Listeria monocytogenes identification method based on image recognition

Abstract

This invention relates to the field of image recognition and detection technology for foodborne pathogens, and discloses a method for identifying Listeria monocytogenes based on image recognition. The identification method includes: acquiring images of chromogenic culture medium plates and converting them to Lab and HSV color spaces; performing adaptive color normalization based on the background region of the culture medium; performing colony instance segmentation on the normalized image and extracting extended regions of interest; extracting saturation distribution sequences along radial rays and detecting halo transition patterns using first-order difference; statistically analyzing halo angle coverage and average transition amplitude, and combining halo quantization feature vectors; inputting the feature vectors into a gradient boosting decision tree classifier to output colony identification results. This invention solves the technical problems of unstable classifier discrimination boundaries caused by color shifts between different batches and the difficulty in detecting halos caused by weak lecithinase reactions.

Description

Technical Field

[0001] This invention relates to the field of image recognition and detection technology for foodborne pathogens, and more specifically, to a method for identifying Listeria monocytogenes based on image recognition. Background Technology

[0002] In the detection of foodborne pathogens at disease control centers, Listeria chromogenic agar is widely used for the preliminary screening and identification of Listeria species. Listeria monocytogenes (Listeria monocytogenes) produces an opaque lecithinase halo around its colonies due to the hydrolysis of substrates in the culture medium by its secreted phospholipase C, resulting in blue-green colonies with this halo; other Listeria species do not produce halos around their colonies. Based on these differences in chromogenic reaction, image recognition methods are commonly used to classify and identify colonies after collecting images of the culture medium plates, a common auxiliary detection method.

[0003] Existing image recognition methods typically use the RGB color mean or dominant hue of the colony region as color features, inputting them into a classifier to determine the colony species. However, existing methods have the following technical problems: First, batch differences in culture media, deviations in culture time, and fluctuations in culture temperature can lead to changes in the intensity of the color reaction, causing significant shifts in the overall color of colonies of the same genus between different batches. When using the overall color mean as a classification feature, the color shifts between different batches make the discrimination boundary of the classifier unstable, resulting in decreased classification accuracy. Second, the characteristic lecithinase reaction halo of Listeria monocytogenes may be extremely weak in the early stages of culture or when there are deviations in the composition of the culture medium. Using the overall color mean as a feature globally averages the color information of the colony region, failing to distinguish the spatial distribution differences between the colony body color and the halo region color. This results in insufficient differentiation between colonies with and without halos in the feature space, leading to misclassification of Listeria monocytogenes with other Listeria species. Summary of the Invention

[0004] This invention provides an image recognition-based method for identifying Listeria monocytogenes colonies on chromogenic culture medium plates, which solves the technical problems in related technologies such as reliance on manual visual judgment, susceptibility to interference from color differences between batches of culture medium, difficulty in quantifying halo features, and insufficient objectivity of identification results.

[0005] This invention discloses a method for identifying Listeria monocytogenes based on image recognition, comprising: acquiring images of chromogenic culture medium plates; converting the original plate images from RGB color space to Lab color space and HSV color space respectively; in Lab color space, dividing pixels into culture medium background region and non-background region by clustering; calculating the color reference value of the culture medium background region; performing color normalization operation on each pixel in the image based on the difference between the color reference value and a preset standard reference value to obtain a color-normalized Lab image; converting the color-normalized Lab image to HSV color space to obtain a color-normalized HSV image; performing colony instance segmentation on the color-normalized Lab image to obtain the segmented region of each colony; calculating the centroid coordinates and equivalent radius of each colony; and truncating the segmented region with the centroid as the center and the equivalent radius as a preset multiple. An extended region of interest (ROI) is extracted, and the corresponding spatial location of the extended ROI is cropped from the color-normalized HSV image. Within the extended ROI corresponding to each colony, saturation channel values ​​are extracted pixel by pixel along multiple radial ray directions starting from the colony centroid to generate a saturation radial distribution sequence. A first-order difference operation is performed on the saturation radial distribution sequence to obtain a saturation gradient sequence. A saturation drop followed by a rise abrupt change pattern is detected in the saturation gradient sequence to obtain a halo abrupt change detection result. Based on the halo abrupt change detection result, the halo angle coverage and average abrupt change amplitude are calculated. Combined with the colony body color histogram distribution features extracted from the color-normalized Lab image, a halo quantization feature vector is formed. The halo quantization feature vector is input into a pre-trained classifier, which outputs the identification result of each colony as Listeria monocytogenes or other Listeria species.

