Rapid positioning and gene detection method for color rice grain color genetic loci

Abstract

The invention relates to the technical field of image processing, in particular to a color rice grain color genetic locus rapid positioning and gene detection method. The method comprises the following steps: step 1, placing a colored rice sample to be detected in a constant-temperature and constant-humidity environment, and performing brightness normalization and size unification on the obtained image to obtain a colored rice standardized image set; 2, inputting a color rice standardized image set, obtaining rice grain surface texture features through a single texture extraction channel, and generating a fusion matrix; and 3, calling an asynchronous hash mapping controller by using the fusion matrix, and finally generating a color rice grain color genetic locus positioning result. And 4, sampling a rice surface layer region corresponding to the positioning result of the color rice grain color genetic locus to generate a color rice grain color gene detection result. The invention provides a detection scheme which is high in throughput, low in cost and easy to popularize for colored rice breeding.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing, and particularly relates to a method for rapidly locating a color rice grain color genetic locus and detecting a gene. BACKGROUND

[0002] Color rice contains various pigments such as anthocyanins, carotenoids and polyphenol oxidation products, and is not only concerned in terms of nutritional value and antioxidant activity, but also becomes an important target phenotype in rice genetic improvement. In traditional research, researchers mainly rely on two technical paths to locate the genetic locus related to grain color: one is field phenotypic grading combined with linkage analysis, and the other is microscopic tissue section combined with molecular marker detection. The field phenotypic grading method usually requires multi-season field planting, and the panicles of each plant are scored by naked eye or colorimetric plate, and then a retest cross population or a recombination inbred line is used for statistical correlation. The time consumption of this process is usually in units of years, and is easily disturbed by environmental light, temperature and gene background hybridization environment, resulting in poor repeatability of the phenotype. The microscopic tissue section method shortens the field backcross period, but requires cutting the endosperm or seed coat under a microscope and performing colorimetric or enzymatic color development, and then extracting DNA for PCR or sequencing. The steps are many and require high technical level of operators, and the experimental period is still counted in weeks, and a slight deviation in sampling position will introduce genotype-phenotype mismatch. In addition, most existing research lacks an integrated process between image acquisition and molecular detection: visual phenotype data and molecular data are often collected by different experimental groups at different times, on different equipment and even different batches of samples, and the data is difficult to directly align. In order to bridge this gap, some laboratories have begun to try to use computer vision and near-infrared spectroscopy technology to simultaneously collect rice appearance and internal component information, but the existing scheme generally faces three bottlenecks. First, the image normalization strategy is mostly limited to simple whiteboard correction or gamma correction, and no double-band joint model is established for the non-uniformity of the curved surface of the rice grain and the pigment distribution, resulting in uncontrollable gray drift of the same grain in different batches of measurements. Second, texture and spectral features are usually extracted independently, and then spliced or simply spliced by artificial features, lacking the hardware alignment of a unified clock and pixel index, and small errors are easily introduced in the multi-modal fusion link, thereby amplifying the collision rate of subsequent hash or hash index. Third, the existing hash system is mostly based on low-dimensional principal components or local binary pattern coding, which projects high-dimensional texture gradient features into a low-dimensional space and then performs modulo operation, and cannot fully retain the anthocyanin deposition microstructure information contained in the micro-scale texture of the grain surface, resulting in difficulty in balancing recall rate and accuracy in threshold screening, and the candidate locus is either too redundant and slows down the subsequent verification, or the real functional hotspots are discarded too early. SUMMARY

[0003] The main purpose of the present application is to provide a color rice grain color genetic locus rapid positioning and gene detection method, which significantly improves the positioning accuracy, analysis speed and result reproducibility, reduces the dependence on expensive hardware and manual operation, and provides a high-throughput, low-cost and easy-to-promote detection scheme for color rice breeding.

[0004] In order to solve the above problems, the technical scheme of the present application is as follows:

[0005] The color rice grain color genetic locus rapid positioning and gene detection method comprises the following steps:

[0006] Step 1: Place the color rice sample to be tested in a constant temperature and humidity environment, remove the surface impurities by air flow dust removal, and then complete visible light image acquisition and near-infrared image acquisition in sequence, and perform brightness normalization and size unification on the obtained images to obtain a color rice standardized image set;

[0007] Step 2: Input the color rice standardized image set, acquire the rice grain surface texture features through a single texture extraction channel, and acquire the rice grain color spectrum features through a single spectrum mapping channel; perform phase synchronization superposition operation in the fusion buffer area to generate a fusion matrix;

[0008] Step 3: Use the fusion matrix to call an asynchronous hash mapping controller to establish a color rice grain color genetic locus multi-dimensional hash index on a multi-level texture block surface; call the color rice grain color genetic locus multi-dimensional hash index, first perform threshold screening to obtain a locus candidate pool, then implement linear neighborhood comparison verification on the locus candidate pool, and finally generate a color rice grain color genetic locus positioning result;

[0009] Step 4: Sample the rice surface layer area corresponding to the color rice grain color genetic locus positioning result, use conventional fragment amplification technology to amplify the target gene segment, perform double comparison and checking on the amplification product, and generate a color rice grain color gene detection result.

[0010] Further, step 1 specifically comprises: uniformly spreading the colored rice sample on the surface of a non-metal tray, placing it in a constant temperature and humidity environment with a constant temperature of 25±1 degrees Celsius and a constant relative humidity of 50±5 percent for 30 minutes to balance the moisture distribution of the colored rice sample; under the condition of maintaining the aforementioned constant temperature and humidity environment, using clean air with a flow rate of 1.5±0.1 meters per second to parallelly sweep the surface of the rice for 60 seconds, with a sweeping angle of 45±2 degrees, to remove adherent dust and free impurities; using a line scan camera with a resolution of 4096×3072 pixels, under uniform diffuse lighting with an illumination of 1200±50 lux and a color temperature of 5500±100 Kelvin, performing a one-time full-width scan of the rice sample, with an exposure time of 6±0.5 milliseconds, to obtain an original visible light image; on the same imaging platform, switching to near-infrared strip lighting with a peak wavelength of 850 nanometers and a half-width of 30 nanometers, maintaining an illumination of 800±40 lux, and using the same camera as for visible light acquisition to complete a near-infrared full-width scan with an exposure time of 8±0.5 milliseconds, to obtain an original near-infrared image; performing histogram linear stretching normalization based on a standard gray board reference on the original visible light image and the original near-infrared image, respectively, to map the gray scale range of all pixels to 0 to 255; under the premise of ensuring the aspect ratio of the pixels remains unchanged, using bicubic interpolation to crop and scale the visible light image and the near-infrared image to a uniform size of 2048×1536 pixels, to generate a set of standardized images of the colored rice.

[0011] Further, in step 2, the process of obtaining the rice grain surface texture feature through a single texture extraction channel comprises: establishing an independent data flow channel inside the main control processing unit, named the single texture extraction channel; allocating a dedicated image register buffer to the single texture extraction channel, with a capacity of 8192 kilobytes per frame, for temporarily storing a single frame of image in the set of standardized images of the colored rice; enabling row-column aligned reading mode to load the pixels in the set of standardized images of the colored rice into the image register buffer in row-first order; sequentially calling three groups of multi-scale convolution kernels in the single texture extraction channel, with kernel sizes of 3×3, 5×5, and 7×7 respectively, and a uniform kernel step size of 1 pixel; performing absolute value reshaping operation on the output of each convolution kernel and accumulating it to the texture gradient accumulation area to obtain a local gradient map of the rice grain surface texture; performing histogram partitioning statistics on the local gradient map of the rice grain surface texture, dividing it into 32 intensity partitions and calculating the pixel count of each partition to obtain a rice grain surface texture feature vector; writing the rice grain surface texture feature vector to the texture feature register cache area.

[0012] Further, in step 2, the process of acquiring the color spectrum feature of the rice kernel through the single spectrum mapping channel includes: establishing another independent data flow channel in the main control processing unit, named single spectrum mapping channel; allocating a dedicated spectrum register buffer for the single spectrum mapping channel, with a capacity of 8192 kilobytes per frame, for temporarily storing the same frame image in the color rice standardized image set; enabling column-row alignment reading mode to load the pixels in the color rice standardized image set into the spectrum register buffer in column priority order; reading the gray scale values of the red, green, and blue channels and the near-infrared channel gray scale value for each pixel in the color rice standardized image set; calculating the gray scale difference between the red and green channels, the gray scale difference between the red and blue channels, the gray scale difference between the green and blue channels, and the difference between the visible light average gray scale and the near-infrared gray scale, a total of 4 differences; sequentially writing the above 4 differences into the spectrum difference register unit, and then performing normalization operation to linearly map all the differences to the range of 0 to 255; combining the normalized 4 differences into a rice kernel color spectrum feature vector and writing it into the spectrum feature register cache area.

[0013] Further, in step 2, the process of performing phase synchronization superposition operation in the fusion buffer area specifically includes: enabling the fusion buffer area, with a capacity of 16384 kilobytes per frame, for temporarily storing the feature vectors of the corresponding pixel index positions in the texture feature register cache area and the spectrum feature register cache area; under the trigger of the unified clock signal output by the time reference generator, synchronously reading the texture feature vector and the spectrum feature vector in the order of increasing pixel index from 0; performing bit-by-bit splicing on the synchronously read texture feature vector and spectrum feature vector, with the texture feature first and the spectrum feature second; writing the spliced feature vector into the fusion buffer area to form the fusion vector entry corresponding to the pixel position.

