An AI-based marking image intelligent collection system, method and terminal device

Abstract

The present application relates to the technical field of examination image acquisition optimization, in particular to an AI-based examination image intelligent acquisition system, method and terminal device, the present application acquires original images through an image acquisition unit, a light source analysis unit generates a color temperature distribution heat map using a convolutional neural network, a dynamic compensation unit divides compensation grids accordingly and performs exposure compensation, an adaptive correction unit outputs a target image through bilinear interpolation and color space conversion, and a quality assessment unit triggers a hierarchical retake mechanism by comparing and positioning the color gamut value, solving the problems of insufficient complexity light adaptation and poor image quality stability in traditional technology, through the cooperation of multiple units driven by AI, high light-shade intelligent compensation, color space adaptive conversion and quality closed-loop control of the answer sheet image are realized, the accuracy of fill-in-the-blank point recognition and the image qualification rate are improved, and the efficient examination needs of large-scale examinations are met.

Description

Technical Field

[0001] The present invention relates to the technical field of optimization of examination paper marking image acquisition, and in particular to an AI-based examination paper marking image intelligent acquisition system, method and terminal device. Background Art

[0002] Marking image acquisition is an important technology. In modern educational examination and assessment scenarios, marking image acquisition technology, by acquiring answer sheet images and optimizing their quality, is the core foundation for achieving automatic scoring of objective questions and intelligent correction of subjective questions.

[0003] With the popularization of mobile marking equipment and the expansion of examination scale, building an intelligent image acquisition system that is adaptive to lighting environments is of key significance to improving marking efficiency and ensuring scoring accuracy. The traditional fixed parameter acquisition mode has been unable to adapt to the needs of high-quality image acquisition under complex lighting conditions. However, the traditional marking image acquisition technology has the core problems of insufficient adaptability to complex lighting and poor image quality stability. The existing scheme adopts fixed exposure parameters and has not established a mechanism for lighting area recognition and dynamic compensation. When the answer sheet is partially reflected by strong light or covered by shadows, such as direct light from the examination room window or blocked by the examinee's hand, the pixel value of the highlight area is saturated, resulting in an increase in the error rate of filling point recognition, and the lack of contrast in the shadow area increases the blurring rate of handwriting. The single color space conversion method cannot take into account both light and shadow. To meet the different needs of character recognition and fill-in point detection, the positioning mark offset error exceeds 10 pixels when converted to RGB space due to the coupling of brightness and chromaticity, affecting the accuracy of subsequent image registration. In addition, the traditional system lacks an image quality closed-loop evaluation mechanism. When the ambient light color temperature changes suddenly, such as when the lights are turned on in the examination room, the captured image color gamut deviates from the standard value and exceeds the preset threshold, but it continues to be used, which increases the misjudgment rate of the automatic scoring system and reduces the efficiency of manual reshooting. This lack of illumination adaptability and insufficient quality closed-loop mechanism ultimately leads to a low image pass rate of the traditional solution in complex examination environments. The reliability and efficiency of the marking system are difficult to meet the needs of large-scale examinations. To solve this problem, we provide an AI-based intelligent marking image acquisition system, method and terminal device. Summary of the Invention

[0004] The purpose of the present invention is to provide an AI-based intelligent image acquisition system, method and terminal device for paper marking to solve the problems raised in the above background technology.

[0005] Due to the complex lighting adaptability of traditional technologies, highlight and shadow areas are not well processed, resulting in a high recognition error rate. Therefore, this case uses a light source analysis unit to identify highlight and shadow areas, and a dynamic compensation unit to perform grid division and exposure compensation, which can reduce the fill point recognition error rate and the handwriting blur rate.

[0006] Due to the poor image quality stability of traditional systems and the lack of a closed-loop evaluation mechanism, resulting in a high misjudgment rate, this case uses the quality assessment unit to extract positioning mark area data and trigger a graded retake mechanism, which can improve the image qualification rate and reduce the misjudgment rate of the automatic scoring system.

[0007] To achieve the above objectives, an AI-based intelligent image acquisition system for examination paper marking is provided, which includes the following units:

[0008] The image acquisition unit obtains the original image of the answer sheet through the mobile terminal;

[0009] The light source analysis unit identifies the highlight reflection areas and shadow coverage areas in the original image based on the convolutional neural network and generates a color temperature distribution heat map;

[0010] The dynamic compensation unit divides the original image into N*N compensation grids according to the color temperature distribution heat map. The exposure compensation algorithm based on regional color temperature perception calculates the brightness compensation coefficient and hue offset for each compensation grid.