[0006] Further, the step of dividing pixels into culture medium background regions and non-background regions through clustering, calculating the color reference value of the culture medium background region, and performing color normalization operation on each pixel in the image based on the difference between the color reference value and the preset standard reference value includes: extracting pixel values ​​of the a-channel and b-channel in the Lab color space; performing K-means clustering on the pixel values ​​using the two-dimensional color vector composed of the a-channel and b-channel as input, setting the number of clusters to at least two, and determining the category with the most pixels in the clustering results as the culture medium background region; calculating the average value of all pixels in the culture medium background region on the L-channel, a-channel, and b-channel to obtain the color reference value of the current batch; for each pixel in the image, subtracting the average value of the culture medium background region of the current batch on the corresponding channel from the value of the pixel on the L-channel, a-channel, and b-channel respectively, and then adding the value of the preset standard reference value on the corresponding channel respectively to obtain the normalized pixel value; the standard reference value is determined based on the Lab average value of the culture medium background region in the reference batch image collected under standard culture conditions.

[0007] Further, the step of performing colony instance segmentation on the color-normalized Lab image to obtain the segmentation region of each colony includes: calculating the mean and standard deviation of pixel values ​​in the culture medium background region on the a channel and b channel of the color-normalized Lab image, respectively; determining pixels outside the interval determined by the mean plus or minus a preset multiple of the standard deviation as candidate colony pixels; performing connected component labeling on the candidate colony pixels; determining connected components with an area greater than a preset minimum area threshold as independent colony instances; and using the pixel mask of each connected component as the segmentation result of the colony.

[0008] Furthermore, the extended region of interest is extracted with the centroid as the center and an equivalent radius of a preset multiple as the extension distance, wherein the preset multiple ranges from 1.5 to 3.0; when extracting the extended region of interest, the pixels in each extended region of interest that belong to other colony segmentation regions are masked, and the values ​​of the pixels are replaced with the average color value of the culture medium background region to obtain an extended region of interest that removes interference from adjacent colonies.

[0009] Furthermore, the multiple radial rays are evenly distributed within a 360-degree range with the colony centroid as the center and fixed angular intervals, the angular intervals ranging from 5 to 15 degrees. Each ray extends from the centroid to the boundary of the extended region of interest. The pixel-by-pixel extraction of saturation channel values ​​includes: for each radial ray, starting from the colony centroid pixel coordinates, calculating the pixel coordinates at each radial distance along the ray direction with a single pixel step size, and using a bilinear interpolation algorithm to obtain the saturation channel values ​​at non-integer coordinates; before performing the first-order difference operation, performing sliding window smoothing on the saturation radial distribution sequence of each radial ray, the width of the sliding window being 3 to 7 pixels, the smoothing method being the mean operation of the pixel values ​​within the window, and performing the first-order difference operation based on the smoothed saturation radial distribution sequence.

[0010] Further, the detection of a jump pattern in the saturation gradient sequence where saturation first decreases and then increases includes: defining a negative value segment as an interval in the saturation gradient sequence where the gradient value is continuously less than a preset gradient amplitude threshold, and defining a positive value segment as an interval where the gradient value is continuously greater than the gradient amplitude threshold; when a negative value segment appears first and then a positive value segment appears in the interval outside the edge of the colony body, a jump pattern is determined to exist; the value of the gradient amplitude threshold is based on the absolute value of the saturation gradient on the radial rays in the background region of the culture medium in the training sample. The sum of the mean and standard deviation of the values ​​is used to determine the value. After detecting the jump mode, the brightness channel values ​​are extracted pixel by pixel along the radial ray to generate a brightness radial distribution sequence. Within the radial distance range corresponding to the saturation jump interval, the values ​​of that interval in the brightness radial distribution sequence are extracted. The relationship between the brightness mean of the first half of the interval and the brightness mean of the second half of the interval is used as the criterion. If the brightness mean of the first half is less than that of the second half, the jump mode is determined to be a valid halo jump. If the brightness mean of the first half is greater than that of the second half, the jump mode is determined to be a pseudo jump.