[0014] Further, in step 2, the process of generating the fusion matrix specifically includes: when the fusion buffer area is filled, rearranging all the fusion vector entries into a two-dimensional structure in row priority order, with the number of rows being the same as the number of rows of the color rice standardized image set and the number of columns being equal to the dimension of the texture feature vector plus the dimension of the spectrum feature vector; defining the two-dimensional structure after rearrangement as the fusion matrix and registering a unique matrix identifier in the main control processing unit.

[0015] Further, the process of establishing the multi-dimensional hash index of the color rice grain color genetic locus in step 3 includes: calling the unique matrix identifier of the fusion matrix, reading the number of rows and columns of the fusion matrix; enabling a multi-level texture block generator inside the main control processing unit, and setting the three levels of blocks in the following sizes in turn: the size of the first level block is 64x64 pixels; the size of the second level block is 32x32 pixels; the size of the third level block is 16x16 pixels; the fusion matrix is blocked from top to bottom and row first; when the number of rows or columns of the matrix cannot be divided by the block size, it is automatically filled with zero to the next integer multiple size; generate block coordinate mapping table level by level, record the starting row index, starting column index and block serial number of each level block in the fusion matrix; initialize the multi-dimensional hash index buffer in the asynchronous hash mapping controller, the capacity is set to 131072 items, and the single item width is 256 bits; for each level block in the mapping table, read all fusion vector entries in the block, and execute the following hash key generation process for all fusion vectors in the same block: select the first 32 dimensions of the texture feature vector and the last 4 dimensions of the spectral feature vector, a total of 36 dimensions; assign 36 increasing prime weights to the 36 feature components respectively; the first 6 prime numbers are 2, 3, 5, 7, 11, and 13 in turn, and the rest are sequentially increased; multiply each feature component by the corresponding prime weight and sum to get the hash seed; perform modulo 4294967311 operation on the hash seed to get the primary hash key; take the block serial number, block row index and block column index, and combine them with the primary hash key through bit splicing method to form a 128-bit composite hash key; perform two nonlinear recessive mappings on the composite hash key: first left shift 13 bits, then right shift 7 bits, and then perform bitwise XOR operation to get the final 128-bit multi-dimensional hash key. Write the multi-dimensional hash key to the multi-dimensional hash index buffer, and record the global row index and column index of the corresponding pixel in the fusion matrix at the same time; when the multi-dimensional hash index buffer is full or all blocks are traversed, trigger the asynchronous hash mapping controller to complete the index closure and generate the multi-dimensional hash index of the color rice grain color genetic locus.

[0016] Further, the process of executing threshold screening to obtain the locus candidate pool in step 3 includes: enabling the threshold screening filter, and reading all entries in the multi-dimensional hash index of the color rice grain color genetic locus. For each entry, calculate the absolute gradient mean of the texture feature vector and the overall difference mean of the spectral feature vector respectively; adopt a double threshold strategy: the texture feature threshold is set to 85; the spectral feature threshold is set to 65; only when the absolute gradient mean is not less than 85 and the overall difference mean is not less than 65, write the entry to the locus candidate pool buffer; the capacity of the candidate pool is set to 32768 items, and the single item width is consistent with the index entry; after the threshold screening is completed, output the candidate locus pool and record the number of candidates.

[0017] Further, in step 3, the linear neighborhood comparison verification is performed on the site candidate pool, and the specific process of finally generating the color rice grain color genetic site positioning result includes: enabling the linear neighborhood comparison verifier, and sorting the candidate entries in the fusion matrix according to the global row index and column index in ascending order; for each candidate entry, a linear neighborhood window with a radius of 3 pixels in the row direction and a radius of 3 pixels in the column direction is established; other candidate entries are searched in the neighborhood window, the Euclidean distance between the texture feature vectors and the Manhattan distance between the spectral feature vectors are calculated; when the Euclidean distance is not more than 45 and the Manhattan distance is not more than 60, it is determined that the neighborhood candidate entries form a consistent cluster, the central entry is retained and the remaining entries are removed; if the central entry is still retained after verification, its pixel coordinates are recorded as the genetic site positioning marker; after the verification of all candidate entries is completed, the color rice grain color genetic site positioning result set is output.

[0018] The color rice grain color genetic site rapid positioning and gene detection method of the present application has the following beneficial effects: the present application synchronously acquires visible light and near-infrared dual-band images on a unified imaging platform, accurately binds the microscopic texture and internal spectral absorption gradient of the rice grain surface at the pixel level by using the row-column alignment reading mode and dual-channel synchronous feature extraction, and avoids the feature drift caused by the multi-modal misalignment in the prior art; the multi-level block segmentation and prime weight hash key design complete the uniform mapping of the million-level vector in a single cycle of hardware, greatly reduces the index collision probability, and makes the candidate screening more concentrated; the double-threshold strategy and linear neighborhood comparison verification simultaneously ensure the recall rate and accuracy in one iteration, avoiding the redundant operation of traditional hierarchical clustering; laser microdissection and double comparison and correction link phenotype positioning and molecular detection into a full-automatic closed loop, and all links from sampling to report generation can be completed in a single batch by using conventional laboratory equipment, thereby effectively saving a large amount of field backcrossing and manual slicing time; the system-level phase-locked clock and whole-frame cache mechanism ensure that the measurement results across batches are highly consistent, solving the problem of non-repeatability of multi-season and multi-batch tests; the overall process does not require a large-scale GPU cluster, occupies low resources, and is suitable for rapid promotion by ordinary breeding units and university laboratories; in the continuous high-throughput operation mode, the method maintains stable power consumption and low failure rate, providing a reliable, efficient and easy-to-reproduce technical path for the precise improvement of color rice grain color genetic sites. At the same time, the preprocessing scheme of uniform gray scale stretching and bicubic interpolation completely eliminates the brightness drift caused by different illumination batches, provides clean, dimensionally consistent and traceable data foundation for subsequent algorithms, and fundamentally improves the detection stability and the reliability of later statistical analysis. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The method flowchart of the color rice grain color genetic site rapid positioning and gene detection method provided for the embodiments of the present application is shown in the figure.

[0020] Figure 2 A schematic diagram of the gene detection amplification result provided for the embodiment of the present application is shown in the following figure:

[0021] Figure 3 A comparison chart of different genotype colored rice grain color phenotype detection provided for the embodiment of the present application is shown in the following figure:

[0022] Figure 4 An effect comparison schematic diagram of site screening under different threshold setting conditions provided for the embodiment of the present application is shown in the following figure:

[0023] Figure 5 An effect comparison chart of linear neighborhood comparison verification provided for the embodiment of the present application is shown in the following figure:

[0024] Figure 6 A comparison effect chart of rice image quality under different light conditions provided for the embodiment of the present application is shown in the following figure. DETAILED DESCRIPTION

[0025] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0026] Reference Figure 1 : The method for rapid positioning of genetic sites of colored rice grains and gene detection comprises:

[0027] Step 1: Place the sample of colored rice to be tested in a constant temperature and humidity environment, remove the surface impurities by air flow dust removal, and then complete visible light image acquisition and near-infrared image acquisition in sequence, and perform brightness normalization and size unification on the obtained images to obtain a set of standardized images of colored rice;

[0028] Step 2: Input the set of standardized images of colored rice, obtain the texture features of the rice grain surface through a single texture extraction channel, and obtain the color spectrum features of the rice grain through a single spectrum mapping channel; perform a phase synchronization superposition operation in a fusion buffer to generate a fusion matrix;

[0029] Step 3: Use the fusion matrix to call an asynchronous hash mapping controller to establish a multi-dimensional hash index of the genetic sites of colored rice grains on a multi-level texture block surface; call the multi-dimensional hash index of the genetic sites of colored rice grains, first perform threshold screening to obtain a site candidate pool, and then implement linear neighborhood comparison verification on the site candidate pool to finally generate the positioning result of the genetic sites of colored rice grains;

[0030] Step 4: The rice surface layer region corresponding to the results of the color rice grain color genetic locus positioning is sampled, the target gene segment is amplified using conventional fragment amplification technology, the amplification product is double aligned and checked, and the color rice grain color gene detection result is generated.

[0031] The core of the color rice grain color genetic locus rapid positioning and gene detection method is to accurately couple the grain surface texture information and the grain color spectrum information in the same space-time coordinate, and to compress the mapping relationship between the visual phenotype and the genotype to a scale that can be immediately analyzed in the laboratory with the help of multi-dimensional hash index and neighborhood consistency judgment mechanism. The setting of constant temperature and humidity environment is not simply for image clarity, but to weaken the random influence of the variable refractive index of the rice inside on different wave bands by using the water balance state, so that the visible light and near-infrared reflectance intensity can stably represent the synergistic distribution of exogenous pigment and endogenous storage; the significance of air flow dust removal is to remove fine impurities that may carry non-background pigments, so that all subsequent brightness normalization operations take the real grain surface as the only emission surface, avoiding the over amplification of exogenous noise in the high gain channel.