[0011] The adaptive correction unit uses a bilinear interpolation algorithm to process the brightness compensation coefficients and hue offsets of adjacent grids, and performs color gamut mapping conversion based on the HSV color space on the adjacent grids to output the target image;

[0012] The quality assessment unit extracts pixel data of the positioning mark area in the target image, calculates its corrected color gamut value, and compares it with the pre-stored standard color gamut value to generate an image usability score and trigger a graded retake mechanism. When the lighting condition changes by more than a threshold, the graded retake mechanism automatically switches to a multi-frame synthesis mode, obtains an exposure difference image group through three samplings, and selects the frame with the best contrast in the filled area in the exposure difference image group as the output result.

[0013] As a further improvement of the present technical solution, the light source analysis unit performs the following steps to generate a color temperature distribution heat map:

[0014] Multi-scale Gaussian filtering is performed on the original image to separate high-frequency texture and low-frequency illumination components. The pre-trained lightweight convolutional neural network model is used to identify the positioning mark edges in the high-frequency texture component and mark them as illumination-invariant areas.

[0015] The low-frequency illumination component is input into the convolutional neural network structure of the convolutional neural network model, and the highlight reflection area and the shadow coverage area are output;

[0016] The high-light reflection area and the shadow coverage area are corrected by void convolution, and the position information of the illumination-invariant area is fused to generate a color temperature distribution heat map.

[0017] As a further improvement of this technical solution, the specific operation of the dynamic compensation unit is as follows:

[0018] The compensation grid density is divided according to the color temperature gradient change amplitude of the color temperature distribution heat map. When the color temperature gradient change amplitude is greater than or equal to the color temperature gradient change threshold, a non-uniform encrypted grid is used in the gradient mutation area. When the color temperature gradient change amplitude is less than the color temperature gradient change threshold, a uniform grid is used.

[0019] Perform area-aware exposure compensation based on the HSV color space on each compensation grid:

[0020] When the proportion of the highlight reflection area in the compensation grid is greater than the first threshold, a compensation instruction for reducing the brightness gain is generated, and the hue offset is locked;

[0021] When the proportion of the shadow coverage area is greater than the second threshold, a compensation instruction for increasing the brightness gain is generated, and at the same time, a hue compensation instruction for shifting toward the warm color reference is generated, and a hue protection strategy is activated for the compensation grid containing the positioning mark, limiting the hue compensation instruction offset range of the compensation grid to not exceed the preset tolerance range.

[0022] As a further improvement of the present technical solution, the operation of the adaptive correction unit specifically includes:

[0023] Using the brightness compensation coefficients of mesh vertices as control points, a bilinear interpolation algorithm is performed on the brightness compensation coefficients to generate a pixel-level V channel gain mapping table. The angular difference of the hue offsets of adjacent meshes is calculated along the radial path of the hue ring, and spherical linear interpolation is used to achieve a smooth transition of the H channel. The V channel and H channel are both contained in the HSV color space.

[0024] Perform inverse color space conversion on the interpolated HSV parameters. When the target image is used for optical character recognition, it is converted to YUV color space output. When it is used for filling point detection, it is converted to Lab color space output.

[0025] As a further improvement of this technical solution, the quality assessment unit performs:

[0026] Extract all the positioning mark areas in the target image, calculate the HSV mean vector of each positioning mark area, and compare the Euclidean distance between the mean vector and the pre-stored standard vector. When any distance exceeds the threshold, a first-level retake is triggered. If the standard is still not met after the first-level retake, a hierarchical retake mechanism is initiated:

[0027] Level 1 retake: Adjust the exposure time of the mobile terminal to the recommended value;

[0028] Secondary retake: activates the anti-shake mode and extends the exposure time to 200%;

[0029] Three-level retake: Switch to multi-frame synthesis mode and capture three frames of images in short / medium / long exposure sequence.

[0030] As a further improvement of the present technical solution, the multi-frame synthesis mode includes:

[0031] The local contrast of the filled area is extracted from the three frames of images collected three times. Local contrast = (local maximum grayscale value - local minimum grayscale value) / average grayscale value. The frame with the highest local contrast is selected as the output. If the contrast difference of multiple frames is less than 5%, the middle exposure frame is selected.

[0032] As a further improvement of this technical solution, the generation logic of the non-uniform encrypted grid is:

[0033] An initial Cartesian grid is generated along the color temperature gradient direction in the color temperature distribution heat map, and additional grid lines are inserted in the area where the gradient change rate is greater than a set value. Delaunay triangulation is performed on the singular quadrilateral formed after inserting the additional grid lines.

