Oil painting brush touch style analysis and teaching guidance system based on AI visual identification

Abstract

The invention relates to the field of cross application of artificial intelligence and computer vision, and particularly discloses an oil painting brush touch style analysis and teaching guidance system based on AI visual identification. The system comprises a multi-mode dynamic sensing unit, a stroke time sequence modeling unit, a style feature decoupling unit, a technique deviation diagnosis unit and an interactive teaching feedback unit. According to the method, a multi-modal dynamic sensing system is constructed, four-dimensional space-time quantization of a stroke generation event in the oil painting drawing process is achieved for the first time, and the fundamental limitation that the dynamic characteristics of wielding cannot be captured through traditional static image analysis is broken through; the system not only identifies the final form of the stroke, but also restores the speed change, the acceleration characteristic and the pressure evolution trajectory during generation, and provides a high-fidelity behavior basis for technical diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary applications of artificial intelligence and computer vision, specifically involving an oil painting brushstroke style analysis and teaching guidance system based on AI visual recognition. Background Technology

[0002] The integration of artificial intelligence and computer vision technology in art education is deepening, especially in the assessment of painting skills and teaching assistance, demonstrating enormous potential. Image recognition-based intelligent systems can classify paintings by style, analyze techniques, and evaluate quality, providing a data-driven new paradigm for personalized teaching. Oil painting, as an important art form emphasizing the expressiveness of brushstrokes, relies heavily on the dynamic characteristics of brush pressure, direction, speed, and layering methods in its creation process. These elements collectively constitute a unique visual language and artistic style.

[0003] AI-based visual recognition technology for analyzing oil painting brushstrokes aims to extract and quantify the morphological and kinematic features of brushstrokes from image or video input, thereby enabling modeling of the painter's style or diagnosis of the learner's technique. This technology typically combines deep learning models to encode features from high-dimensional visual data and attempts to establish a mapping between the appearance of brushstrokes and their generating actions, supporting subsequent teaching feedback generation.

[0004] Current technologies primarily rely on single-frame image analysis of completed static paintings. While they can identify brushstroke thickness, direction, and overlapping relationships, they fail to capture the temporal dynamics of the painting process, such as wrist movement inertia, brush speed variations, and the connection logic between consecutive brushstrokes. Therefore, instructional guidance systems struggle to provide accurate, real-time feedback based on learners' actual brushwork, leading to a disconnect between suggested content and practical application. This is particularly problematic in beginner training scenarios, where ignoring the temporal characteristics of movements prevents the system from recognizing brushstroke distortions caused by improper force application or rhythmic instability, thus weakening the effectiveness of instructional intervention. Therefore, there is an urgent need for an oil painting brushstroke style analysis and instructional guidance system that integrates dynamic visual perception and temporal behavior modeling. Summary of the Invention

[0005] The purpose of this invention is to provide an oil painting brushstroke style analysis and teaching guidance system based on AI visual recognition, so as to solve the problem in the prior art that the dynamic temporal information of the painting process cannot be obtained due to the reliance on static image analysis, resulting in a disconnect between teaching feedback and learners' actual brushwork behavior.

[0006] This invention provides an oil painting brushstroke style analysis and teaching guidance system based on AI visual recognition, including: a multimodal dynamic perception unit, a brushstroke temporal modeling unit, a style feature decoupling unit, a technique deviation diagnosis unit, and an interactive teaching feedback unit; The multimodal dynamic perception unit is used to synchronously collect high frame rate video streams, canvas surface depth images, and brush end motion trajectory data of the learner during the oil painting creation process. The pen stroke timing modeling unit is used to perform spatiotemporal alignment and fusion processing on the data output by the multimodal dynamic perception unit to construct a four-dimensional pen stroke event sequence that includes the pen stroke generation time, spatial position, movement speed, acceleration and pressure change. The style feature decoupling unit is used to separate the stable features representing the artistic style and the instantaneous features representing the operational habits from the four-dimensional brushstroke event sequence, and to calculate the style matching degree based on a preset database of classic painter styles. The technique deviation diagnosis unit is used to compare the learner's instantaneous characteristics with the standard technique template corresponding to the target style item by item, and identify structural deviations in pen rhythm, force mode, pen tip conversion or superposition order. The interactive teaching feedback unit is used to generate visual guidance instructions with action correction orientation based on the output results of the technique deviation diagnosis unit, and then superimposes them on the learner's canvas field of view in real time through augmented reality projection.

[0007] Furthermore, the multimodal dynamic sensing unit includes a high-speed industrial camera mounted above the easel, a flexible pressure sensor array embedded in the back of the canvas, and a miniature inertial measurement unit fixed to the brush handle. The high-speed industrial camera continuously captures high-definition video of the area where the brush tip contacts the canvas at a frame rate of more than 120 frames per second. The flexible pressure sensor array has pressure sensing points distributed at 10 mm intervals and a sampling frequency of 200 Hz, which is used to record the local pressure distribution map of each stroke on the canvas. The miniature inertial measurement unit includes a three-axis accelerometer and a three-axis gyroscope, with a sampling frequency of 500 Hz, and is used to output the angular velocity and linear acceleration data of the paintbrush in three-dimensional space in real time.