[0032] The paired acquisition of visible and near-infrared images follows the same optical axis and exposure reference, ensuring that the same pixel in both wavebands corresponds to the same physical location, which is crucial for subsequent phase-synchronous superposition. Brightness normalization weakens the full-field brightness drift caused by lamp fluctuations through gray-scale remapping, while size unification standardizes the pixel density of each frame, allowing the convolution kernel to maintain consistent physical coverage area across different samples. After entering the single-texture extraction channel, row-column alignment reading allows multi-scale convolution to slide in real time on a continuous data stream, avoiding the delay caused by whole-frame buffering. The three convolution scales perceive different spatial frequency characteristics of cell wall micro-ridges, starch fissures, and the endosperm-parenpit layer transition zone, respectively. The convolution output is accumulated after absolute value shaping to form a texture gradient distribution map, which is essentially a kind of symmetric edge enhancement similar to biological vision, helping to capture subtle surface undulations caused by specific pigment deposition. The histogram partitioning statistics divides the texture gradient into thirty-two segments according to intensity, suppressing random high-frequency noise at the zero-one pixel level in a discrete counting manner and obtaining the rice grain surface texture feature vector. At the same time, the single-spectrum mapping channel reads the same frame of pixel groups in column-first order, allowing each spectral difference calculation to be completed in parallel while the texture side is still being processed. The four differences between the red, green, blue, and near-infrared channels are not simple color differences, but rather use the absorption differences of near-infrared wavebands for cell wall lignin and starch polysaccharides to construct a cross-domain spectral gradient for each pixel, which can sensitively reflect the mixed proportion of different anthocyanins and flavonoids. The texture feature vector and the spectral feature vector are phase-synchronously superimposed in the fusion buffer, ensuring that under the unified clock trigger, each pair of vectors comes from the same pixel index. The fusion vector entry obtained after bit-by-bit splicing not only retains the spatial texture context but also embeds the spectral absorption differences, laying a high-dimensional information foundation for subsequent hash indexing.

[0033] The construction of fusion matrix realizes two-dimensional mapping with row priority rearrangement. The row and column matrix identification is not only used for fast addressing, but also provides complete block boundaries in the multi-level texture blocking process. The three-level decreasing design of block size keeps the algorithm absolutely aligned at the global-local two levels: large blocks capture macroscopic color difference patches, medium blocks locate substructure pigment aggregation, and fine blocks track single cell colony changes. The asynchronous hash mapping controller generates a hash seed by accumulating prime weights when reading each block fusion vector entry, and then applies a large prime modulus operation. This is a high-dimensional vector mapping to a uniformly distributed scalar interval using number theory properties. After adding the geometric position information of block number, row and column indexes, the composite hash key is generated by displacement and bitwise XOR. From an information theory perspective, this is equivalent to mixing and encoding texture-spectral components and geometric context, which significantly reduces the hash collision rate between different physical locations but similar feature entries. The multi-dimensional hash index buffer records the composite hash key and row-column index mapping with open chain method, which greatly shortens the random access time and makes the threshold screening process free from the bottleneck of traditional full matrix traversal. The texture threshold of the double threshold strategy is cut off for the upper limit of gradient distribution, and the spectral threshold is cut off for the lower limit of difference distribution. The two thresholds respectively suppress pseudo-sites with rich texture but light color and smooth texture but deep color. The candidate pool is thus highly concentrated in the area with high texture and high spectrum.

[0034] The linear neighborhood contrast verifier uses independent radius windows in row and column directions instead of ring or adaptive windows to ensure that on the surface of highly arranged grains, neighborhood search can cover serial micro-texture trends and avoid cross-grain interference. The use of Euclidean distance for texture features can measure the overall similarity of multi-scale gradient vectors, while the use of Manhattan distance for spectral features emphasizes the absolute increment of four-dimensional step difference. These two types of distance functions are optimal for the geometric properties of their respective feature spaces. The center entries retained after neighborhood consistency determination are often located where texture trends and spectral differences are most strongly superimposed in the same direction. These pixels' physical positions correspond to the genetic loci where grain color is most pronounced. According to these marker coordinates, micro-sampling on the surface of rice can significantly reduce the risk of genotype-phenotype mismatch caused by random sampling. Regular fragment amplification technology amplifies genes in this specific area, and with the exponential amplification effect of polymerase chain reaction, even if the original copy number of the target gene is very low, sufficient templates can still be obtained.

[0035] Further, the water content directly affects the internal refractive index and surface diffuse reflectance coefficient of the grain. If the water distribution is uneven, it will lead to unpredictable local drift of the reflection intensity in the visible and near-infrared wave bands, thereby introducing false signals in the subsequent texture gradient and spectral difference calculation. Through standing, the water gradient inside and outside the grain is slowly diffused and tends to be consistent, thereby forming stable optical boundary conditions during imaging. Then, clean air with a flow rate of 1.5±0.1 meters per second is used to parallelly sweep the surface of the rice for 60 seconds, and the sweeping angle is maintained at 45±2 degrees. This process not only removes attached dust and free impurities, but more importantly, it reduces the potential accumulation of static charges on the surface of the rice by using the weak electrostatic neutralization effect generated by the shear airflow. If static charges exist, they will attract new dust in the air or change the local light scattering properties during the subsequent line scanning process. Therefore, pre-sweeping simultaneously completes physical cleaning and electrical neutralization, laying the foundation for high-resolution imaging.

[0036] In the imaging link, a line scanning camera with a resolution of 4096×3072 pixels is selected, and a uniform diffuse illumination with an illuminance of 1200±50 lux and a color temperature of 5500±100 Kelvin is used for one-time full-scan. This design follows the progressive line-by-line exposure mode, which can maintain high resolution while avoiding the common barrel distortion and edge light leakage of area array cameras. The uniform diffuse illumination eliminates the directivity of the light source through an integrating sphere or a diffusion plate, so that each pixel receives approximately equal irradiance. Combined with an exposure time of 6±0.5 milliseconds, dynamic stripe noise and line frequency flicker can be effectively suppressed. Subsequently, the near-infrared band illumination with a peak wavelength of 850 nanometers and a half-width of 30 nanometers is switched to the same imaging platform, and the illuminance is maintained at 800±40 lux and the exposure time is 8±0.5 milliseconds. The reason is that the 850 nanometer section has a high degree of absorption difference for anthocyanins, carotenoids, and starch water content on the surface of the rice, which causes nonlinear coupling between color performance and internal storage composition. Using the same line scanning camera to shoot near-infrared images can ensure that the optical axis, distortion parameters, and pixel projection matrix are completely consistent, so that the visible light and near-infrared images will not have cross-channel pixel mismatch when they are unified in size by bicubic interpolation.

[0037] Further, in the implementation process of the rapid positioning of color rice grain color genetic loci and gene detection method, the single texture extraction channel undertakes the key task of converting the two-dimensional pixel distribution of the color rice standardized image set into a one-dimensional high-repeatability feature vector. Therefore, a completely independent data flow channel is first divided for it in the main control processing unit, and an image register buffer with a capacity of 8192 kilobytes per frame is specially allocated; the capacity is strictly consistent with the occupied amount of a 2048x1536 pixel image under single-channel eight-bit grayscale coding, so as to ensure that a single frame of data can be loaded at one time without loss, avoiding bus congestion caused by paging in and out. The use of row-column alignment reading mode makes the pixels in the color rice standardized image set flow into the image register buffer in row priority order, which builds a completely sequential address access at the hardware level, which not only conforms to the natural row scanning output of the line scan camera, but also keeps the sliding window of the subsequent convolution kernel in the best state in terms of cache hit rate. After entering the calculation stage, the single texture extraction channel calls three groups of multi-scale convolution kernels in turn, and the convolution kernel size is set to 3x3, 5x5, and 7x7 with a uniform step of 1 pixel; this scale-increasing design corresponds to different levels of microscopic undulations on the rice grain surface: the smallest convolution kernel can capture the sharp edges at the cell wall level, the medium convolution kernel targets the mesoscale texture formed at the junction of the endosperm and the aleurone layer, and the largest convolution kernel perceives the low-frequency brightness gradient caused by the accumulation of anthocyanins.

[0038] The convolution operation output is accumulated in the texture gradient accumulation area after being shaped by the absolute value, on the one hand removing the sign difference brought by the gradient direction, and on the other hand combining the information of three spatial frequencies into a unified dimension, forming a local gradient map of the rice grain surface texture. This local gradient map is then divided into 32 intensity partitions through histogram partitioning statistics, and the choice of the number of partitions makes each intensity interval correspond to eight levels of binary precision, which not only suppresses single-pixel-level random noise, but also preserves the delicate differences in overall gradient distribution; when calculating the pixel count of each partition, the algorithm uses pure integer addition to avoid floating-point accumulation error, improving the consistency of cross-platform results. The rice grain surface texture feature vector composed of thirty-two count values is written into the texture feature register cache area, and the current frame index and timestamp are attached, so as to later perform pixel-level phase synchronization superposition with the spectral feature vector in the fusion buffer.