[0034] As a further improvement to this technical solution, the initial H channel mean of the positioning mark grid is detected. If the H channel offset after compensation causes the mean to deviate from the standard value by ±10 degrees, it is forcibly reset to the value before the offset, and the compensation redistribution algorithm is started for the affected adjacent grids.

[0035] As a further improvement of the present technical solution, the convolutional neural network model training adopts an answer sheet image dataset containing diffuse reflection, direct light and mixed light sources, and marks the mirror reflection outline of the highlight area and the penumbra transition boundary of the shadow area. Finally, a multi-task loss function is introduced to synchronously optimize the regional segmentation accuracy and color temperature prediction error. The multi-task loss function is used to measure the difference between the model prediction results and the true label.

[0036] A second object of the present invention is to provide a method for implementing the above-mentioned AI-based intelligent image acquisition system for examination papers, comprising the following steps:

[0037] S1. Capture the answer sheet image through a mobile terminal, use a convolutional neural network to separate high-frequency texture and low-frequency illumination components, identify the edge of the positioning mark as the illumination-invariant area, combine the low-frequency components to generate a probability map of the highlight and shadow areas, and correct the output color temperature distribution heat map through void convolution;

[0038] S2. Dynamically divide the compensation grid based on the color temperature gradient of the heat map. For grids with a highlight ratio greater than a threshold, reduce the brightness gain and lock the hue. For grids with a shadow ratio greater than a threshold, increase the brightness gain and compensate the hue to a warm tone baseline. A hue protection mechanism with a ±10-degree tolerance is enabled for grids containing positioning marks.

[0039] S3. Use bilinear interpolation to generate a pixel-level V channel gain mapping table, and spherical linear interpolation to achieve a smooth transition of the H channel. Convert the corrected image into YUV color space or Lab color space according to the application scenario;

[0040] S4. Extract the HSV mean vector of the target image positioning mark and compare the Euclidean distance with the pre-stored standard value. If it exceeds the threshold, start the graded re-shooting.

[0041] The third object of the present invention is to provide a terminal device for implementing the above-mentioned AI-based intelligent image acquisition system for marking examinations, including: a teacher's PC, a scanner, a camera, an image acquisition workstation, a server, a processor, a memory bar, a storage, a network interface and a GPU.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] The present invention uses a light source analysis unit to identify highlight reflections and shadow areas based on a convolutional neural network, generates a color temperature distribution heat map, and provides an accurate basis for dynamic compensation. The dynamic compensation unit divides the compensation grid according to the heat map, reduces the brightness gain and locks the hue of the highlight area, increases the brightness gain of the shadow area and compensates the hue to the warm color benchmark, and at the same time enables a hue protection mechanism for the grid containing positioning marks, effectively solving the problems of highlight saturation and shadow blur under complex lighting, and improving image quality. The adaptive correction unit realizes a smooth transition of compensation parameters through bilinear interpolation and spherical linear interpolation, and converts it to YUV or Lab color space according to the application scenario to meet the different needs of optical character recognition and filling point detection, reducing the positioning mark offset error. The quality assessment unit extracts the HSV mean vector of the positioning mark and compares it with the standard value, triggering a graded reshoot mechanism, and ensuring that the collected image meets the examination requirements by adjusting the exposure time, activating the anti-shake mode and the multi-frame synthesis strategy, thereby improving the image pass rate and the accuracy of automatic scoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is an overall block diagram of the present invention.

[0045] Figure 2 It is the overall flow chart of the present invention.

[0046] Figure 3 It is the overall structure diagram of the hardware system of the present invention.

[0047] The meaning of each number in the figure is:

[0048] 1. Image acquisition unit; 2. Light source analysis unit; 3. Motion compensation unit; 4. Adaptive correction unit; 5. Quality assessment unit. DETAILED DESCRIPTION

[0049] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] The present invention provides an AI-based intelligent image acquisition system for marking papers. Figure 1 As shown, it includes the following units:

[0051] The image acquisition unit 1 obtains the original image of the answer sheet through the mobile terminal;

[0052] The light source analysis unit 2 identifies the highlight reflection area and shadow coverage area in the original image based on the convolutional neural network and generates a color temperature distribution heat map;