[0011] Further, the calculation of halo angle coverage and average jump amplitude based on the halo jump detection results includes: the halo angle coverage is the ratio of the number of radial rays with detected effective jump modes to the total number of radial rays; the average jump amplitude is the mean of the saturation jump amplitudes on all radial rays with detected effective jump modes, and the saturation jump amplitude is the sum of the absolute value of the minimum gradient value of the negative segment and the maximum gradient value of the positive segment in the saturation gradient sequence on a single radial ray; the average jump amplitude is normalized to the mean based on the range, scaling the average jump amplitude to the range of 0 to 1; for colonies without detected effective jump modes, the halo angle coverage is set to zero, and the average jump amplitude is set to zero.

[0012] Furthermore, the combination of halo quantization feature vectors further includes: for each radial ray that detects an effective transition mode, recording the difference between the starting radial distance of the saturation decrease segment and the ending radial distance of the increase segment as the halo width estimate in that direction, and calculating the mean and standard deviation of the halo width estimates on all effective radial rays; after scaling the mean and standard deviation of the halo width estimates to the range of 0 to 1 using mean normalization based on the range, concatenating them into the halo quantization feature vector; the extraction method of the colony body color histogram distribution features is as follows: in the color-normalized Lab image, extracting the color values ​​of all pixels in the colony body segmentation area in the a channel and b channel, dividing the value range of the a channel and b channel into a preset number of uniform intervals, counting the frequency of pixels falling into each interval in the colony body segmentation area, obtaining the normalized frequency distribution vector, and using it as the color histogram distribution features of the a channel and the b channel, respectively.

[0013] Furthermore, the pre-trained classifier is a gradient boosting decision tree classifier. This classifier uses a cross-entropy loss function to measure classification error, gradually adding weak learners to the decision trees in a gradient boosting manner. Each decision tree fits the negative gradient residual of the current ensemble model on the training samples. The weighted sum of the outputs of all decision trees is mapped to class probabilities using a sigmoid function. These class probabilities are then thresholded and converted into class labels for Listeria monocytogenes or other Listeria species. The cross-entropy loss function takes the true class labels of each sample in the training sample set and the class probabilities output by the gradient boosting decision tree classifier as input, and averages the cross-entropy of all training samples as the objective function to measure classification error. During the training phase, halo quantization feature vectors of labeled colony samples cultured on various brand culture media are used as training data. During deployment, the output threshold of the gradient boosting decision tree classifier is fine-tuned based on labeled samples in the target detection environment.

[0014] This invention discloses a Listeria monocytogenes colony identification system based on radial color distribution analysis, comprising: an image acquisition and color space conversion module for acquiring images of chromogenic culture medium plates and converting the original plate images from RGB color space to Lab color space and HSV color space respectively; an adaptive color normalization module for dividing pixels into culture medium background region and non-background region in Lab color space through clustering, calculating the color reference value of the culture medium background region, and performing color normalization operation on each pixel in the image based on the difference between the color reference value and a preset standard reference value to obtain a color-normalized Lab image and a color-normalized HSV image; and a colony segmentation and extended region extraction module for performing colony instance segmentation on the color-normalized Lab image to obtain the segmented region of each colony and calculating the centroid of each colony. The system uses coordinates and equivalent radius to extract extended regions of interest (ROIs) and HSV extended ROIs. A radial color distribution analysis and halo detection module is used to extract saturation channel values ​​pixel-by-pixel along multiple radial ray directions within the HSV extended ROI corresponding to each colony, generating a saturation radial distribution sequence. A first-order difference operation is performed on the saturation radial distribution sequence to obtain a saturation gradient sequence. A jump pattern where saturation first decreases and then increases is detected in the saturation gradient sequence to obtain halo jump detection results. A feature quantization module is used to calculate the halo angle coverage and average jump amplitude based on the halo jump detection results, combining them with the colony body color histogram distribution features to form a halo quantification feature vector. A classification and identification module is used to input the halo quantification feature vector into a pre-trained classifier and output the identification result of each colony as *Listeria monocytogenes* or other *Listeria* species.