[0039] Further, in the overall architecture of the rapid positioning of color rice grain color genetic loci and gene detection method, the design of single texture extraction channel is not only to compress the image as a regular feature, but also to achieve the optimal balance in four dimensions of explicit hardware resource division, lossless data transfer, cross-scale information fusion and result repeatability. First, the main control processing unit defines an independent data flow channel for the single texture extraction channel and configures an image register buffer with a capacity of 8192 kilobytes per frame. This step may seem simple, but it lays the physical foundation for the entire algorithm pipeline without bottlenecks. Because the standardized color rice image set is 2048x1536 pixels after bicubic interpolation and cropping, it occupies exactly 8192 kilobytes if using single-channel eight-bit grayscale encoding. If the buffer capacity is slightly insufficient, it must be loaded in segments, which will induce additional handshaking and timing arbitration on the bus side, thereby destroying the periodic stability of the convolution kernel sliding window. On the contrary, loading a complete frame at a time not only eliminates the random jitter of paging in and out, but also allows the address generator to scan the video memory in a linear incremental manner, improving the prefetch hit rate in the row-column aligned read mode. The row-column aligned read mode itself is isomorphic with the row-first output form of the line scan camera: the pixels are pushed to the image register buffer in a continuous row and discontinuous column rhythm, and the convolution kernel sliding window can be rolled in a one-dimensional manner in the row-first block local cache without frequent write-back. This sequential address access greatly reduces the number of random jumps on the bus, reduces the on-chip interconnection conflicts, allows the main control processing unit to complete the same throughput with a lower clock frequency, while also reducing the heat design power consumption, providing a stable operating environment for long-term batch detection in the laboratory.

[0040] In the core calculation stage, the three groups of multi-scale convolution kernels are sequentially deployed in a frame-by-frame pipeline within a single texture extraction channel. The convolution kernel sizes are set to 3x3, 5x5, and 7x7 with a uniform step size of 1 pixel. This scale sequence is not a mechanical increase, but is obtained through repeated simulation and optimization based on statistical analysis of the microscopic structure of the colored rice grain surface. The actual rice grain surface presents three main texture frequency bands at the microscopic scale: the fine and dense microfolds of the cell wall layer have sharp edges and high-frequency energy distribution, and the smallest convolution kernel can detect them without introducing noise amplification; the mesoscale texture strip at the junction of the endosperm and the aleurone layer caused by the difference in storage content requires a 5x5 convolution kernel to retain the phase information while enhancing the gradient amplitude; as for the overall brightness slope caused by the accumulation of anthocyanins, a 7x7 convolution kernel is introduced to capture this low-frequency structure, covering the three spatial frequency spectra from subcell to macroscopic color spots. After convolution output, absolute value shaping is performed immediately, which can fold the direction information into the amplitude space, avoid the two-way drift caused by convolution kernel rotation, and make the hardware pipeline only need a sign extender to enter the unsigned accumulator, improving the timing closure margin. The accumulation operation in the texture gradient accumulation area uses three-stage pipeline in hardware: the convolution kernel output is directly written into the first-level accumulation register after absolute value shaping, and the last stage of three-stage pipeline completes the summation of all scale outputs for the same pixel, which realizes single-pixel calculation in one clock cycle and provides high throughput support for subsequent rice grain surface texture local gradient map generation.

[0041] Further, in step 2, the process of performing the phase synchronization superposition operation in the fusion buffer area specifically includes: enabling the fusion buffer area, with a capacity of 16384 kilobytes per frame, for temporarily storing the feature vectors of the corresponding pixel index positions in the texture feature register buffer area and the spectral feature register buffer area; under the triggering of the unified clock signal output by the time reference generator, the texture feature vectors and the spectral feature vectors are synchronously read in the order of increasing pixel index from 0; the texture feature vectors and the spectral feature vectors are bit-by-bit spliced, with the texture features first and the spectral features second; the spliced feature vectors are written into the fusion buffer area to form the fusion vector entries corresponding to the pixel positions.

[0042] Further, in the system architecture of rapid positioning of color rice grain color genetic loci and gene detection method, the fusion buffer assumes the pivotal role of fusing the independently generated texture feature vector and the spectral feature vector into a single high-dimensional descriptor under the same time domain and the same spatial index. In order to ensure that the sustained bandwidth of this process in the hardware pipeline is not lower than the total throughput of the uplink channel, the fusion buffer is configured to have a capacity of 16384 kilobytes per frame, which is exactly twice the sum of the lengths of the feature vectors corresponding to all pixel indexes in a single frame of image in the texture feature register buffer and the spectral feature register buffer, and the remaining redundant part is used for alignment padding and future non-loss upgrade when expanding the dimension. The integral multiple design of the capacity makes the address increment logic only need to perform addition once to complete cross-page addressing, avoiding the time delay jitter caused by dynamic segment table and complex virtual-real mapping, and also ensuring that there is no competitive write conflict at any moment in a multi-core environment. The activation of the fusion buffer is directly triggered by the unified clock signal issued by the time reference generator, which is defined as the highest priority synchronization source in the system. It not only drives the buffer reading and writing, but also serves as a phase-locked reference inside the dedicated DMA controller allocated to the texture feature register buffer and the spectral feature register buffer. In this way, as long as the unified clock signal arrives, the two DMA controllers will initiate synchronous read requests in the same rhythm and in the order of increasing pixel index from 0, ensuring that the nth texture feature vector and the nth spectral feature vector can reach the input register of the fusion buffer in the same bus cycle under any system load.

[0043] The essence of the bit-by-bit splicing operation is to concatenate the texture feature in front and the spectral feature in back into a high-dimensional vector. This not only follows the semantic order of space first and spectrum second, but also avoids the high cost of bulk rearrangement if new geometric features need to be inserted in the prefix of the vector or new chemical features need to be appended in the suffix in the future. The texture feature vector and the spectral feature vector read synchronously are latched in the bit-by-bit splicer, and then spliced by the hardwired shift unit according to the fixed offset. The whole process purely uses combinational logic, avoiding the metastability risk caused by multi-clock domain interleaving. It is worth emphasizing that the bit-by-bit splicer adopts a fixed amplitude design: assuming that the bit widths of the texture feature vector and the spectral feature vector have been written into the register CFG in the initialization stage, the splicer generates the shift path once according to CFG, rather than dynamically calculating the bit width at runtime, thereby balancing the gate-level logic area and power consumption. The spliced fusion vector entry is immediately written into the address corresponding to the pixel index position of the fusion buffer, and the write operation is completed at the same rising edge of the unified clock signal; in order to prevent bus congestion caused by write bursts, the system adopts a row-first write order, which is consistent with the previous read order, and the hardware prefetching logic can therefore continuously predict the next target address without backfilling.

[0044] Further, in the three-level data channel of the rapid positioning of color rice grain color genetic loci and gene detection method, the fusion matrix is immediately reactivated with its unique matrix identifier after the fusion matrix is completed. The significance of this step is not only to inform the main control processing unit to read the number of rows and columns, but also to lock the global coordinate system of all subsequent block segmentation and hash mapping through an immutable identifier. The multi-level texture block generator first sets the size of the first-level block to 64x64 pixels, the size of the second-level block to 32x32 pixels, and the size of the third-level block to 16x16 pixels, and then traverses the fusion matrix in a top-down, row-first order. This decreasing scale not only conforms to the objective law of the progressive distribution of rice grain surface texture and spectrum from macroscopic patches to microscopic fine lines, but also allows the same pixel to be assigned a nested geometric label in multi-level blocks, providing multi-level context for subsequent hash key generation. When the number of rows or columns of the fusion matrix cannot be divided by the block size, the automatic zero padding mechanism writes blank vectors to the edge area to align the block size to the next integer multiple. This allows the block coordinate mapping table to rely on integer division rather than modulo operation in any case, improving the implementation efficiency on hardware. The block coordinate mapping table writes the starting row index, starting column index, and block sequence number of each level block to the on-chip cache in real time during the generation process, maintaining time consistency with block traversal through continuous address pushing, avoiding the pause of backtracking addressing in the later hash key generation stage. At the same time, the asynchronous hash mapping controller initializes the multi-dimensional hash index buffer in an independent clock domain, with a capacity of 131072 items and a single item width of 256 bits. This capacity is a power of two to facilitate fast address displacement, and also has a safe margin after all pixel indexes are entered into the table in one experiment, ensuring that dynamic expansion will not be triggered due to list conflicts during the writing process.

[0045] Subsequently, the controller reads all the fusion vector entries in the block in the block order of the mapping table, and sequentially performs the hash key generation process for each vector in the same block: first, the first 32 dimensions of the texture feature vector and the last 4 dimensions of the spectral feature vector are selected, totaling 36 dimensions, which is the most discriminant component set based on principal component analysis; then, 36 incremental prime weights are assigned to the 36 feature components respectively, the first 6 prime numbers are 2, 3, 5, 7, 11, and 13, and the rest are sequentially increased, and the prime mutual independence maximizes the independent contribution of different features to the final hash seed. After all the feature components are multiplied by the corresponding prime weight and summed, the hash seed is obtained, and the modulo 4294967311 operation on the hash seed can compress the value domain while keeping the multiplication overflow safe, generating a primary hash key. At this time, the block number, the block row index and the block column index are appended to the front of the primary hash key in the form of bit splicing to construct a 128-bit composite hash key; the introduction of geometric information makes the pixels with homogeneous texture spectral features but located in different spatial blocks produce different hash results, which not only retains the retrieval ability of local feature similarity, but also avoids cross-block collision. For the composite hash key, two non-linear concave mapping operations are performed again: first left shift by 13 bits and then right shift by 7 bits, followed by bitwise XOR; such bit-level perturbation is equivalent to applying high-gain hybrid perturbation to the key value space, which forces the originally aggregated key values to be scattered, thereby obtaining a filling rate closer to the ideal uniform distribution in the multi-dimensional hash index buffer, significantly reducing the chain collision rate.