[0053] In order to accurately generate a color temperature distribution heat map reflecting the illumination distribution in the original image, the light source analysis unit 2 needs to process the original image through the steps of multi-scale Gaussian filtering, lightweight convolutional neural network recognition, and hole convolution correction. The specific implementation method of the light source analysis unit 2 performing the following steps to generate the color temperature distribution heat map is as follows:

[0054] Multi-scale Gaussian filtering is performed on the original image to separate high-frequency texture and low-frequency illumination components. Since the original image contains both texture and illumination information, multi-scale Gaussian filtering is required to separate the two. Gaussian kernels with different standard deviations are used to perform convolution operations on the original image to generate multi-scale filtering results. The low-frequency illumination component is obtained by weighted fusion of the filtering results of different scales. The high-frequency texture component is then subtracted from the original image to obtain the text and fill-in dots on the answer sheet. Compared with traditional single-scale Gaussian filtering, this multi-scale filtering method can more accurately separate illumination and texture. The positioning marker edges in the high-frequency texture component are identified through a pre-trained lightweight convolutional neural network model. In order to determine the areas in the image that are not affected by illumination changes, the high-frequency texture component is processed using a pre-trained lightweight convolutional neural network model. The model input is the high-frequency texture component image, and the output is a probability map of the positioning marker edges. The probability map is converted into a binary image through threshold segmentation, and the edge contours are extracted and marked as illumination-invariant areas to ensure the accuracy of the marking of illumination-invariant areas.

[0055] In order to analyze the highlight and shadow areas in the illumination component, the low-frequency illumination component is input into the convolutional neural network structure of the convolutional neural network model for semantic segmentation. The network takes the low-frequency illumination component as input, extracts illumination features through the encoder-decoder structure, and outputs the highlight reflection area and the shadow coverage area. It should be further explained that the convolutional neural network model training adopts the answer card image dataset containing diffuse reflection, direct light and mixed light sources, and annotates the specular reflection outline of the highlight area and the penumbra transition boundary of the shadow area. Finally, a multi-task loss function is introduced to simultaneously optimize the regional segmentation accuracy and color temperature prediction error, so as to accurately distinguish between highlight, shadow and normal illumination areas. The multi-task loss function is used to measure the difference between the model prediction result and the true label.

[0056] In order to optimize the segmentation results of highlight and shadow areas and integrate the position information of illumination-invariant areas, we first perform dilated convolution correction on the highlight reflection area and the shadow coverage area to generate a color temperature distribution heat map. The edges of the highlight and shadow areas are smoothed by dilated convolution kernels with different expansion rates to eliminate segmentation noise and small holes, making the area boundaries more consistent with actual illumination changes. Then, the corrected highlight and shadow areas are fused with the position information of the illumination-invariant area, and a color temperature distribution heat map is generated according to the color temperature characteristics of different areas. The pixel value in the heat map reflects the estimated color temperature value of the area. The bluer the color, the higher the color temperature, and the redder the color, the lower the color temperature. This method can accurately reflect the color temperature differences between different areas in the image and provide a reliable illumination distribution basis for subsequent dynamic compensation.

[0057] The dynamic compensation unit 3 divides the original image into N*N compensation grids according to the color temperature distribution heat map, and calculates the brightness compensation coefficient and hue offset for each compensation grid based on the exposure compensation algorithm based on regional color temperature perception;

[0058] To achieve accurate compensation for different lighting areas, the dynamic compensation unit 3 needs to perform adaptive grid division and regional perception exposure adjustment based on the color temperature distribution heat map. The specific operations of the dynamic compensation unit 3 are as follows:

[0059] In order to make the compensation grid density match the degree of illumination change, the color temperature gradient change amplitude of the color temperature distribution heat map is first calculated. The gradient components of the heat map in the horizontal and vertical directions are calculated using the Sobel operator, and the gradient amplitude of each pixel is synthesized. The mean and standard deviation of the gradient amplitude of the entire image are calculated, and the color temperature gradient change threshold is set to the mean of the gradient amplitude of the entire image + the empirical coefficient. The empirical coefficient is set according to historical experience. The compensation grid density is divided according to the color temperature gradient change amplitude of the color temperature distribution heat map. When the color temperature gradient change amplitude of more than 50% in a certain area is greater than or equal to the color temperature gradient change threshold, it is determined to be a gradient mutation area. A non-uniform mesh is used in the gradient mutation area, and an initial Cartesian mesh is generated along the main direction of the color temperature gradient. Additional grid lines are inserted at the position where the gradient change rate exceeds the threshold, such as the color temperature change per pixel is greater than 5K, so that the grid density increases adaptively with the gradient amplitude. For the singular quadrilateral formed after the grid line is inserted, the Delaunay triangulation algorithm is used to decompose it into a regular triangular mesh to ensure that the color temperature change gradient of each sub-area does not exceed 10K. When the color temperature gradient change amplitude is less than the color temperature gradient change threshold, an equal mesh is used, unified into a 32×32 pixel square grid, to ensure computational efficiency and improve the compensation accuracy in the illumination mutation area.