[0015] This invention addresses the technical problem of classifier boundary drift caused by unstable color features under different batches and culture conditions, by using adaptive color normalization based on the culture medium background region to correct the culture medium background color of each image to a unified standard reference value. This achieves the technical effect of ensuring a consistent benchmark for images from different batches in the color space. Furthermore, this invention transforms the detection of lecithinase reaction halos into the detection of saturation spatial gradient changes by extracting saturation channel values ​​pixel-by-pixel along the radial direction and performing first-order difference operations. This solves the technical problem of insensitivity to weak lecithinase reaction halos when using the overall color mean as a feature, achieving the technical effect of extracting the saturation drop-then-rise transition pattern even when the halo is extremely weak. Finally, this invention solves the technical problem of insufficient distinguishability between halos and halo-free colonies in the feature space by quantifying the halo angle coverage and average transition amplitude into numerical features and combining them with the colony body color histogram distribution features to form a halo quantification feature vector. This achieves the technical effect of providing the classifier with a quantified index of optical characterization of lecithinase active products with clear physical meaning. Attached Figure Description

[0016] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0017] Figure 1 This is a flowchart of the image recognition-based Listeria monocytogenes identification method provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of a typical pixel color space conversion comparison (Lab channel) provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the Lab channel offset before and after color normalization provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the segmentation feature parameters of the colony instance 001 provided in the embodiment of the present invention; Figure 5 This is a schematic diagram of the radial ray saturation distribution profile of colony CO30 degrees provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the radial ray saturation gradient sequence of colony CO30 provided in an embodiment of the present invention; Figure 7 This is a schematic diagram comparing the normalization of the quantitative features of colony CO3 halo before and after normalization, provided in an embodiment of the present invention. Figure 8 This is a schematic diagram of the scatter distribution of the halo angle coverage and normalized average jump amplitude of the plate 001 colony provided in the embodiment of the present invention. Figure 9 This is a schematic diagram of the output probability and identification results of the plate 001 colony classifier provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Example like Figures 1-9 As shown, in the detection of foodborne pathogens at the Centers for Disease Control and Prevention (CDC), Listeria monocytogenes chromogenic medium is widely used for the preliminary screening and identification of Listeria species. On the chromogenic medium, Listeria monocytogenes (hereinafter referred to as Listeria monocytogenes) produces an opaque lecithinase reaction halo around its colonies due to the hydrolysis of the substrate in the medium by its secreted phospholipase C, resulting in a blue-green colony with the lecithinase reaction halo. Other Listeria species, however, only appear blue-green and do not produce a halo around their colonies. Based on these differences in chromogenic reaction, image recognition methods are commonly used to classify and identify colonies after acquiring images of the culture medium plates, a common auxiliary detection method.

[0020] Existing image recognition methods typically use the RGB color mean or dominant hue of the colony region as color features, inputting them into a classifier to determine the colony species. However, in actual disease control and prevention testing operations, the following technical problems exist: Firstly, batch variations in culture media, deviations in culture time, and fluctuations in culture temperature can all lead to changes in the intensity of the colorimetric reaction, causing significant shifts in the overall color of colonies of the same genus between different batches. When using the average overall color as a classification feature, these color shifts between batches make the classifier's discrimination boundary unstable, resulting in decreased classification accuracy.

[0021] Secondly, the characteristic lecithinase reaction halo of Listeria monocytogenes may be extremely weak in the early stages of culture or when there are deviations in the culture medium composition. The difference between the lecithinase reaction halo and the colony body color in the RGB space is easily masked by the interference of the culture medium background color. The method of using the overall color mean as a feature performs a global average of the color information of the colony area, which cannot distinguish the spatial distribution difference between the colony body color and the halo area color. This results in insufficient differentiation between colonies with and without halos in the feature space, leading to misclassification of Listeria monocytogenes from other Listeria species.

[0022] According to an embodiment of this invention, an image recognition-based method for identifying Listeria monocytogenes is provided. This method is applied to the data processing stage after acquiring images of chromogenic culture medium plates. Before performing the steps of this embodiment, it should be understood that the computing device executing the image recognition-based method for identifying Listeria monocytogenes is equipped with at least an image acquisition device (for acquiring images of culture medium plates) and a processor with image processing capabilities.

[0023] The steps of this implementation method are as follows: Step 1: Acquire images of chromogenic culture medium plates and perform multi-color space conversion. Images of chromogenic culture medium plates were acquired to obtain the original plate images in the RGB color space. The original plate images were then converted from the RGB color space to the Lab and HSV color spaces, respectively, to obtain the Lab and HSV color space representations of the same image.