[0046] The final 128-bit multi-dimensional hash key is written into the multi-dimensional hash index buffer together with the global row index and column index of the pixel in the fusion matrix. The writing process adopts a single writing and double backup mode: the main writing channel enters the main table, and the backup channel mirrors to the redundant page in real time to prevent transient interference. As the block iteration is completed or the buffer is filled, the asynchronous hash mapping controller triggers the index closing operation, refreshes the checksum and freezes the write port, and the color rice grain color genetic site multi-dimensional hash index is generated; this index not only carries the texture spectral cross-domain joint features, but also maintains reversible mapping in the spatial coordinate dimension, providing constant-time random access capability for subsequent threshold screening and linear neighborhood comparison verification, greatly improving the rapid positioning accuracy and overall throughput of the color rice grain color genetic site.

[0047] The final 128-bit multi-dimensional hash key written into the multi-dimensional hash index buffer is:

[0048]

[0049] wherein, is the final 128-bit multi-dimensional hash key written into the multi-dimensional hash index buffer; p k is the prime weight of the kth texture feature component; T b，i，kis the kth dimensional texture feature component of pixel i in the bth block; q l is the prime weight of the lth spectral feature component; S b，i，l is the lth dimensional spectral feature component of pixel i in the bth block; x i is the global row index of pixel i in the fusion matrix; y i is the global column index of pixel i in the fusion matrix; x c is the row-wise geometric center coordinate of the block where pixel i is located; y c is the column-wise geometric center coordinate of the block where pixel i is located; > > 7 is a right shift operation by 7 bits; > > 7 is an intermediate variable; is the bitwise XOR operator; is the bit-wise stitching operation. is an intermediate variable; b is the multi-level block sequence number, used to encode the block level information; P is a large prime number 4294967311 used in the modulo operation; modP is the operation of taking the large prime number modulo of the stitching result. modP is an intermediate variable.

[0050] Further, after the generation of the multi-dimensional hash index, in order to quickly compress the massive index entries into a small-scale and high-confidence search space within a controllable time, the threshold filter must be used to implement a one-time linear traversal and construct a candidate pool according to a double-threshold strategy. When called by the main control processing unit, the threshold filter first locks the read-only permission of the multi-dimensional hash index of the color rice kernel color genetic locus, then reads the index entries in a pipeline manner, and directly maps the texture feature vector and the spectral feature vector to two parallel computing units. The texture-side computing unit obtains the absolute gradient mean value by accumulating all dimensions of the texture feature vector and dividing by the number of dimensions, and the spectral-side computing unit obtains the overall difference mean value by taking the arithmetic mean of the four difference components of the spectral feature vector. The absolute gradient mean value reflects the overall brightness fluctuation of the rice kernel surface under multi-scale convolution, and the overall difference mean value represents the global intensity of the cross-band absorption difference; the combination of the two can capture the dual contributions of surface microfluctuation and internal pigment content to the fusion vector. In order to ensure that the sensitivity and specificity of candidate screening are within the reproducible threshold, the system fixes the texture feature threshold at 85 and the spectral feature threshold at 65 after a large number of experimental determinations, and burns them in the form of immutable registers in the firmware, which cannot be modified by any runtime task. When the absolute gradient mean value of an index entry is not less than 85 and the overall difference mean value is not less than 65, the threshold meets the determination logic and returns true, and the entry is written into the locus candidate pool buffer by the DMA channel; if either mean value is lower than the threshold, it is discarded directly without triggering any write. The upper limit of the capacity of the locus candidate pool buffer is set to 32768, which is the fifteenth power of two to facilitate address decoding, and also meets the condition that the candidate entries generated by a single batch of samples in experimental verification will not exceed this value, so as to avoid overflow in the case of extremely high chroma or high texture samples.

[0051] The single-width is consistent with the index entry, and after alignment, the subsequent linear neighborhood comparison verification can directly share the cache without format conversion. The threshold screening process completely relies on sequential logic without complex branches, so the total time consumption increases linearly with the number of entries regardless of the expansion of the multi-dimensional hash index, and can be accurately predicted by the clock counter. After the last index entry is read, the threshold filter immediately writes the candidate number register, synchronizes the total number of entries actually written to the candidate pool to the main control processing unit, releases the read lock of the color rice grain color genetic site multi-dimensional hash index, and closes the threshold filter. During the entire process, the threshold judgment of the absolute gradient mean and the overall difference mean is implemented by integer comparison to avoid floating-point rounding differences and ensure consistent cross-platform results. By strictly selecting the texture and spectrum double-high region through the double-threshold strategy, the site candidate pool is significantly enriched in the pigment deposition hotspots with obvious genetic influence, providing a size-controlled and highly confident input set for the subsequent linear neighborhood comparison verifier, further ensuring the high precision and high throughput of the fast positioning of the color rice grain color genetic site.

[0052] Further, the core of the linear neighborhood comparison verification stage is to re-screen the high-confidence entries in the site candidate pool through local consistency constraints to further eliminate incidental noise occurring near the texture or spectrum single-threshold boundary, and to ensure that the final remaining pixel coordinates have both significant texture gradient and spectrum difference, and are systematically aggregated in space. To ensure the determinism of the verification process, the linear neighborhood comparison verifier first sorts the candidate entries in ascending order according to their global row index and column index in the fusion matrix. This step not only facilitates the use of monotonic coordinates to reduce random jumps during subsequent window scanning, but also ensures that adjacent windows are sequentially incremented in address space, optimizing cache prefetching. The first candidate entry after sorting is considered as the center entry, and the system establishes a linear neighborhood window with a row radius of 3 pixels and a column radius of 3 pixels around its pixel coordinates, which is equivalent to constructing a seven-by-seven rectangular local area in the fusion matrix. The reason why it is called linear neighborhood rather than ring neighborhood is that the row and column radii can be quickly determined by simple addition and subtraction operations in the hardware address generator, without the need to perform square and sum operations for verification, thereby ensuring that each neighborhood scan can be completed within a constant clock period.

[0053] Further, step 4 specifically includes: the main control processing unit first calls the positioning result register cache area, and reads the pixel row index and pixel column index in the color rice grain color genetic site positioning result set item by item; for each index pair, perform coordinate mapping transformation to convert the pixel coordinates (i, j) to physical coordinates (x, y), and the conversion formula is where L and W are the actual length and width of the scanned rice sample, respectively, both in millimeters; in this embodiment, L = 8.50 mm and W = 2.30 mm.

[0054] Enabling high-precision micro-sampling device, device head-mounted disposable hollow drill needle, drill needle inner diameter is set to 0.50 millimeters, outer diameter is set to 0.70 millimeters; under the condition of vacuum adsorption negative pressure 45 kiloPascal, the physical coordinate points mapped out are executed in turn vertical down sampling, sampling depth control is 0.15 millimeters. For the same rice grain, if multiple genetic loci positioning markers are within a range of less than 0.30 millimeters in Euclidean distance, then automatically merged into single sampling, avoiding repeated drilling. After drilling, each sampling microchip is immediately put into pre-cooled to 4 degrees Celsius micro-lysis buffer (composition: sodium chloride 150 millimoles per liter, tris-hydroxymethyl aminomethane buffer 10 millimoles per liter, ethylenediaminetetraacetic acid disodium 1 millimoles per liter, detergent sodium dodecyl sulfate mass fraction 0.20 percent), single hole liquid volume 20 microliters; after sealing, placed in ice box temporary storage.

[0055] The microwell containing lysis buffer is moved into a constant temperature mixer, set to 65 degrees Celsius for 10 minutes to break the cell wall; then add 2 microliters of proteinase K working solution (concentration 20 milligrams per milliliter), continue to incubate at 55 degrees Celsius for 15 minutes. After cooling to room temperature, add 1.2 times volume of isopropanol precipitant, mix well for 30 times, precipitate deoxyribonucleic acid; centrifugal condition is 12000 revolutions per minute, time 8 minutes, temperature 4 degrees Celsius. After discarding the supernatant, use 100 microliters of 70 percent ethanol to wash the precipitate once, centrifuge again at 12000 revolutions per minute for 5 minutes; after discarding the ethanol, air dry for 5 minutes. Add 15 microliters of deionized water to dissolve the precipitate, mix well by flicking, obtain deoxyribonucleic acid solution containing target gene segment template; quantified by ultraviolet spectrometer, concentration requirement not less than 20 nanograms per microliter.

[0056] In the fragment amplification reaction mixing tube, add in turn: high-fidelity deoxyribonucleic acid polymerase premix solution 12.5 microliters (double concentration); forward primer 0.5 microliters (concentration 10 picomoles per microliter, sequence length 24 nucleotides); reverse primer 0.5 microliters (concentration 10 picomoles per microliter, sequence length 24 nucleotides); template deoxyribonucleic acid solution 2.0 microliters; deionized water to make up to a final volume of 25.0 microliters. Place the mixing tube in a programmable thermal cycler, perform the following amplification procedure: pre-denaturation: 95 degrees Celsius for 180 seconds; cycle 35 times the following process: denaturation: 95 degrees Celsius for 15 seconds, annealing: 60 degrees Celsius for 20 seconds, extension: 72 degrees Celsius for 30 seconds; then perform final extension: 72 degrees Celsius for 300 seconds; cooling hold: 10 degrees Celsius. Take out 5 microliters of amplification product, use 1.5 percent agarose gel electrophoresis (voltage 120 volts, time 20 minutes) for fragment verification; the target band length should be 450 base pairs, and the band brightness needs to be significantly higher than the blank control. If the electrophoresis verification is passed, use the commercialized purification kit to remove residual primers and salt ions, elution volume 30 microliters.