[0060] Perform area-aware exposure compensation based on the HSV color space on each compensation grid:

[0061] For each compensation grid, the pixel ratio of the highlight reflection area is calculated. The first threshold is set to the mean ratio of the highlight area of ​​the image block to which the grid belongs plus 1.2 times the standard deviation. The highlight ratio distribution of the 10×10 grid block is determined by sliding window statistics. When the ratio of the highlight reflection area in the compensation grid is greater than the first threshold, a compensation instruction to reduce the brightness gain is generated. In the HSV color space, a nonlinear attenuation function is applied to the V channel, that is, the brightness channel. The expression of the nonlinear attenuation function is: ,in is the highlight ratio, is the first threshold, ensuring that the brightness attenuation amplitude is related to the highlight ratio. At the same time, the hue offset of the H channel is locked. The H channel is the hue channel.

[0062] When the shadow coverage area ratio in the compensation grid is greater than the second threshold, a compensation instruction for increasing the brightness gain is generated, wherein the second threshold is set to 1.5 times the average value of the shadow ratio. At the same time, a hue compensation instruction for shifting to the warm tone reference is generated, and the brightness increase and hue compensation are started. A piecewise linear gain function is applied to the V channel. When the shadow ratio is between 50% and 70%, the gain coefficient is 1.2. When it exceeds 70%, the gain coefficient is 1.5 to avoid excessive enhancement of noise. At the same time, a compensation amount for shifting to the warm tone reference is generated in the H channel. The angle of the warm tone reference shift is 60 degrees hue angle, and the shift angle is 1.5. ;in , To prevent the positioning mark from causing recognition errors due to hue shift, the hue protection strategy is activated for the compensation grid containing the positioning mark. First, the initial H channel mean of the positioning mark pixels in the grid is detected. When the H channel offset after compensation causes the mean to deviate from the standard value by more than When the hue compensation value of the grid is reset to the value before the offset, the compensation redistribution algorithm is started for the adjacent grids. The redistribution algorithm distributes the reset compensation value to the surrounding grids according to the distance weight through bilinear interpolation to ensure the smoothness of the overall color adjustment and limit the offset range of the hue compensation instruction of the compensation grid to not exceed the preset tolerance interval, providing high-quality input data for subsequent image correction and recognition.

[0063] The adaptive correction unit 4 uses a bilinear interpolation algorithm to process the brightness compensation coefficients and hue offsets of adjacent grids, and performs color gamut mapping conversion based on the HSV color space on the adjacent grids to output the target image;

[0064] To achieve smooth transitions between compensation grids and adapt to different recognition requirements, the adaptive correction unit 4 needs to perform pixel-level interpolation and color space conversion. The operations of the adaptive correction unit 4 specifically include:

[0065] Since the brightness compensation coefficient of the compensation grid needs to be smoothly transitioned from the vertex to the internal pixel, the brightness compensation coefficient of the grid vertex is used as the control point, and the bilinear interpolation algorithm is performed on the brightness compensation coefficient for pixel level mapping. The grid where the compensation coefficients of its four vertices are The vertex coordinates are , calculate the horizontal interpolation coefficient and , and then vertically interpolate to get Point gain factor , generate the pixel-level V channel gain mapping table. Given the annular characteristics of the hue ring, it is necessary to calculate the angle difference of the hue offset of adjacent grids along the radial path of the hue ring. First, the hue value is normalized to Degrees, hue offset for adjacent grids and , calculate the shortest path angle difference , ensuring that the angle difference is range, and spherical linear interpolation is used to achieve smooth transition of the H channel, avoiding the jump problem of linear interpolation at the 360-degree boundary. The V channel and H channel are both included in the HSV color space. The inverse color space conversion is performed on the interpolated HSV parameters. When the target image is used for optical character recognition, HSV is converted to YUV color space output to separate brightness and chromaticity information. After conversion, the Y channel retains the brightness information of the text, and the UV channel compresses color redundancy. When used for filling point detection, it is converted to Lab color space output to match the uniformity of human eye perception. The conversion steps are to first convert HSV to RGB, and then convert to Lab through the CIEXYZ intermediate space. The adaptive correction process forms a closed loop with the grid parameters output by the dynamic compensation unit. The compensation mutation at the grid boundary is first eliminated through bilinear interpolation and spherical linear interpolation, and then converted to the optimal color space according to application requirements, thereby improving image availability in complex lighting scenes.