[0008] Furthermore, the pen stroke timing modeling unit performs the following processing flow: First, based on the pixel coordinates of the pen tip in the video frame of the high-speed industrial camera, combined with the physical size of the canvas, spatial calibration is performed to obtain the two-dimensional position sequence of the pen tip in the canvas coordinate system. Secondly, the pressure peak trigger signal of the flexible pressure sensor array is used as the start and end timestamp of the brushstroke to segment the position sequence and form discrete brushstroke segments. Next, the inertial measurement unit data for each stroke segment within the corresponding time period are interpolated and aligned to calculate the velocity vector, acceleration vector, and equivalent pressure value of the stroke at each millisecond. Finally, the above parameters are organized in chronological order into a structured four-dimensional pen stroke event sequence, where each event contains five fields: timestamp, spatial coordinates, velocity amplitude and direction, acceleration amplitude and direction, and equivalent pressure value.

[0009] Furthermore, the style feature decoupling unit adopts a two-branch convolutional recurrent neural network architecture; the first branch is a style encoder, which receives a four-dimensional brush stroke event sequence as input, extracts local temporal patterns through a one-dimensional convolutional layer, and then aggregates the global context through a bidirectional long short-term memory network to output a 128-dimensional style embedding vector. The second branch is the operation decoder, which reconstructs the original brushstroke event sequence with the same input and introduces adversarial loss constraints during the training phase, forcing the style encoder to retain only features that are strongly correlated with artistic style and are independent of individual operation noise. The style embedding vector is compared with the baseline style vector of each painter in the classic painter style database using cosine similarity calculation, and the top 3 matching painters and their similarity scores are output.

[0010] Furthermore, the database of classic painter styles pre-stores the benchmark style vectors of at least 50 representative oil painting masters; Each baseline style vector is obtained by performing temporal modeling of the same brushstrokes on reconstructed videos of the painting process of no less than 20 representative works of the painter, and then extracting and taking the average value through a style encoder. The restored video is generated by professional artists copying key areas of the original work in a controlled environment, ensuring the authenticity and repeatability of the dynamic features of the brushstrokes.

[0011] Furthermore, the technique deviation diagnosis unit has a built-in library of standard templates for different oil painting techniques; The standard template library includes five basic techniques: dry painting, wet painting, thick painting, thin painting, and scraping. Each technique has a defined standard brush rhythm range, pressure change curve, brush speed decay rate threshold, and upper limit of the time interval between adjacent brush strokes. When a learner selects a target style, the system automatically associates the two most commonly used techniques of that style with their corresponding standard templates. The technique deviation diagnosis unit classifies the learner's four-dimensional brushstroke event sequence according to technique type, aligns it with the corresponding standard template through dynamic time adjustment, and calculates the normalized deviation index of each dimension parameter. When the deviation index of any dimension exceeds the preset tolerance threshold, it is marked as a structural deviation.

[0012] Furthermore, the visual guidance instructions generated by the interactive teaching feedback unit include dynamic arrow trajectories, pressure heatmaps, and rhythmic pulse prompts; The dynamic arrow trajectory is projected onto the canvas surface as a green dashed line along the ideal path of the current stroke, and the arrow density increases with the increase of the ideal stroke speed; The pressure heatmap is overlaid on the corresponding area of ​​the canvas with semi-transparent color levels, where red indicates that pressure needs to be increased and blue indicates that pressure needs to be reduced. The rhythmic pulse cue appears as a periodic halo around the projection of the brush tip, with the halo flashing frequency synchronized with the standard brush stroke rhythm of the target technique. All guiding elements are rendered and overlaid in real time via an augmented reality projection module fixed in front of the easel, with a latency of less than 80 milliseconds.

[0013] Furthermore, the system operates within a hierarchical decision-making framework, comprising a perception layer, a cognition layer, and an execution layer; the perception layer consists of multimodal dynamic perception units and is responsible for raw data acquisition. The cognitive layer integrates the brushstroke temporal modeling unit, the style feature decoupling unit, and the technique deviation diagnosis unit to complete the reasoning from data to diagnostic conclusions; The execution layer is implemented by interactive instructional feedback units, which are responsible for transforming the output of the cognitive layer into actionable visual guidance; The three layers communicate via a low-latency message bus, ensuring an end-to-end response time of less than 150 milliseconds from the moment a stroke occurs to the presentation of feedback.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention, by constructing a multimodal dynamic perception system, has for the first time realized the four-dimensional spatiotemporal quantification of brushstroke generation events during the oil painting process, breaking through the fundamental limitation of traditional static image analysis that cannot capture the dynamic characteristics of brushstrokes; the system not only identifies the final form of the brushstroke, but also accurately restores the speed changes, acceleration characteristics and pressure evolution trajectory during its generation, providing high-fidelity behavioral evidence for technique diagnosis.

