanning cycle number corresponding to the center coordinates of the next touch area when the center coordinate distance is greater than the distance threshold, are determined as segmentation positions. The touch feature sequence is divided according to these segmentation positions to obtain touch event segments. The corresponding standard score is determined based on the scanning cycle corresponding to each touch event segment. Each touch event segment includes the area of ​​the touch area corresponding to each scanning cycle between two adjacent segmentation positions, the peak capacitance change of the touch area, the movement speed of the touch trajectory, and the duration of the touch event. The event time range and event coordinate range of each touch event segment are determined based on the sampling timestamp and the center coordinates of the touch area, and the standard scores are arranged to obtain a standard score sequence. The standard score sequence is input into a pre-trained long short-term memory network to calculate the probability of writing, the probability of palm touch, and the probability of wiping. The touch event segment with the highest probability of writing is selected as a candidate writing segment, and the touch event segment with the highest probability of palm touch or wiping is selected as a non-writing segment.

4. The deep learning-based blackboard conductive thin film touch calibration method according to claim 3, characterized in that, The event time range is the time interval between the first sampling timestamp and the last sampling timestamp in the touch event segment, and the event coordinate range is the coordinate area defined by the minimum horizontal coordinate, maximum horizontal coordinate, minimum vertical coordinate, and maximum vertical coordinate corresponding to the center coordinates of each touch area in the touch event segment.

5. The deep learning-based blackboard conductive thin film touch calibration method according to claim 3, characterized in that, The Long Short-Term Memory Network is trained based on multiple sets of labeled touch samples. Each set of touch samples consists of the touch area, capacitance change peak, touch trajectory movement speed, and touch event duration corresponding to a continuous scanning cycle, and is labeled as writing category, palm touch category, or wiping category.

6. The deep learning-based blackboard conductive thin film touch calibration method according to claim 3, characterized in that, Step S3 specifically includes: Obtain the coordinates of the touch calibration point representing the touch position in the blackboard conductive film touch coordinate system and the coordinates of the display calibration point representing the display position in the blackboard display area display coordinate system, and calculate the homography matrix; The center coordinates of the touch area corresponding to each candidate writing segment are converted into touch trajectory coordinates using the homography matrix. The handwriting coordinates are collected according to the event time range corresponding to each candidate writing segment. The touch trajectory coordinates and the handwriting coordinates are matched with the sampling timestamp to obtain corresponding coordinate pairs. The Euclidean distance between the touch trajectory coordinates and the handwriting coordinates in the coordinate pairs and the average value of each Euclidean distance are calculated to obtain the average trajectory distance. Obtain a trajectory distance threshold, and use the candidate writing segments whose average trajectory distance is less than or equal to the trajectory distance threshold as actual writing segments, and use the remaining candidate writing segments as isolated segments.

7. The deep learning-based blackboard conductive thin film touch calibration method according to claim 6, characterized in that, Step S4 specifically includes: The capacitance change of the corresponding capacitance sensing node is extracted according to the event time range and event coordinate range corresponding to each actual writing segment. The capacitance change is arranged into two-dimensional capacitance data according to the corresponding node coordinates. Adjacent positions where the capacitance change is greater than or equal to the touch detection threshold are connected to form a connected region. The connected region containing the center coordinates of the touch area in the corresponding actual writing segment is taken as the handwriting connected region. The capacitance sensing nodes outside the handwriting connected region where the capacitance change is greater than or equal to the touch detection threshold are taken as shield nodes. The reference node closest to each shield node is selected from the handwriting connected region. The capacitance change of the reference node is weighted by the inverse distance to obtain the replacement capacitance change. The replacement capacitance change is used to replace the capacitance change of the corresponding shield node in the two-dimensional capacitance data. Then the replaced two-dimensional capacitance data is normalized to obtain normalized capacitance data. Based on the sampling timestamp, match the display handwriting coordinates and touch trajectory coordinates corresponding to each actual writing segment, calculate the difference between the display handwriting coordinates and the touch trajectory coordinates in each matched coordinate pair, obtain the actual offset of the touch trajectory coordinates, and then construct incremental learning samples based on the normalized capacitance data corresponding to each actual writing segment and the actual offset. A non-writing event feature vector is constructed based on the standard scores corresponding to each non-writing segment and each isolated segment. A single-class support vector machine is trained using the non-writing event feature vector to determine the non-writing event decision boundary.