[0066] The quality assessment unit 5 extracts pixel data of the positioning mark area in the target image, calculates its corrected color gamut value, and compares it with the pre-stored standard color gamut value to generate an image usability score and trigger a graded retake mechanism. When the lighting condition changes by more than a threshold, the graded retake mechanism automatically switches to a multi-frame synthesis mode, obtains an exposure difference image group through three samplings, and selects the frame with the best contrast of the filled area in the exposure difference image group as the output result.

[0067] To ensure that the captured image meets the examination requirements, the quality assessment unit 5 triggers a graded retake mechanism by analyzing the color characteristics of the positioning mark area. The quality assessment unit 5 performs the following operations:

[0068] Since the color stability of the positioning mark area is a key indicator for evaluating image quality, first, based on the illumination-invariant area marked by the light source analysis unit 2, all positioning mark areas in the target image are extracted through the contour detection algorithm. For each marked area, the three channel values ​​of the HSV color space are read pixel by pixel, namely hue, saturation, and brightness. The H channel mean of each positioning mark area is obtained by accumulating the H value of all pixels and dividing it by the total number of pixels. Then, the mean of the S channel and the V channel are calculated respectively to form the HSV mean vector to ensure the accuracy of the color feature. The Euclidean distance between the HSV mean vector and the pre-stored standard vector is compared. When any distance exceeds the threshold, a first-level retake is triggered. When the first-level retake still does not meet the standard, the graded retake mechanism is started:

[0069] Level 1 retake: Adjusts the exposure time of the mobile terminal to the recommended value. When the Euclidean distance between the HSV mean vector of the positioning marker and the pre-stored standard vector exceeds the threshold, a level 1 retake is triggered. The pre-stored standard vector is derived from the laboratory data collected on the standard answer sheet. The adjustment logic is to calculate the recommended exposure time using the PID algorithm based on the deviation between the current exposure time and the standard vector, and send a command to the camera module of the mobile terminal. The camera module controls the exposure time through the electronic shutter, and the adjustment step size is an integer multiple of 1 / 1000s. After the adjustment, the image is recaptured.

[0070] Level 2 retake: Activates stabilization mode and extends exposure time to 200%. If the distance still exceeds the threshold after the first retake, the second retake is initiated. First, the mobile terminal's stabilization mode is activated. The gyroscope monitors device vibration in real time, driving the micro-electromechanical system actuator to adjust the lens position to compensate for image blur caused by hand shake. At the same time, the exposure time is extended to 200% of the current value, increasing the amount of light entering to improve details in shadow areas. This level of retake uses dual measures to address motion blur and underexposure. The stabilization angle compensation range is ±2 degrees.

[0071] Three-level retake: Switch to multi-frame synthesis mode and capture three frames using short, medium, or long exposure sequences. If the second-level retake still fails to meet the requirements, switch to multi-frame synthesis mode and capture three frames using short, medium, or long exposure parameters. Short exposure (e.g., 1 / 1000s) captures highlight details, medium exposure (1 / 400s) balances overall brightness, and long exposure (1 / 100s) enhances shadow details. During acquisition, the shutter is triggered three times synchronously by hardware, with an interval of less than 50ms to minimize the impact of scene changes. The local contrast of the filled area is extracted from each of the three frames. This is achieved by calculating the difference between the local maximum and minimum grayscale values, and the frame with the highest contrast is selected as the output. If the contrast difference between the three frames is less than 5%, the middle exposure frame is selected to balance the performance of all regions. This quality assessment and retake mechanism form a closed-loop feedback loop. From color feature analysis of positioning marks to three-level progressive capture optimization, the system improves image pass rate in complex lighting environments. It is suitable for high-precision marking scenarios using non-professional cameras such as mobile terminals.