[0015] 2. By using a style feature decoupling mechanism, the stable core of artistic style is separated from the random noise of individual operations, so that style matching is no longer affected by learner's temporary mistakes, thus improving the robustness and accuracy of style recognition.

[0016] 3. The technique deviation diagnosis unit introduces a standardized template library built based on real restoration data and adopts a dynamic time regularization alignment strategy, which can accurately locate learners' structural errors in key dimensions such as rhythm, intensity, and connection, rather than just a vague evaluation based on surface form.

[0017] 4. The interactive teaching feedback unit uses augmented reality technology to transform abstract techniques into intuitive, real-time visual guidance, enabling learners to perceive and correct deviations in their actions during the creative process. This forms a closed-loop teaching mechanism of "perception-diagnosis-feedback-correction," which greatly improves the efficiency and scientific nature of oil painting skills training. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall technical solution architecture of the oil painting brushstroke style analysis and teaching guidance system based on AI visual recognition proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the style feature decoupling unit in this invention; Figure 3 This is a logical flow diagram of the multimodal dynamic sensing unit and the pen stroke timing modeling unit in this invention; Figure 4 This is a logical framework diagram of the standard template matching and structural deviation identification of the technique deviation diagnosis unit in this invention; Figure 5 This is a schematic diagram illustrating the generation and overlay principle of augmented reality visualization guidance instructions in the interactive teaching feedback unit of this invention; Figure 6 This is a schematic diagram of the hierarchical decision-making framework and low-latency data flow interaction relationship of the perception layer, cognition layer and execution layer in this invention. Detailed Implementation

[0019] Please refer to Figures 1 to 6 The overall technical architecture of the oil painting brushstroke style analysis and teaching guidance system based on AI visual recognition proposed in this invention is shown in the attached figure. Figure 1 As shown, it consists of five core functional modules: a multimodal dynamic perception unit, a brushstroke temporal modeling unit, a style feature decoupling unit, a technique deviation diagnosis unit, and an interactive teaching feedback unit, and achieves efficient collaboration through a hierarchical decision-making framework.

[0020] During system operation, the multimodal dynamic perception unit first performs high-precision synchronous acquisition of the learner's dynamic behavior during the oil painting creation process; Subsequently, the stroke timing modeling unit performs spatiotemporal alignment and structuring processing on the original multi-source data to generate a four-dimensional stroke event sequence. Based on this, the style feature decoupling unit separates the stable features of art style from the instantaneous features of individual operation, and completes the matching calculation with the database of classic painter styles; The technique deviation diagnosis unit compares the learner's actual penmanship behavior item by item based on the standard technique template associated with the target style to identify structural deviations. Ultimately, the interactive teaching feedback unit transforms the diagnostic results into augmented reality-based visual guidance instructions, projecting them onto the canvas surface in real time, thus forming a closed-loop teaching feedback mechanism.

[0021] Please refer to the attached document. Figure 3 The multimodal dynamic sensing unit includes a high-speed industrial camera mounted above the easel, a flexible pressure sensor array embedded in the back of the canvas, and a miniature inertial measurement unit fixed to the brush handle.

[0022] The high-speed industrial camera continuously captures high-definition video of the area where the brush tip contacts the canvas at a frame rate of more than 120 frames per second. The video resolution is 1920×1080 pixels, and the exposure time is controlled within 1 / 2000 second to effectively freeze motion blur in high-speed brush strokes.

[0023] The camera ensures temporal consistency of each frame through a global shutter mechanism and achieves vertical alignment with the canvas plane through a physical mounting bracket, with a field of view covering the entire standard 40 cm × 50 cm canvas area.

[0024] The flexible pressure sensing array is made of piezoresistive thin film material, with pressure sensing points evenly distributed at 10 mm intervals to form a 40×50 grid-like sensing matrix, with a total of 2000 sensing points.

[0025] The array's sampling frequency is set to 200 Hz, and each sensing point outputs an analog voltage signal of 0 to 5 volts, corresponding to a pressure range of 0 to 5 Newtons. After quantization by a 16-bit analog-to-digital converter, a two-dimensional pressure distribution map containing 2,000 pressure values ​​per frame is generated.

[0026] The miniature inertial measurement unit is integrated inside the pen handle and includes a three-axis accelerometer and a three-axis gyroscope. The sampling frequency is 500 Hz, the accelerometer range is ±16g, and the gyroscope range is ±2000 degrees per second. The output data is transmitted to the main control processor via a digital I2C interface after temperature compensation and zero bias calibration.

[0027] The three sub-components achieve strict time synchronization through hardware trigger signals: The frame start signal of the high-speed industrial camera serves as the master clock source, synchronously triggering the data acquisition of the flexible pressure sensing array and the miniature inertial measurement unit, ensuring that the three data streams are aligned on a millisecond time scale.