8. The deep learning-based blackboard conductive film touch calibration method according to claim 7, characterized in that, The non-written event decision boundary is specifically the feature space boundary corresponding to the function value of the discriminant function, which is composed of support vectors, support vector coefficients, kernel function parameters, and bias, when the function value is equal to zero.

9. The deep learning-based blackboard conductive thin film touch calibration method according to claim 7, characterized in that, Step S5 specifically includes: A convolutional neural network is established, and the network weight parameters before the convolutional neural network update are saved as backup network weight parameters. The convolutional neural network is called to calculate the predicted offset of the corresponding touch trajectory coordinates based on the normalized capacitance data in each incremental learning sample. The network weight parameters are updated based on the mean square error between each predicted offset and the corresponding actual offset. The system obtains the center coordinates of the touch area corresponding to the touch event to be calibrated on the conductive film of the blackboard, the touch feature sequence, the capacitance change, and the node coordinates. It then recalculates the standard score based on the touch feature sequence to obtain the feature vector of the event to be calibrated. The feature vector of the event to be calibrated is input into a single-class support vector machine using the decision boundary of the non-writing event to obtain the discriminant value of the event to be calibrated. The touch events to be calibrated with a discriminant value less than zero are treated as writing touch events, while the remaining non-writing touch events are removed. The capacitance change of the writing touch events is arranged according to the node coordinates and normalized to obtain the normalized capacitance data to be calibrated. The normalized capacitance data to be calibrated is input into a convolutional neural network using the updated network weight parameters to calculate the coordinate offset of the touch trajectory to be calibrated. The center coordinates of the touch area corresponding to the touch event to be calibrated are converted into initial touch trajectory coordinates in the display coordinate system based on the homography matrix. The initial touch trajectory coordinates are added to the offset of the touch trajectory coordinates to be calibrated to obtain the calibrated touch trajectory coordinates in the display coordinate system. The root mean square error between the calibrated touch trajectory coordinates and the display handwriting coordinates with the same sampling timestamp is calculated to obtain a coordinate error threshold. When the root mean square error is less than or equal to the coordinate error threshold, the calibrated touch trajectory coordinates are used as the touch trajectory calibration result. When the root mean square error is greater than the coordinate error threshold, the convolutional neural network using the backup network weight parameters is called to recalculate the offset of the touch trajectory coordinates to be calibrated. The recalculated offset of the touch trajectory coordinates to be calibrated is added to the initial touch trajectory coordinates to obtain the backtracked touch trajectory coordinates, which are used as the touch trajectory calibration result.

10. A blackboard conductive thin film touch calibration system based on deep learning, characterized in that, The system includes: The touch feature extraction module determines the touch area of ​​the blackboard conductive film within the scanning cycle and acquires the touch feature sequence; wherein, the touch features include the area of ​​the touch area and the peak value of capacitance change, the movement speed of the touch trajectory, and the duration of the touch event; The event segmentation module divides touch event segments according to the touch area and touch feature sequence, and determines the event time range and event coordinate range of the touch event segments; then, it selects candidate writing segments and non-writing segments from the touch event segments based on the pre-trained long short-term memory network. The writing segment confirmation module obtains the coordinates of the touch calibration point in the blackboard conductive film touch coordinate system, which represents the touch position of the touch calibration point, and the coordinates of the display calibration point in the blackboard display area display coordinate system, which represents the display position of the display calibration point. Referring to the touch calibration point coordinates and the display calibration point coordinates, the actual writing segment is selected from the candidate writing segments, and the remaining candidate writing segments are used as isolated segments. The boundary construction module builds incremental learning samples based on actual written fragments; it calculates the decision boundary for non-written events based on non-written fragments and isolated fragments. The trajectory calibration module establishes a convolutional neural network, then updates the convolutional neural network using incremental learning samples, and calibrates the touch trajectory by combining the non-writing event decision boundary and the updated convolutional neural network to obtain the touch trajectory calibration result.

Images