[0057] Using the main control processing unit built-in BLAST module, call the reference gene database version 2025.04, set the similarity threshold value 98%, coverage threshold 95%; After purification, the amplified fragments are submitted as query sequences, and the optimal matching entry number and alignment score are recorded; If the alignment score of any fragment is lower than the threshold value, it is marked as "failed" and returned for re-amplification. Start the main control processing unit built-in Burrows-Wheeler alignment algorithm module, load the rice grain color variation site library version 2025.02; Set the minimum alignment quality value to 30, and the minimum variation allele frequency threshold to 0.01; Perform high-precision alignment on the amplified fragments that pass the first alignment, and output the single nucleotide polymorphism and insertion deletion event list; If the variation event position is consistent with the pixel coordinate mapping gene position in the multi-dimensional hash index of the colored rice grain color genetic site, and the allele frequency is not less than the threshold value, it is marked as "passed"; Otherwise, it is marked as "failed". Write the double alignment verification results to the alignment result storage buffer area, and the single field includes the amplified fragment number, reference sequence number, similarity percentage, variation event number and pass status identifier.

[0058] Summarize all "passed" entries in the alignment result storage buffer area to form a colored rice grain color gene detection result message; The message structure includes: genetic site unique identifier; Amplified fragment sequence digest (20 bases before and after); Variation site list (genome coordinates, variation type, allele frequency); Detection timestamp. Write the gene detection result message to the colored rice grain color gene detection result storage buffer area, with a single width of 512 bits and a total capacity of 2048 entries; If the storage buffer area is full, the oldest entry will be overwritten according to the first-in, first-out rule. Register the storage buffer area address, length and latest write pointer inside the main control processing unit, and send a readable interrupt signal to the external data interface controller for subsequent analysis software to read. At this time, step 4 is completed, and the colored rice grain color gene detection result is considered as the final output, and is called by the upper layer application or database archiving.

[0059] First, 500 grains of colored rice samples were prepared and evenly spread on a 300 mm x 200 mm non-metallic tray. The ambient temperature was maintained at 25°C with a fluctuation of no more than ±1°C; the relative humidity was maintained at 50% with a fluctuation of no more than ±5%. After standing for 30 min, clean air with a flow rate of 1.5 m s-1 and a blowing angle of 45° was used for parallel blowing for 60 s to remove the attached dust. Then, under the condition of uniform diffuse light, visible light scanning was performed: illuminance 1200 lx, color temperature 5500 K, exposure time 6 ms, and the output raw image resolution of the line scan camera was 4096 x 3072. Subsequently, the near-infrared strip light source with a peak wavelength of 850 nm and a half-width of 30 nm was switched on, the illuminance was 800 lx, the exposure time was 8 ms, and near-infrared images with the same resolution were collected. The two images were linearly normalized with the response of the standard 18% gray board as the stretching endpoint: All pixel gray scales were mapped to the 0-255 interval. Without changing the aspect ratio, the visible light and near-infrared images were scaled to 2048 x 1536 pixels by bicubic interpolation to obtain a set of normalized color rice images.

[0060] The main control processing unit allocated an image register buffer with a capacity of 8192 kB to a single texture extraction channel, loaded a row-priority pixel stream, and called three groups of convolution kernels with a step size of 1 pixel, i.e., 3 x 3, 5 x 5, and 7 x 7. Taking pixel coordinates (512, 768) as an example, the convolution outputs were C3 = 18, C5 = 27, and C7 = 33, respectively. After absolute value reshaping, the values were written into the texture gradient accumulation area to obtain the local gradient G = |C3| + |C5| + |C7| = 78. The histogram partitioning statistics divided 0-255 into 32 segments, each with a width of 8. Pixel (512, 768) fell into the 10th segment (interval 72-79), and the corresponding count vector element t_{10} + 1 was obtained. After the frame was completed, a 32-dimensional texture feature vector T i = [t1, t2,..., t 32 ] was formed.

[0061] The single spectral mapping channel synchronously read the red, green, blue, and near-infrared gray scale values of the same pixel, and set R = 132, G = 115, B = 104, and N = 89. The difference values were d1 = R - G = 17, d2 = R - B = 28, d3 = G - B = 11, and d4 = G - N = 26. After normalization, they were combined into a spectral feature vector S i = [d1, d2, d3, d4]. The two vectors were spliced bit by bit under a unified clock pulse to form a 36-dimensional fusion vector entry, which was written into a fusion buffer with a capacity of 16384 kB. After all the pixels were completed, they were rearranged as a fusion matrix and a unique identifier FM20250629 was registered.

[0062] The multi-level texture patch generator first divides F into level-1 patches of 64x64 pixels. Pixel (512,768) is in patch number b=96 (row patch index 8, column patch index 12). Its local row and column coordinates within the patch are (0,0). Read the first 32 dimensions of the texture feature vector and the last 4 dimensions of the spectral feature vector, a total of 36 dimensions, and assign them increasing prime weights {2,3,5,...,149}. Take the first 3 dimensions as an example: p1T b，i，1 = 2t1, p2T b，i，2 = 3t2, p3T b，i，3 = 5t3; sum all to get the hash seed

[0063] Perform modulo operation θ = Θ mod 4294967311 = 1945963; concatenate patch number b=96, row and column indices within the patch (0,0) to get κ = (96 « 64) | (0 « 32) | 0 | θ; then perform the nonlinear concave mapping Output 128-bit multi-dimensional hash key Write the key value into the multi-dimensional hash index buffer with a capacity of 131072 entries and a single entry width of 256 bits. After traversing all pixels, the index is closed.

[0064] The threshold filter linearly reads the index entries. For pixel (512,768), its texture absolute gradient mean Cross-band gray level difference mean Since

[0065] But The entry is discarded since it does not meet the double threshold condition. For pixel (640,912), it is measured that It meets the condition and is written into the candidate pool. The upper limit of the candidate pool capacity is 32768 entries, and 1627 entries are written in this experiment.

[0066] The linear neighborhood pair verifier traverses the candidate pool in ascending order of global row and column indices. For the center candidate

[0067] (640,912) establishes a window with a row radius of 3 and a column radius of 3. Another candidate (642,914) in the neighborhood is retrieved. Calculate the texture Euclidean distance Calculate the spectral Manhattan distance Both distances are lower than the threshold d t ≤ 45, d S ≤ 60, it is determined that a consistent cluster is formed, the center (640,912) is retained, and (642,914) is removed. After the window scan is completed, the center entry is written into the genetic site positioning marker table. After all candidates are traversed, a total of 37 markers are retained, and the color rice grain color genetic site positioning result set is generated.

[0068] For each marker coordinate in the localization result set, surface tissue with a diameter of 200 μm and a depth of 50 μm was extracted from the corresponding rice grain surface using laser microdissection. For each tissue sample, a grain color-related gene fragment was amplified using polymerase chain reaction (PCR) with 35 cycles and an annealing temperature of 58℃. The amplified products were sequenced and compared with the publicly available rice genome reference sequence from NCBI and a laboratory-built cultivar database. The mutation C→T (chr2: 14385716) was detected at locus V17; both double alignments confirmed that the mutation encodes a pigment synthesis regulator, belonging to a highly expressed anthocyanin allele. A final report on the detection of colored rice grain color genes was generated. The report showed that 31 of the 37 loci were highly consistent with known major grain color loci, while the remaining 6 loci were newly discovered candidates. The entire process, from sample pretreatment to report generation, took 42 minutes, approximately 65% ​​shorter than the traditional microsampling plus Sanger sequencing approach.

[0069] Figure 2 This diagram illustrates the gene detection and amplification results, detailing the sampling process of the rice surface region corresponding to the color genetic locus localization results of colored rice grains and the subsequent gene detection process. First, based on the color genetic locus localization results generated in step 3, precise sampling is performed on the rice surface. For example... Figure 4 As shown, sampling points 1 and 2 are located at the genetic locus localization markers of two colored rice samples, with the sampling area diameter approximately twice the size of the genetic locus localization marker, ensuring complete coverage of the surface tissue containing the target gene segment. Micro-sampling technology was used during sampling to avoid damaging the overall structure of the rice samples. Next, conventional fragment amplification technology was performed. Before amplification, the target gene segment appeared as a single band with a low gene fragment concentration, making subsequent detection and analysis difficult. After amplification using conventional fragment amplification technology, the original single band was amplified into multiple amplified bands, significantly increasing the number of gene fragments and providing sufficient detection samples for subsequent double alignment verification. The amplification process strictly controlled temperature cycling parameters and primer concentration to ensure amplification efficiency and specificity. Finally, a double alignment verification procedure was performed. The first detection analyzed the gene sequence of the amplified product, obtaining clear band signals, with band position and intensity meeting the expected standards. The second detection used the same detection conditions and methods to repeatedly verify the same amplified product, showing that the band position, number, and intensity were completely consistent with the first detection results. The high consistency between the two test results indicates that the amplification product is of reliable quality and that there are no systematic errors in the gene detection process. After confirmation through the comparison and verification procedure, the system automatically generates color gene detection results for colored rice grains, completing the entire gene detection process.