[0028] After receiving the above three synchronous data streams, the pen stroke timing modeling unit executes a four-stage processing flow.

[0029] The first stage is spatial calibration: Based on the pixel coordinates of the brush tip in the video frame of a high-speed industrial camera, combined with the known physical size of the canvas (400 mm × 500 mm) and the camera intrinsic parameter matrix, the pixel coordinates are mapped to the two-dimensional physical position in the canvas coordinate system through a perspective transformation model.

[0030] This mapping relationship is determined using a pre-calibrated checkerboard pattern, with the calibration error controlled within 0.2 mm. This yields a continuous time-position sequence. ,in For the first Frame timestamp and These are the corresponding canvas coordinates.

[0031] The second stage is stroke segmentation: By utilizing the pressure distribution map output by the flexible pressure sensor array, the presence of local pressure peaks at each moment can be detected.

[0032] When the pressure value at a certain sensing point is greater than a preset threshold (e.g., 0.5 Newtons) and the duration is greater than 10 milliseconds, it is determined as the start of a valid stroke; when the pressure value falls back below the threshold and lasts for 20 milliseconds, it is determined as the end of the stroke.

[0033] In this way, a continuous sequence of positions is divided into several discrete stroke segments, each segment corresponding to a complete stroke action.

[0034] The third stage is multi-source data alignment: for each stroke segment, extract its start and end time intervals. , The initial time, The cutoff time is set; and the raw acceleration and angular velocity data output by the micro inertial measurement unit within this interval are linearly interpolated to align its time resolution with the 200 Hz sampling rate of the pressure sensing array.

[0035] Subsequently, by numerically integrating the acceleration data, the instantaneous velocity vector of the brush in three-dimensional space is calculated; then, combined with the contact geometry between the brush tip and the canvas, it is projected onto the canvas plane to obtain the two-dimensional velocity vector. , yes Directional components, yes Directional component, velocity amplitude ,direction .

[0036] Simultaneously, the three-axis acceleration vectors are combined to obtain the total acceleration amplitude. . , , The outputs of the miniature inertial measurement unit are respectively axis, axis, Triaxial acceleration components in the axial direction, equivalent pressure value This is calculated using the average pressure within the projection area of ​​the flexible pressure sensor array on the pen tip, i.e., for... The pressure values ​​of all sensing points within a circular area with a radius of 5 mm centered are taken as the arithmetic mean. This represents the two-dimensional coordinate position of the pen tip on the canvas plane. In order to be in The x-coordinate of the point In order to be in The ordinate of the point.

[0037] The fourth stage is the construction of the event sequence: The above parameters are organized into a structured four-dimensional pen stroke event sequence with a time granularity of 1 millisecond, where each event contains five fields: timestamp. (Unit: milliseconds), spatial coordinates (Unit: mm), velocity amplitude (Unit: millimeters per second) and direction (Unit: radians), acceleration amplitude (Unit: millimeters per second squared) and direction (Calculated from three-dimensional acceleration vector projection), and equivalent pressure value (Unit: Newtons). This sequence is stored in JSON format, supporting efficient parsing and processing by subsequent modules.

[0038] Please refer to the attached document. Figure 2 The style feature decoupling unit adopts a dual-branch convolutional recurrent neural network architecture, which aims to separate the stable features of artistic style from the instantaneous features of individual operations from the four-dimensional brush stroke event sequence.

[0039] The input to this unit is the previously constructed sequence of brushstroke events. First, its five-dimensional feature vector is normalized to the [0,1] interval, and then arranged in chronological order into a sequence of variable length. ,in The duration of the stroke (in milliseconds). It is arranged in chronological order. The single stroke event characteristic.

[0040] The first branch is the style encoder, which consists of three one-dimensional convolutional layers and one bidirectional long short-term memory network.

[0041] The first convolutional layer has a kernel size of 7 and 64 output channels; the second convolutional layer has a kernel size of 5 and 128 output channels; the third convolutional layer has a kernel size of 3 and 256 output channels.

[0042] Each convolutional layer employs the ReLU activation function and batch normalization to extract local temporal patterns, such as short-term acceleration, pressure abrupt changes, or directional reversals.

[0043] Subsequently, a bidirectional long short-term memory network performs global context aggregation on the convolutional features. After concatenating the forward and backward hidden states, the dimensionality is reduced to 128 dimensions through a fully connected layer, outputting a style embedding vector. ; For the set of real numbers, It represents a 128-dimensional real space.

[0044] The second branch is the operation decoder, which consists of a single-layer unidirectional long short-term memory network and three deconvolutional layers. Its goal is to... The original input sequence S is reconstructed conditionally. During the training phase, the system introduces adversarial loss constraints: An additional discriminator network is trained to attempt to distinguish between the real brushstroke sequence and the decoder-reconstructed sequence.