[0024] It should be noted that the Lab color space mentioned above is the CIE Lab color space, where the L channel represents lightness, the a channel represents the color components from green to red, and the b channel represents the color components from blue to yellow. In the HSV color space mentioned above, the H channel represents hue, the S channel represents saturation, and the V channel represents brightness. The reason for using the Lab and HSV color spaces is that the Lab color space separates lightness information from color information, which is suitable for background segmentation and color normalization processing based on color information in subsequent steps; the HSV color space expresses saturation independently, which is suitable for halo feature detection based on saturation changes in subsequent steps.

[0025] Step 2: Adaptive color normalization based on the culture medium background region In the Lab color space representation, pixel values ​​from channels a and b are extracted. K-means clustering is then performed on these values, with at least two clusters. This divides the image pixels into a culture medium background region and a non-background region, obtaining the segmentation result for the culture medium background region. K-means clustering uses a two-dimensional color vector composed of channels a and b as input and the category label of each pixel as output. Since the culture medium background region occupies the largest connected region in the plate image, the category with the most pixels in the clustering result is determined as the culture medium background region.

[0026] Calculate the average value of all pixels in the background area of ​​the culture medium in the L, a, and b channels to obtain the color reference value for the current batch. Based on the difference between the color reference value of the current batch and the preset standard reference value, perform the following color normalization operation on each pixel in the Lab color space representation of the entire image to obtain the color-normalized Lab image.

[0027] Let the color base value of the current batch be... The preset standard reference value is Lab value for each pixel in the image Perform the following operations to obtain the normalized pixel values. :

[0028]

[0029]

[0030] in, , , These represent the average values ​​of the background region of the current batch of culture medium in channels L, a, and b, respectively. , , These are the preset standard reference values ​​in the L channel, a channel, and b channel, respectively. , , These represent the normalized pixel values ​​in the L, a, and b channels, respectively. In the formula above, L, ... , , All are dimensionless values ​​of the L channel in the Lab color space, a, , , All are dimensionless values ​​for channel a, b, , , All values ​​are dimensionless values ​​from the b channel, with consistent dimensions, and addition and subtraction operations are valid. The above standard reference values ​​are determined based on the Lab average value of the culture medium background area in reference batch images acquired under standard culture conditions.

[0031] Convert the color-normalized Lab image to the HSV color space to obtain the color-normalized HSV image.

[0032] Step 3: Perform colony instance segmentation and extract extended regions of interest. Colony instance segmentation is performed on the color-normalized Lab image to obtain the segmented region of each colony and its corresponding pixel mask. The colony instance segmentation takes the color-normalized Lab image as input, utilizes the color difference between the colony region and the background region in the a and b channels, and combines threshold segmentation with connected component labeling to segment and label each independent colony region as a separate instance, with the pixel mask of each colony as the output.

[0033] Furthermore, the specific implementation of the threshold segmentation is as follows: on the a channel and b channel of the color-normalized Lab image, the mean and standard deviation of the pixel values ​​in the background region of the culture medium are calculated respectively. Pixels outside the interval determined by the mean plus or minus a preset multiple of the standard deviation are determined as non-background pixels, i.e., colony candidate pixels. Then, connected component labeling is performed on the colony candidate pixels. Connected components with an area greater than a preset minimum area threshold are determined as independent colony instances. The pixel mask of each connected component is used as the segmentation result of the colony.

[0034] For each colony segmentation region, the centroid coordinates and equivalent radius are calculated, where the equivalent radius is the radius of a circle with the same area as the segmented colony region. Using the centroid of each colony as the center and an equivalent radius of a preset multiple as the expansion distance, a rectangular image patch containing the colony body and its surrounding culture medium area is extracted to obtain the extended region of interest (ROI) for each colony. The preset multiple ranges from 1.5 to 3.0, and its purpose is to ensure that the extended ROI covers the lecithinase reaction halo region that may exist around the colony body, as well as the culture medium background region outside the lecithinase reaction halo. Simultaneously, the corresponding spatial location region is extracted from the color-normalized HSV image to obtain the HSV extended ROI for each colony.

[0035] In this embodiment of the application, in order to avoid the interference information of adjacent colonies caused by the overlap of the extended regions of interest when colonies are close to each other, when the extended regions of interest are extracted, the pixels in each extended region of interest that belong to other colony segmentation regions are masked, and the values ​​of these pixels are replaced with the average color of the culture medium background region, thereby obtaining the extended regions of interest that have removed the interference of adjacent colonies.