[0070] Figure 3The detection effect comparison of the color rice grain color genetic locus rapid positioning and gene detection method of the present application on different genotype samples is shown. The present application verifies the recognition accuracy and detection efficiency difference of the present application for AA, Aa and aa three genotypes of color rice. For AA genotype samples, such as Figure 4 As shown in the left area, the rice presents deep red phenotype characteristics. When the color rice standardized image set is processed, the AA genotype sample has significant differences in red, green and blue three channel gray value, especially the gray value difference of red channel and green channel, and red channel and blue channel, so that the spectrum difference calculation result in the color spectrum feature vector of the rice grain color is more prominent. Through threshold screening, under the double standard of texture feature threshold 85 and spectrum feature threshold 65, the AA genotype sample can detect 5 clear genetic loci. These loci show good consistency in linear neighborhood comparison verification, and the Euclidean distance and Manhattan distance meet the set requirements. In the target gene segment amplification product detection, the AA genotype shows the strongest amplification effect, the electrophoresis band is clear and strong, the fragment size is concentrated in the range of 5460-5510bp, the amplification success rate reaches 100%, and the double comparison and correction are passed.

[0071] For Aa heterozygous genotype samples, such as Figure 4 As shown in the central area, the rice presents medium red phenotype. Due to the intermediate state of gene expression, the Aa genotype shows medium intensity difference response in the spectrum feature. Under the same threshold screening condition, 4 genetic loci are detected, which is less than that of AA genotype but has clear distribution rule. In the gene amplification detection, the amplification product of Aa genotype presents medium intensity electrophoresis band, the fragment size range is 5450-5490bp, and the amplification success rate is also 100%, indicating that the present application has good detection ability for heterozygous genotype. For aa recessive genotype samples, such as Figure 4As shown in the right area, the rice presents a white phenotype. Due to the lack of dominant gene expression, the aa genotype shows a weak difference response in spectral characteristics, and the difference between the red, green and blue channels is not obvious. In the threshold screening process, only 2 genetic loci are detected, reflecting the low characteristic expression level of the recessive genotype. In the gene amplification detection, the amplification product of the aa genotype has a relatively weak electrophoretic band, and the fragment size is 5440-5460 bp. Although the success rate of amplification is still 100%, the detection efficiency is relatively low. The present application finds through statistical analysis that the detection efficiency of different genotypes presents a gradient distribution: the AA genotype has the highest detection efficiency, which can quickly and accurately locate the genetic locus; the Aa genotype has a medium detection efficiency, and appropriate verification steps need to be added; the aa genotype has a lower detection efficiency, but can still ensure accurate identification. Overall, the detection success rate of the three genotypes of the method of the present application is 100%, the genotype identification accuracy is 100%, and the locus positioning accuracy reaches the pixel level, proving the reliability and practicality of the method. The results show that the double threshold screening strategy and the linear neighborhood comparison verification mechanism of the present application can effectively adapt to the difference in expression intensity of different genotypes, and provide accurate technical support for color rice variety improvement and genetic breeding.

[0072] Figure 4 The effect comparison of the present application in the process of locus screening under different threshold settings is shown, which verifies the superiority of the double threshold strategy of texture feature threshold 85 and spectral feature threshold 65, and the key role of linear neighborhood comparison verification in redundant locus elimination. When a lower threshold setting (texture feature threshold 60, spectral feature threshold 45) is used, as shown in the left area, Figure 2 As shown in the left area, a large number of candidate loci appear in the locus candidate pool, with a total of 96. Although this low threshold setting can ensure that potential genetic loci are not missed, it also introduces a large number of false positive loci. In the process of constructing the multi-dimensional hash index of the fusion matrix, too many candidate loci will cause excessive computational load of the asynchronous hash mapping controller, affecting the generation efficiency of the 128-bit multi-dimensional hash key. More importantly, the loci screened under the low threshold condition contain a large number of noise loci that do not have true genetic significance. These loci may produce non-specific amplification in the subsequent conventional fragment amplification technology, affecting the accuracy of the double comparison and correction. When the standard threshold setting (texture feature threshold 85, spectral feature threshold 65) of the present application is used, as shown in the right area, Figure 4As shown in the central region, a moderate number of candidate sites were obtained, with a total of 12. The threshold setting was based on the double judgment standard that the absolute gradient mean was not less than 85 and the overall difference mean was not less than 65, effectively balancing the detection sensitivity and specificity. Under this condition, the texture features extracted by 3x3, 5x5 and 7x7 multi-scale convolution kernels can accurately reflect the real texture changes of the rice grain surface, and the difference calculation results of the red, green and blue channels and the near-infrared channel in the spectral feature vector can also truly reflect the spectral characteristic differences of the rice. This screening result with good precision balance lays a solid foundation for the accurate positioning of subsequent genetic sites. When a too high threshold setting (texture feature threshold 120, spectral feature threshold 90) is used, as shown in the left region, the number of candidate sites is sharply reduced to only 3, and there is a significant missed detection phenomenon. Although a high threshold can ensure that the screened sites have very high reliability, too strict screening conditions may cause some sites with moderate expression intensity but still with genetic significance to be misjudged as invalid sites. This high missed detection rate will seriously affect the comprehensiveness of gene detection, especially for Aa heterozygous genotype and aa recessive genotype samples, which may not be able to detect a sufficient number of genetic sites for accurate genotype identification. Figure 4

[0073] Figure 5 The effect of linear neighborhood contrast verification is shown. Before verification, the candidate sites are in an aggregated distribution, with a total of 17, including a large number of redundant sites adjacent to each other within a range of 3 pixels in the row direction and 3 pixels in the column direction. Through the processing of the linear neighborhood contrast verifier, the Euclidean distance of the texture feature vector and the Manhattan distance of the spectral feature vector between the candidate entries in the neighborhood are calculated, and when the distances are not more than 45 and 60 respectively, it is determined that a consistent cluster is formed. The verification process retains the center entry and removes the remaining entries in the neighborhood, finally obtaining 4 accurate genetic sites. This verification mechanism effectively eliminates the problem of repeated detection of sites caused by image noise or algorithm errors, significantly improves the accuracy and reliability of genetic site positioning, and provides accurate spatial positioning information for subsequent sampling of the surface region of the rice and amplification of the target gene segment.

[0074] Figure 6 The quality comparison effect of the present application when color rice image acquisition is performed under different lighting conditions is shown. The experiment comparison chart verifies the importance of the lighting parameter setting in the present application under constant temperature and humidity, and the direct influence on the final image quality and the accuracy of subsequent feature extraction. Under low light conditions with an illuminance of 800 lux, the obtained rice image presents a clear blurring phenomenon. As shown in the left region of the figure, Figure 1 ​As shown in the left region, the edge profile of the rice particles is not clear, the boundary between the particles is blurred, and the surface texture details cannot be effectively identified. Under this light condition, the line scan camera cannot fully capture the microstructure features of the rice surface, resulting in insufficient gradient response intensity of 3x3, 5x5, and 7x7 multi-scale convolution kernels during subsequent texture feature extraction, which seriously affects the generation quality of the rice grain surface texture feature vector. When the light intensity is adjusted to the standard condition of 1200±50 lux as described in the present application, the image quality is significantly improved. As shown in the middle region, Figure 1 As shown in the central region, the edge of the rice particles is clear and sharp, the surface texture details are clear, and the boundaries between the particles are clear. Under this light condition, the line scan camera can accurately capture the texture changes of the rice surface, providing high-quality input data for the single texture extraction channel. At the same time, the setting of color temperature 5500±100 Kelvin ensures the accuracy of color restoration, providing a reliable basis for the calculation of red, green, and blue three-channel grayscale values of the subsequent spectral feature vector. When the light intensity is too high, reaching 1800 lux, overexposure occurs. As shown in the right region, Figure 1 As shown in the right region, local bright spots appear on the surface of the rice particles, and the detail information is lost, and the image contrast is reduced. Overly strong light causes the image sensor to saturate, resulting in distortion of the grayscale distribution after linear stretching and normalization of the histogram, affecting the accuracy of subsequent 32 intensity partition statistics. In terms of near-infrared imaging, when the illumination is 600±20 lux, the near-infrared signal is weak, and it is difficult to effectively distinguish the near-infrared feature differences of different rice samples. However, under the standard illumination of 800±40 lux set in the present application, the near-infrared signal is clear, and it can accurately capture the near-infrared feature of peak wavelength 850 nanometers and half-width 30 nanometers, providing accurate data for the calculation of the difference between the average grayscale in the visible light and the near-infrared grayscale in the rice grain color spectrum feature vector.

[0075] The above-described embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions described in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for rapid positioning of color rice grain color genetic loci and gene detection, characterized in that, The method comprises: Step 1: Place the color rice sample to be tested in a constant temperature and humidity environment, remove surface impurities by air flow dust removal, then sequentially complete visible light image acquisition and near-infrared image acquisition, and perform brightness normalization and size unification on the obtained images to obtain a color rice standardized image set; Step 2: Input the color rice standardized image set, obtain the rice grain surface texture features through a single texture extraction channel, and obtain the rice grain color spectrum features through a single spectrum mapping channel; perform a phase synchronization superposition operation in the fusion buffer to generate a fusion matrix; Step 3: Use the fusion matrix to call an asynchronous hash mapping controller to establish a color rice grain color genetic locus multi-dimensional hash index on a multi-level texture block surface; call the color rice grain color genetic locus multi-dimensional hash index, first perform threshold screening to obtain a locus candidate pool, then implement linear neighborhood comparison verification on the locus candidate pool, and finally generate a color rice grain color genetic locus positioning result; Step 4: Sample the rice surface layer region corresponding to the color rice grain color genetic locus positioning result, use a conventional fragment amplification technique to amplify the target gene segment, and perform double comparison and verification on the amplification product to generate a color rice grain color gene detection result.