[0072] The present invention obtains the original image through the image acquisition unit 1, the light source analysis unit 2 uses the convolutional neural network to generate the color temperature distribution heat map, the dynamic compensation unit 3 divides the compensation grid and performs exposure compensation accordingly, the adaptive correction unit 4 outputs the target image through bilinear interpolation and color space conversion, and the quality assessment unit 5 compares the positioning mark color gamut value to trigger the graded reshoot mechanism, which solves the problems of insufficient adaptability to complex lighting and poor image quality stability in traditional technologies. Through AI-driven multi-unit collaboration, it realizes the highlight-shadow intelligent compensation, color space adaptive conversion and quality closed-loop control of the answer sheet image, improves the accuracy of filling point recognition and the image pass rate, and meets the efficient marking requirements of large-scale examinations.

[0073] See also Figure 2 As shown, the second object of the present invention is to provide a method for implementing an AI-based intelligent image acquisition system for examination papers, comprising the following steps:

[0074] S1. Capture the answer sheet image through a mobile terminal, use a convolutional neural network to separate high-frequency texture and low-frequency illumination components, identify the edge of the positioning mark as the illumination-invariant area, combine the low-frequency components to generate a probability map of the highlight and shadow areas, and correct the output color temperature distribution heat map through void convolution;

[0075] S2. Dynamically divide the compensation grid based on the color temperature gradient of the heat map. For grids with a highlight ratio greater than a threshold, reduce the brightness gain and lock the hue. For grids with a shadow ratio greater than a threshold, increase the brightness gain and compensate the hue to a warm tone baseline. A hue protection mechanism with a ±10-degree tolerance is enabled for grids containing positioning marks.

[0076] S3. Use bilinear interpolation to generate a pixel-level V channel gain mapping table, and spherical linear interpolation to achieve a smooth transition of the H channel. Convert the corrected image into YUV color space or Lab color space according to the application scenario;

[0077] S4. Extract the HSV mean vector of the target image positioning mark and compare the Euclidean distance with the pre-stored standard value. If it exceeds the threshold, start the graded re-shooting.

[0078] See also Figure 3 As shown, the third object of the present invention is to provide a terminal device for realizing the above-mentioned AI-based intelligent image acquisition system for marking examinations, including: a teacher's PC, a scanner, a camera, an image acquisition workstation, a server, a processor, a memory bar, a storage, a network interface and a GPU.

[0079] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-based intelligent image acquisition system for examination papers, characterized in that: The following units are included: The image acquisition unit (1) obtains the original image of the answer sheet through the mobile terminal; The light source analysis unit (2) identifies the highlight reflection area and shadow coverage area in the original image based on the convolutional neural network and generates a color temperature distribution heat map; The dynamic compensation unit (3) divides the original image into N*N compensation grids according to the color temperature distribution heat map, and calculates the brightness compensation coefficient and hue offset for each compensation grid based on the exposure compensation algorithm based on regional color temperature perception; The adaptive correction unit (4) uses a bilinear interpolation algorithm to process the brightness compensation coefficient and hue offset of adjacent grids, and performs color gamut mapping conversion based on the HSV color space on the adjacent grids to output the target image; The operation of the adaptive correction unit (4) specifically includes: Using the brightness compensation coefficients of mesh vertices as control points, a bilinear interpolation algorithm is performed on the brightness compensation coefficients to generate a pixel-level V channel gain mapping table. The angular difference of the hue offsets of adjacent meshes is calculated along the radial path of the hue ring, and spherical linear interpolation is used to achieve a smooth transition of the H channel. The V channel and H channel are both contained in the HSV color space. Perform inverse color space conversion on the interpolated HSV parameters. When the target image is used for optical character recognition, it is converted to the YUV color space output. When it is used for filling point detection, it is converted to the Lab color space output. The quality assessment unit (5) extracts pixel data of the positioning mark area in the target image, calculates its corrected color gamut value, and compares and calculates it with the pre-stored standard color gamut value to generate an image usability score and trigger a graded retake mechanism. When the illumination condition changes beyond a threshold, the graded retake mechanism automatically switches to a multi-frame synthesis mode, obtains an exposure difference image group through three samplings, and selects the frame with the best contrast of the filled area in the exposure difference image group as the output result.

2. The AI-based intelligent image acquisition system for examination paper according to claim 1, characterized in that: The light source analysis unit (2) performs the following steps to generate a color temperature distribution heat map: Multi-scale Gaussian filtering is performed on the original image to separate high-frequency texture and low-frequency illumination components. The pre-trained lightweight convolutional neural network model is used to identify the positioning mark edges in the high-frequency texture component and mark them as illumination-invariant areas. The low-frequency illumination component is input into the convolutional neural network structure of the convolutional neural network model, and the highlight reflection area and the shadow coverage area are output; The high-light reflection area and the shadow coverage area are corrected by void convolution, and the position information of the illumination-invariant area is fused to generate a color temperature distribution heat map.