[0045] By minimizing the reconstruction loss of the decoder and maximizing the discrimination difficulty of the discriminator, the style encoder is forced to retain only features that are strongly correlated with artistic style and independent of individual operational noise.

[0046] This training strategy ensures It exhibits a high degree of consistency across different works by the same painter, while demonstrating robustness in imitating the same style by different learners.

[0047] The database of classic painters' styles pre-stores baseline style vectors for at least 50 representative oil painting masters, with each painter corresponding to a 128-dimensional baseline vector. The baseline vector was obtained by extracting and averaging the brushstrokes from reconstructed videos of the painting process of at least 20 of the painter's representative works.

[0048] The restoration videos are generated by professional artists who, in a controlled experimental environment, meticulously copy key areas of the original artwork (such as facial features, clothing textures, or landscape brushstrokes). The copying process strictly adheres to the techniques described in historical documents and uses paints and canvas materials similar to the original artwork. Each restoration video is captured by a multimodal dynamic perception unit and processed into a four-dimensional brushstroke event sequence by a brushstroke temporal modeling unit before being input into a style encoder to obtain a single-time style embedding vector.

[0049] The arithmetic mean of the embedding vectors obtained from more than 20 independent imitations is used to obtain the final embedding vector. Style matching is calculated using the cosine similarity formula: ; The formula calculates the cosine of the angle between the style embedding vector of the learner's current brushstroke sequence and the baseline vectors of each painter in the database. The output range is [-1, 1], and the larger the value, the closer the styles are.

[0050] The system returns the top 3 painters with the highest similarity scores and their corresponding scores, and locks the technique templates to be used for subsequent diagnosis based on the target style selected by the user (such as "Van Gogh style").

[0051] Please refer to the attached document. Figure 4 The technique deviation diagnosis unit has a built-in library of standard templates for different oil painting techniques, including five basic techniques: dry painting, wet painting, impasto, thin painting, and scraping.

[0052] Each technique category is defined with four dimensions of standardized parameters: The range of pen stroke rhythm (unit: strokes per second), pressure change curve (pressure as a function of time), pen speed decay rate threshold (upper limit of speed decrease rate, unit: millimeters per square second), and upper limit of time interval between adjacent strokes (unit: milliseconds).

[0053] For example, the standard template for the thick coating method specifies a brush stroke rhythm of 1.5 to 2.5 strokes per second, with pressure changes showing an initial increase followed by stabilization, a brush speed decay rate of less than 50 millimeters per square second, and an interval of less than 800 milliseconds between adjacent strokes.

[0054] When a learner selects a target style, the system automatically associates it with the two most commonly used techniques of that style (e.g., Van Gogh style is associated with impasto and dry brush techniques) and loads the corresponding standard templates.

[0055] The technique deviation diagnosis unit first categorizes the learner's four-dimensional brushstroke event sequence according to technique type: By analyzing the duration, pressure amplitude, and speed variation patterns of brushstrokes, a support vector machine classifier is used to determine the technique category to which each brushstroke belongs.

[0056] Subsequently, each type of stroke sequence is dynamically time-aligned with the corresponding specification template.

[0057] The dynamic time warping algorithm constructs a cost matrix and finds the optimal nonlinear alignment path, enabling learners' strokes and templates to achieve flexible matching on the time axis.

[0058] After alignment, calculate the normalized deviation index for each dimension parameter: ; Indicates the first Several dimensions (such as speed, pressure, etc.). This represents the number of aligned sampling points. For the first Under the first dimension, the actual data collected was the first... The value of each sampling point For the first Under the first dimension, the first in the specification template The standard reference values ​​for each sampling point (belonging to the template data sequence). This represents the standard deviation of this dimension within the canonical template.

[0059] Normalized bias index The influence of dimensions is eliminated, making cross-dimensional comparisons easier.

[0060] The system presets tolerance thresholds for each dimension (such as...). (Considered as exceeding the standard) When the deviation index of any dimension is greater than the threshold, it is marked as a structural deviation, and the deviation type, the time of occurrence, and the severity level (mild, moderate, severe) are recorded.

[0061] Please refer to the attached document. Figure 5 The interactive teaching feedback unit generates three types of visual guidance instructions based on the output of the technique deviation diagnosis unit: Dynamic arrow trajectory, pressure heatmap, and rhythmic pulse cues. The dynamic arrow trajectory is projected onto the canvas surface as a green dashed line along the ideal path of the current stroke. The ideal path is generated from a sequence of spatial coordinates aligned in the standard template.

[0062] Arrow density increases with the ideal pen speed; for example, at a speed of 100 millimeters per second, one arrow is drawn every 10 millimeters. At a speed of 200 millimeters per second, draw an arrow every 5 millimeters to visually reflect the speed of the stroke.