[0036] Step 4: Generate radial color distribution profile and detect halo transition modes Within each colony's corresponding HSV extended region of interest, starting from the colony centroid, the S and V channel values ​​are extracted pixel-by-pixel along multiple radial rays, generating a saturation radial distribution sequence and a brightness radial distribution sequence for each radial ray, thus obtaining a radial color distribution profile. These multiple radial rays are evenly distributed within a 360-degree range, centered on the colony centroid, with fixed angular intervals. Each ray extends from the centroid to the boundary of the extended region of interest, with the angular interval ranging from 5 to 15 degrees.

[0037] Furthermore, the specific implementation of the above pixel-by-pixel extraction is as follows: For each radial ray, according to the given ray direction angle, with the centroid pixel coordinates as the starting point, the pixel coordinates at each radial distance are calculated sequentially along the direction with a single pixel step size. The S-channel and V-channel values ​​at non-integer coordinates are obtained by using a bilinear interpolation algorithm. All sampled values ​​from the centroid to the boundary of the extended region of interest are recorded step by step to form the saturation radial distribution sequence and brightness radial distribution sequence corresponding to the ray.

[0038] A first-order difference operation is performed on the saturation radial distribution sequence of each radial ray in the radial color distribution profile to obtain the saturation gradient sequence. The first-order difference operation uses the saturation radial distribution sequence... As input, the difference between adjacent pixels is calculated based on the radial distance r, using the saturation gradient sequence. As output:

[0039] Where r is the radial pixel distance from the centroid. The saturation value at a radial distance r. This represents the saturation value at a radial distance r+1. Let be the saturation gradient value at a radial distance r. In the above formula, and All values ​​are dimensionless values ​​of the S channel in the HSV color space, with a range of [value range missing]. Since the two quantities have the same dimensions, the difference operation is legal.

[0040] In the saturation gradient sequence, the presence of a saturation transition pattern—a decrease followed by an increase—as the colony extends outwards is detected. The halo transition detection results on each radial ray are obtained. This transition pattern manifests in the saturation gradient sequence as follows: within a certain interval outside the colony edge, First, a series of negative values ​​(decreasing saturation) appear, followed by a series of positive values ​​(increasing saturation). This jump pattern corresponds to the optical characteristics of the opaque lecithinase reaction halo formed around the colony by the lecithinase reaction products. That is, the saturation of the lecithinase reaction halo area decreases due to the presence of opaque substances, while the saturation of the culture medium area outside the lecithinase reaction halo recovers.

[0041] Furthermore, to ensure that the detection of abrupt change modes has a clear criterion, the negative value segment in the saturation gradient sequence is defined as continuously satisfying... The interval, the positive value segment is defined as a continuous interval satisfying The interval, where The preset gradient magnitude threshold is used to exclude the interference of small gradient fluctuations caused by image noise on the detection of abrupt patterns. The value of is determined based on the statistical distribution of the saturation gradient sequence in the background region of the training samples. Specifically, the sum of the mean and standard deviation of the absolute values ​​of the saturation gradient values ​​on the radial rays in the background region of the culture medium in the training samples is used as . The initial value is determined and optimized along with other model parameters during the classifier training phase.

[0042] In this embodiment of the application, in order to reduce the impact of noise on the detection of abrupt change modes on a single radial ray, before performing the first-order difference operation, a sliding window smoothing process is performed on the saturation radial distribution sequence of each radial ray. The width of the sliding window is 3 to 7 pixels, and the smoothing method is the mean operation of the pixel values ​​within the window. After the smoothing process, the first-order difference operation is performed to obtain the smoothed saturation gradient sequence, and the abrupt change mode detection is performed based on the smoothed saturation gradient sequence.

[0043] In this embodiment of the application, to further eliminate spurious transitions caused by the color transition of the colony edge itself, after detecting the transition pattern, the brightness change trend of the corresponding interval in the brightness radial distribution sequence is also verified. When the brightness radial distribution within the saturation transition interval simultaneously shows a trend of first decreasing and then increasing, the transition pattern is determined to be a valid halo transition; otherwise, the transition pattern is determined to be a spurious transition and is not included in the halo detection result.