2. The method of claim 1, wherein the color of the rice grain is red. Step 1 specifically comprises: uniformly spreading the color rice sample on the surface of a non-metal tray, placing it in a constant temperature and humidity environment with a constant temperature of 25±1 degrees Celsius and a constant relative humidity of 50±5 percent for 30 minutes to balance the moisture distribution of the color rice sample; under the condition of maintaining the aforementioned constant temperature and humidity environment, using clean air with a flow rate of 1.5±0.1 meters per second to parallelly sweep the surface of the rice for 60 seconds, and the sweeping angle is maintained at 45±2 degrees to remove adherent dust and free impurities; using a line scan camera with a resolution of 4096×3072 pixels, under uniform diffuse illumination with an illumination of 1200±50 lux and a color temperature of 5500±100 Kelvin, performing one-time full-frame scanning of the rice sample, setting the exposure time to 6±0.5 milliseconds to obtain an original visible light image; on the same imaging platform, switch to near-infrared band illumination with a peak wavelength of 850 nanometers and a half-width of 30 nanometers, maintain the illumination at 800±40 lux, and use the same camera as the visible light acquisition to complete the near-infrared full-frame scanning under an exposure time of 8±0.5 milliseconds to obtain an original near-infrared image; perform histogram linear stretching normalization based on a standard gray board reference on the original visible light image and the original near-infrared image respectively, so that the gray scale range of all pixels is mapped to 0 to 255; under the premise of ensuring the aspect ratio of the pixels, using a bicubic interpolation method, the visible light image and the near-infrared image are cropped and scaled to a unified size of 2048×1536 pixels to generate a color rice standardized image set.

3. The method for rapid location of color genetic locus and gene detection of colored rice grain according to claim 2, wherein, In step 2, the process of obtaining the rice grain surface texture feature through the single texture extraction channel includes: establishing an independent data flow channel in the main control processing unit, named as the single texture extraction channel; allocating a dedicated image register buffer for the single texture extraction channel, with a capacity of 8192 kilobytes per frame, for temporarily storing a single frame of image in the color rice standardization image set; enabling row-column alignment reading mode, so that the pixels in the color rice standardization image set are loaded into the image register buffer in row priority order; sequentially calling three groups of multi-scale convolution kernels in the single texture extraction channel, with the multi-scale convolution kernel sizes being 3x3, 5x5 and 7x7 in turn, and the convolution kernel step being uniform as 1 pixel; performing absolute value shaping operation on the output of each convolution kernel and accumulating to the texture gradient accumulation area to obtain the local gradient map of the rice grain surface texture; performing histogram partitioning statistics on the local gradient map of the rice grain surface texture, dividing it into 32 intensity partitions and calculating the pixel count of each partition to obtain the rice grain surface texture feature vector; and writing the rice grain surface texture feature vector into the texture feature register cache area.

4. The method for rapid positioning of color rice kernel color genetic locus and gene detection according to claim 3, characterized in that, In step 2, the process of obtaining the rice grain color spectrum feature through the single spectrum mapping channel includes: establishing another independent data flow channel in the main control processing unit, named as the single spectrum mapping channel; allocating a dedicated spectrum register buffer for the single spectrum mapping channel, with a capacity of 8192 kilobytes per frame, for temporarily storing the same frame of image in the color rice standardization image set; enabling column-row alignment reading mode, so that the pixels in the color rice standardization image set are loaded into the spectrum register buffer in column priority order; reading the gray scale values of the red, green and blue channels and the near-infrared channel gray scale value for each pixel in the color rice standardization image set; calculating the gray scale difference between the red and green channels, the gray scale difference between the red and blue channels, the gray scale difference between the green and blue channels, and the difference between the visible light average gray scale and the near-infrared gray scale, a total of 4 difference values; sequentially writing the above 4 difference values into the spectrum difference value register unit, and then performing normalization operation to linearly map all the difference values to the range of 0 to 255; combining the normalized 4 difference values into a rice grain color spectrum feature vector, and writing it into the spectrum feature register cache area.

5. The method of claim 4, wherein the color of the rice grain is red. In step 2, the process of performing phase synchronization superposition operation in the fusion buffer area includes: enabling the fusion buffer area, with a capacity of 16384 kilobytes per frame, for temporarily storing the feature vectors at the corresponding pixel index positions in the texture feature register cache area and the spectrum feature register cache area; under the triggering of the unified clock signal output by the time reference generator, synchronously reading the texture feature vector and the spectrum feature vector in the order of increasing pixel index from 0; performing bit-by-bit splicing on the synchronously read texture feature vector and spectrum feature vector, with the splicing order being texture feature first and spectrum feature second; writing the spliced feature vector into the fusion buffer area to form the fusion vector entry corresponding to the pixel position. ​ 6. The method for rapid positioning of color rice kernel color genetic locus and gene detection according to claim 5, characterized in that, In step 2, the process of generating the fusion matrix specifically includes: when the fusion buffer is filled, rearranging all the fusion vector entries into a two-dimensional structure in row-major order, the number of rows being the same as the number of rows of the color rice standardization image set, and the number of columns being equal to the dimension of the texture feature vector plus the dimension of the spectral feature vector; defining the two-dimensional structure after the rearrangement as the fusion matrix, and registering a unique matrix identifier in the main control processing unit.

7. The method of claim 6, wherein the method is used for rapid mapping and detection of a gene for color of rice grain. In step 3, the process of establishing the multi-dimensional hash index of the color rice kernel color genetic locus includes: calling the unique matrix identifier of the fusion matrix, reading the number of rows and columns of the fusion matrix; enabling a multi-level texture block generator in the main control processing unit, and setting the sizes of three levels of blocks in the following order: the size of the first-level block is 64x64 pixels; the size of the second-level block is 32x32 pixels; the size of the third-level block is 16x16 pixels; block the fusion matrix from top to bottom in row-major order; when the number of rows or columns of the matrix cannot be divided by the block size, automatically fill in zero to the next integer multiple size; generate block coordinate mapping table level by level, record the starting row index, starting column index and block serial number of each level block in the fusion matrix; initialize the multi-dimensional hash index buffer in the asynchronous hash mapping controller, with a capacity of 131072 items and a single item width of 256 bits; for each level block in the mapping table, read all the fusion vector entries in the block, and perform the following hash key generation process on all fusion vectors in the same block: select the first 32 dimensions of the texture feature vector and the last 4 dimensions of the spectral feature vector, a total of 36 dimensions; assign 36 increasing prime weights to the 36 feature components respectively; the first 6 prime numbers are 2, 3, 5, 7, 11, and 13 in order, and the rest are sequentially increased; multiply each feature component by the corresponding prime weight and sum to get the hash seed; perform modulo 4294967311 operation on the hash seed to get the primary hash key; take the block serial number, block row index and block column index, and combine them with the primary hash key through bit splicing to form a 128-bit composite hash key; perform two nonlinear concave mappings on the composite hash key: first left shift 13 bits, then right shift 7 bits, and then perform bitwise XOR operation to get the final 128-bit multi-dimensional hash key; write the multi-dimensional hash key into the multi-dimensional hash index buffer, and record the global row index and column index of the corresponding pixel in the fusion matrix at the same time; when the multi-dimensional hash index buffer is full or all blocks are traversed, trigger the asynchronous hash mapping controller to complete the index closure to generate the multi-dimensional hash index of the color rice kernel color genetic locus.

8. The method of claim 7, wherein the color of the rice grain is red. The process of performing threshold screening in step 3 to obtain the site candidate pool includes: enabling the threshold screening device, reading all entries in the multi-dimensional hash index of the color rice kernel color genetic site; for each entry, respectively calculating the absolute gradient mean of the texture feature vector and the overall difference mean of the spectral feature vector; using a double threshold strategy: the texture feature threshold is set to 85; the spectral feature threshold is set to 65; only when the absolute gradient mean is not less than 85 and the overall difference mean is not less than 65, the entry is written into the site candidate pool buffer; the candidate pool capacity is set to 32768 entries, and the single entry width is consistent with the index entry; after the threshold screening is completed, the candidate site pool is output, and the number of candidates is recorded.

9. The method of claim 8, wherein the color of the rice grain is red. In step 3, the linear neighborhood comparison verification is performed on the site candidate pool, and the specific process of finally generating the color rice kernel color genetic site positioning result includes: enabling the linear neighborhood comparison verification device, and sorting the candidate entries in ascending order according to the global row index and column index in the fusion matrix; for each candidate entry, a linear neighborhood window with a row radius of 3 pixels and a column radius of 3 pixels is established; other candidate entries are searched in the neighborhood window, and the Euclidean distance between the texture feature vectors and the Manhattan distance between the spectral feature vectors are calculated; when the Euclidean distance is not more than 45 and the Manhattan distance is not more than 60, it is determined that the neighborhood candidate entries form a consistent cluster, the central entry is retained and the remaining entries are removed; if the central entry is still retained after verification, its pixel coordinates are recorded as the genetic site positioning marker; after the verification of all candidate entries is completed, the color rice kernel color genetic site positioning result set is output. ​

Images