[0063] The pressure heatmap is overlaid on the corresponding area of ​​the canvas with semi-transparent color levels, using a four-color gradient mapping of red-yellow-green-blue: red (>1.2 times the target pressure) indicates that the pressure needs to be reduced, orange (1.0 to 1.2 times) indicates slight overpressure, green (0.8 to 1.0 times) indicates the ideal pressure range, and blue (<0.8 times) indicates that the pressure needs to be increased. The transparency of the heatmap is set to 60% to avoid obscuring the original artwork.

[0064] The rhythmic pulse cue appears as a periodic halo around the brush tip projection, with a diameter of 15 mm and an initial transparency of 80%. It decays exponentially to 0 over time, with the decay period synchronized with the standard brush stroke rhythm of the target technique.

[0065] For example, if the standard rhythm is 2 strokes per second, the halo will flash once every 500 milliseconds.

[0066] All guiding elements are rendered in real time by an augmented reality projection module fixed in front of the easel. This module uses DLP projection technology with a resolution of 1280×800 and a refresh rate of 60 Hz. It tracks the real-time position of the brush tip through an infrared positioning system to ensure that the projected content is precisely aligned with the physical coordinates of the canvas.

[0067] By optimizing the rendering pipeline and hardware acceleration, the system can control the latency from the occurrence of a brush stroke to the presentation of feedback to within 80 milliseconds.

[0068] Please refer to the attached document. Figure 6 The entire system operates within a hierarchical decision-making framework, comprising a perception layer, a cognition layer, and an execution layer. The perception layer, composed of multimodal dynamic sensing units, is responsible for acquiring raw data, and its output is transmitted to the cognition layer via gigabit Ethernet.

[0069] The cognitive layer integrates the brushstroke temporal modeling unit, style feature decoupling unit, and technique deviation diagnosis unit, and is deployed on a high-performance embedded GPU platform (such as NVIDIA Jetson AGX Orin) to complete end-to-end reasoning from raw data to diagnostic conclusions.

[0070] The execution layer is implemented by an interactive teaching feedback unit, which runs on an independent real-time operating system node to ensure the determinism and timeliness of graphics rendering.

[0071] The three layers communicate with each other via a low-latency message bus, using a publish-subscribe pattern to transmit structured data packets.

[0072] The data packet contains a timestamp, event type, parameter payload, and checksum. The transmission protocol is based on UDP optimization, and the packet loss retransmission mechanism is triggered only when critical diagnostic results are lost.

[0073] The system uses timestamp alignment and sliding window buffering to ensure that the end-to-end response time from the occurrence of a pen stroke to the presentation of feedback is less than 150 milliseconds, which meets the physiological perception threshold requirements for real-time teaching interaction.

[0074] In actual teaching scenarios, after learners start the system, they first select a target learning style (such as "Monet's Water Lilies style"), and the system automatically loads the corresponding technique templates (wet-on-wet and thin-on-thin techniques). Learners begin painting, and the multimodal dynamic sensing unit continuously collects data. The stroke timing modeling unit constructs event sequences in real time; The style feature decoupling unit updates the style matching result after each local area (such as a petal) is drawn; The technique deviation diagnosis unit compares each brushstroke in real time. Once a structural deviation is detected (such as an interval of more than 1200 milliseconds between adjacent brushstrokes in the wet-on-wet technique, which causes the paint to dry too quickly), the interactive teaching feedback unit is immediately triggered to generate a correction instruction.

[0075] For example, a blue heatmap is projected onto the canvas to indicate insufficient pressure, while a green arrow trail guides the correct direction of the next stroke, and a halo flashes at a frequency of 1.8 times per second to indicate the ideal rhythm.

[0076] Learners adjust their penmanship in real time based on guidance, and the system continuously monitors and corrects the results, forming a closed loop of "perception-diagnosis-feedback-correction". This mechanism transforms abstract techniques into perceptible and actionable visual cues, lowering the learning threshold and improving training efficiency.

[0077] The system also possesses adaptive learning capabilities: if the same learner repeatedly exhibits the same deviation in a certain dimension (such as pressure control in thick painting), the system will automatically increase the feedback intensity for that dimension (e.g., increase the transparency of the heatmap to 80%) and prioritize strengthening related training content in subsequent lessons. Furthermore, all brushstroke event sequences and diagnostic records are encrypted and stored in a local database, supporting post-lesson playback and long-term skill development trajectory analysis. Teachers can view the overall technique mastery heatmap for their class through the management interface, identify common problems, and optimize teaching plans.

[0078] In summary, this embodiment constructs a high-precision, low-latency, and highly interactive oil painting teaching and guidance system through five core technical aspects: multimodal dynamic perception, four-dimensional brushstroke modeling, style feature decoupling, standardized template comparison, and augmented reality feedback. It fundamentally solves the technical bottlenecks of traditional static analysis methods, such as the inability to capture brushstroke dynamics, delayed feedback, and vague guidance, and provides a complete and engineering-practical solution for the intelligentization of art education.

[0079] In another implementation, the system further expands the dimensions of multimodal perception to enhance the ability to analyze complex oil painting techniques.