3. The AI-based intelligent image acquisition system for examination paper according to claim 1, characterized in that: The specific operation of the dynamic compensation unit (3) is as follows: The compensation grid density is divided according to the color temperature gradient change amplitude of the color temperature distribution heat map. When the color temperature gradient change amplitude is greater than or equal to the color temperature gradient change threshold, a non-uniform encrypted grid is used in the gradient mutation area. When the color temperature gradient change amplitude is less than the color temperature gradient change threshold, a uniform grid is used. Perform area-aware exposure compensation based on the HSV color space on each compensation grid: When the proportion of the highlight reflection area in the compensation grid is greater than the first threshold, a compensation instruction for reducing the brightness gain is generated, and the hue offset is locked; When the proportion of the shadow coverage area is greater than the second threshold, a compensation instruction for increasing the brightness gain is generated, and at the same time, a hue compensation instruction for shifting toward the warm color reference is generated, and a hue protection strategy is activated for the compensation grid containing the positioning mark, limiting the hue compensation instruction offset range of the compensation grid to not exceed the preset tolerance range.

4. The AI-based intelligent image acquisition system for examination paper according to claim 1, characterized in that: The quality assessment unit (5) performs: Extract all the positioning mark areas in the target image, calculate the HSV mean vector of each positioning mark area, and compare the Euclidean distance between the mean vector and the pre-stored standard vector. When any distance exceeds the threshold, a first-level retake is triggered. If the standard is still not met after the first-level retake, a hierarchical retake mechanism is initiated: Level 1 retake: Adjust the exposure time of the mobile terminal to the recommended value; Secondary retake: activates the anti-shake mode and extends the exposure time to 200%; Three-level retake: Switch to multi-frame synthesis mode and capture three frames of images in short / medium / long exposure sequence.

5. The AI-based intelligent image acquisition system for examination paper according to claim 4, characterized in that: The multi-frame synthesis mode includes: The local contrast of the filled area is extracted from the three frames of images collected three times. Local contrast = (local maximum grayscale value - local minimum grayscale value) / average grayscale value. The frame with the highest local contrast is selected as the output. If the contrast difference of multiple frames is less than 5%, the middle exposure frame is selected.

6. The AI-based intelligent image acquisition system for examination paper marking according to claim 3, characterized in that: The generation logic of the non-uniform encrypted grid is: An initial Cartesian grid is generated along the color temperature gradient direction in the color temperature distribution heat map, and additional grid lines are inserted in the area where the gradient change rate is greater than a set value. Delaunay triangulation is performed on the singular quadrilateral formed after inserting the additional grid lines.

7. The AI-based intelligent image acquisition system for examination paper marking according to claim 1, characterized in that: Detect the initial H channel mean of the positioning mark grid. If the H channel offset after compensation causes the mean to deviate from the standard value by ±10 degrees, it is forced to reset to the value before the offset, and the compensation redistribution algorithm is initiated for the affected adjacent grids.

8. A method for implementing an AI-based intelligent image acquisition system for examination paper images according to any one of claims 1 to 7, characterized in that: The following steps are involved: S1. Capture the answer sheet image through a mobile terminal, use a convolutional neural network to separate high-frequency texture and low-frequency illumination components, identify the edge of the positioning mark as the illumination-invariant area, combine the low-frequency components to generate a probability map of the highlight and shadow areas, and correct the output color temperature distribution heat map through void convolution; S2. Dynamically divide the compensation grid based on the color temperature gradient of the heat map. For grids with a highlight ratio greater than a threshold, reduce the brightness gain and lock the hue. For grids with a shadow ratio greater than a threshold, increase the brightness gain and compensate the hue to a warm tone baseline. A hue protection mechanism with a ±10-degree tolerance is enabled for grids containing positioning marks. S3. Use bilinear interpolation to generate a pixel-level V channel gain mapping table, and spherical linear interpolation to achieve a smooth transition of the H channel. Convert the corrected image into YUV color space or Lab color space according to the application scenario; S4. Extract the HSV mean vector of the target image positioning mark and compare the Euclidean distance with the pre-stored standard value. If it exceeds the threshold, start the graded re-shooting.

9. A terminal device for implementing the AI-based intelligent image acquisition system for examination paper images according to any one of claims 1 to 7, comprising: Teacher PC, scanner, camera, image acquisition workstation, server, processor, memory bar, storage, network interface and GPU.

Images