[0080] Specifically, the multimodal dynamic sensing unit adds a near-infrared spectral imaging module and an environmental temperature and humidity sensor array to the existing high-speed industrial camera, flexible pressure sensing array, and miniature inertial measurement unit.

[0081] The near-infrared spectroscopy imaging module operates in the 900 to 1700 nanometer band and simultaneously captures the moisture content and oil content distribution of the pigment layer on the canvas surface at a frame rate of 30 frames per second.

[0082] This module analyzes the reflectance changes at specific wavelengths to invert the degree of drying of the pigment layer, thereby helping to determine whether the time window for pigment layering in wet painting is reasonable.

[0083] The ambient temperature and humidity sensor array consists of four distributed sensors located at the four corners of the easel, with a sampling frequency of 1 Hz, used to monitor the studio temperature (accuracy ±0.5 degrees Celsius) and relative humidity (accuracy ±2%).

[0084] These environmental parameters are incorporated into the correction factors of the technique deviation diagnosis unit: for example, in a high temperature and low humidity environment, the paint dries faster, and the system automatically shortens the upper limit of the time interval between adjacent brushstrokes in wet painting from the standard 1000 milliseconds to 700 milliseconds to adapt to actual physical conditions.

[0085] The brushstroke temporal modeling unit is correspondingly expanded to a five-dimensional event sequence, with a new "pigment state" field added. Its value is obtained by mapping near-infrared spectral data through a pre-trained regression model, and its range is 0 (completely wet) to 1 (completely dry). The input sequence of the style feature decoupling unit is adjusted to six dimensions (including timestamps), but the network structure of the style encoder remains unchanged. Only during the training phase is pigment state introduced as an auxiliary supervision signal to enhance the recognition ability of styles such as wet-on-wet painting that depend on the physical state of the pigment.

[0086] The interactive teaching feedback unit also includes a "drying warning" feature: when the system predicts that the paint in the current brushstroke area will become non-layerable within the next 2 seconds, a slowly rotating yellow warning ring is projected at the edge of that area, reminding learners to complete the next brushstroke in time. This warning mechanism reduces technical errors caused by environmental factors and improves the system's adaptability and robustness in uncontrolled teaching environments.

[0087] Furthermore, the database of classic painter styles has been expanded to include environmental adaptation parameters. Each baseline style vector is associated with a set of typical creative environment conditions (e.g., Monet's preference for high humidity environments), ensuring that style matching is based not only on brushstroke dynamics but also on the consistency of the environmental context. This embodiment, while retaining the core architecture, further enhances the system's practicality and accuracy in real-world teaching scenarios by expanding the perception dimension and introducing an environmental adaptation mechanism.

[0088] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0089] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A system for analyzing and teaching oil painting brushstroke styles based on AI visual recognition, characterized in that, include: The multimodal dynamic sensing unit is used to simultaneously acquire high frame rate video streams, canvas surface depth images, and brush end motion trajectory data of learners during the oil painting creation process. The stroke timing modeling unit is used to perform spatiotemporal alignment and fusion processing on the data output by the multimodal dynamic perception unit, and to construct a four-dimensional stroke event sequence that includes the stroke generation time, spatial position, movement speed, acceleration and pressure changes. The style feature decoupling unit is used to separate the stable features representing artistic style and the instantaneous features representing operational habits from the four-dimensional brush stroke event sequence, and to calculate the style matching degree based on a preset classic painter style database. The technique deviation diagnosis unit is used to compare the learner's instantaneous characteristics with the standard technique template corresponding to the target style item by item, and identify structural deviations in pen rhythm, force mode, pen tip transition or stacking order. The interactive teaching feedback unit generates visual guidance instructions with action correction orientation based on the output of the technique deviation diagnosis unit, and overlays them on the learner's canvas field of view in real time through augmented reality projection.

2. The oil painting brushstroke style analysis and teaching guidance system based on AI visual recognition according to claim 1, characterized in that, The multimodal dynamic sensing unit includes a high-speed industrial camera mounted above the easel, a flexible pressure sensor array embedded in the back of the canvas, and a miniature inertial measurement unit fixed to the brush handle. The high-speed industrial camera continuously captures high-definition video of the area where the brush tip contacts the canvas. The flexible pressure sensor array is distributed with pressure sensing points to record the local pressure distribution map of each stroke on the canvas. The miniature inertial measurement unit includes a three-axis accelerometer and a three-axis gyroscope, with a sampling frequency of 500 Hz, and is used to output the angular velocity and linear acceleration data of the paintbrush in three-dimensional space in real time.

3. The oil painting brushstroke style analysis and teaching guidance system based on AI visual recognition according to claim 2, characterized in that, The stroke timing modeling unit performs the following processing flow: Based on the pixel coordinates of the pen tip in the video frame of a high-speed industrial camera, spatial calibration is performed in combination with the physical size of the canvas to obtain the two-dimensional position sequence of the pen tip in the canvas coordinate system. The pressure peak trigger signal of the flexible pressure sensor array is used as the start and end timestamp of the brush stroke to segment the position sequence and form discrete brush stroke segments. Interpolate and align the inertial measurement unit data for each stroke segment within the corresponding time period to calculate the velocity vector, acceleration vector, and equivalent pressure value of the stroke at each millisecond. The above parameters are organized in chronological order into a structured four-dimensional pen stroke event sequence, where each event contains five fields: timestamp, spatial coordinates, velocity amplitude and direction, acceleration amplitude and direction, and equivalent pressure value.

4. The oil painting brushstroke style analysis and teaching guidance system based on AI visual recognition according to claim 1, characterized in that, The style feature decoupling unit adopts a dual-branch convolutional recurrent neural network architecture; The first branch is the style encoder, which receives a four-dimensional brush stroke event sequence as input, extracts local temporal patterns through a one-dimensional convolutional layer, and then aggregates the global context through a bidirectional long short-term memory network to output a style embedding vector. The second branch is the operation decoder, which reconstructs the original brushstroke event sequence with the same input and introduces adversarial loss constraints during the training phase, forcing the style encoder to retain only features that are strongly correlated with artistic style and are independent of individual operation noise. The style embedding vector is compared with the baseline style vector of each painter in the classic painter style database using cosine similarity calculation, and the top 3 matching painters and their similarity scores are output.

5. The oil painting brushstroke style analysis and teaching guidance system based on AI visual recognition according to claim 4, characterized in that, The database of classic painter styles pre-stores the benchmark style vectors of several representative oil painting masters; Each baseline style vector is obtained by performing temporal modeling of the same brushstrokes in the reconstruction video of the painting process of multiple representative works of the painter, and then extracting and taking the average value through the style encoder; The restored video is generated by professional artists copying key areas of the original work in a controlled environment, ensuring the authenticity and repeatability of the dynamic features of the brushstrokes.

6. The oil painting brushstroke style analysis and teaching guidance system based on AI visual recognition according to claim 1, characterized in that, The technique deviation diagnosis unit has a built-in library of standard templates for different oil painting techniques; The standard template library includes five basic techniques: dry painting, wet painting, thick painting, thin painting, and scraping. Each technique has a defined standard brush rhythm range, pressure change curve, brush speed decay rate threshold, and upper limit of the time interval between adjacent brush strokes. When a learner selects a target style, the system automatically associates the two most commonly used techniques of that style with their corresponding standard templates. The technique deviation diagnosis unit categorizes the learner's four-dimensional brushstroke event sequence according to technique type, performs dynamic time warping and alignment with the corresponding standard template, and calculates the normalized deviation index of each dimension parameter. When the deviation index of any dimension exceeds the preset tolerance threshold, it is marked as a structural deviation.

7. The oil painting brushstroke style analysis and teaching guidance system based on AI visual recognition according to claim 6, characterized in that, The normalized deviation index is obtained by calculating the mean absolute deviation of each dimension parameter after the learner's stroke sequence is aligned with the canonical template, and dividing it by the standard deviation of that dimension in the canonical template. The system presets tolerance thresholds for each dimension. When the normalized deviation index is greater than the threshold, it determines that there is a structural deviation and records the deviation type, the time of occurrence, and the severity level.

8. The oil painting brushstroke style analysis and teaching guidance system based on AI visual recognition according to claim 1, characterized in that, The interactive teaching feedback unit generates visual guidance instructions including dynamic arrow trajectories, pressure heatmaps, and rhythm pulse prompts. The dynamic arrow trajectory is projected onto the canvas surface as a green dashed line along the ideal path of the current stroke, and the arrow density increases with the increase of the ideal stroke speed; The pressure heatmap is overlaid on the corresponding area of ​​the canvas with semi-transparent color levels, where red indicates that pressure needs to be increased and blue indicates that pressure needs to be reduced. The rhythmic pulse cue appears as a periodic halo around the projection of the brush tip, with the halo flashing frequency synchronized with the standard brush stroke rhythm of the target technique. All guiding elements are rendered and overlaid in real time via an augmented reality projection module fixed in front of the easel.

9. The oil painting brushstroke style analysis and teaching guidance system based on AI visual recognition according to claim 8, characterized in that, The pressure heat map uses a four-color gradient mapping of red-yellow-green-blue.

10. The oil painting brushstroke style analysis and teaching guidance system based on AI visual recognition according to claim 1, characterized in that, The system operates within a hierarchical decision-making framework, comprising a perception layer, a cognition layer, and an execution layer. The perception layer consists of multimodal dynamic perception units and is responsible for raw data acquisition; The cognitive layer integrates the brushstroke temporal modeling unit, the style feature decoupling unit, and the technique deviation diagnosis unit to complete the reasoning from data to diagnostic conclusions; The execution layer is implemented by interactive instructional feedback units, which are responsible for transforming the output of the cognitive layer into actionable visual guidance; The three layers communicate with each other via a low-latency message bus.

Images