[0044] Furthermore, the verification method for the above-mentioned brightness change trend is as follows: within the radial distance range corresponding to the saturation jump interval, extract the value of that interval from the radial distribution sequence of brightness. Use whether the difference between the average brightness of the pixels in the first half of the interval and the average brightness of the pixels in the second half of the interval is greater than zero as the criterion. If the average value of the first half is greater than the average value of the second half, it is determined that the trend of first decreasing and then increasing does not hold. If the average value of the first half is less than the average value of the second half, it is determined that the trend of first decreasing and then increasing holds. This verifies the effectiveness of the saturation jump mode.

[0045] Step 5: Combine halo quantization feature vectors based on halo detection results The results of halo transition detection on all radial rays of each colony were statistically analyzed, and the ratio of the number of radial rays with detected effective transition modes to the total number of radial rays was calculated to obtain the halo angle coverage. The mean value of the saturation transition amplitude on all radial rays with detected effective transition modes was calculated to obtain the average transition amplitude; the saturation transition amplitude is the sum of the absolute value of the minimum gradient value in the negative segment and the maximum gradient value in the positive segment of the saturation gradient sequence on a single radial ray.

[0046] It should be noted that the halo angle coverage rate is the ratio of the number of radial rays, ranging from 0 to 1, and is a dimensionless value; the average jump amplitude is the mean of the saturation gradient values, and its dimension is the same as the S-channel value, also a dimensionless value. Before concatenating the halo angle coverage rate, average jump amplitude, and color histogram distribution features into a feature vector, the average jump amplitude is normalized based on the range, scaling it to the same 0 to 1 range as the halo angle coverage rate, thus eliminating the influence of differences in the numerical ranges of each feature component on subsequent classifier calculations.

[0047] Meanwhile, in the color-normalized Lab image, the color values ​​of all pixels in the colony body segmentation area in the a and b channels are extracted. The value range of the a or b channel is divided into a preset number of uniform intervals. The frequency of pixels falling into each interval in the colony body segmentation area is counted, and the normalized frequency distribution vector is obtained, which are used as the color histogram distribution features of the a channel and the color histogram distribution features of the b channel, respectively.

[0048] The normalized halo angle coverage, normalized average hopping amplitude, and color histogram distribution features of channel a and channel b are concatenated to obtain the halo quantization feature vector for the colony. For colonies where no effective hopping mode is detected, the halo angle coverage and average hopping amplitude are set to zero, while retaining the color histogram distribution features of channels a and b of the colony itself. These are then combined into the halo quantization feature vector using the same concatenation method described above.

[0049] In this embodiment, to enhance the ability of the feature vector to distinguish halos of different degrees, the following features are extracted: For each radial ray that detects a valid transition mode, the difference between the starting radial distance of the saturation decrease segment and the ending radial distance of the increase segment is recorded as the halo width estimate in that direction. The mean and standard deviation of the halo width estimates on all valid radial rays are calculated, and the mean and standard deviation of the halo width estimates are concatenated into the halo quantization feature vector. The mean of the halo width estimates reflects the degree of halo development, and the standard deviation of the halo width estimates reflects the spatial uniformity of the halo. The unit of the halo width estimates is pixels. Before concatenating them into the halo quantization feature vector, mean normalization based on the range is also used to scale the halo width estimates to the range of 0 to 1.

[0050] Step 6: Input the halo quantification feature vector into the classifier and output the colony identification results. The halo quantization feature vector of each colony is input into a pre-trained gradient boosting decision tree classifier. The gradient boosting decision tree classifier takes the halo quantization feature vector as input and the category label of Listeria monocytogenes or other Listeria species as output. It performs classification based on the joint discrimination conditions of halo angle coverage, average jump amplitude and colony body color histogram distribution characteristics, and outputs the identification result of each colony as Listeria monocytogenes or other Listeria species.

[0051] The gradient boosting decision tree classifier described above uses the cross-entropy loss function to measure classification error. It progressively adds weak learners to the decision trees using a gradient boosting approach. Each decision tree fits the negative gradient residual of the current ensemble model on the training samples. Finally, the outputs of all decision trees are weighted and summed, then mapped to class probabilities using the sigmoid function. These class probabilities are then thresholded and converted into class labels for *Listeria monocytogenes* or other *Listeria* species, serving as the final colony identification result. During the training phase, the halo quantization feature vectors of labeled colony samples are used as training data. The Adam optimization algorithm is employed to update the model parameters, enabling the gradient boosting decision tree classifier to learn discrimination rules that comprehensively consider halo angle coverage, average jump amplitude, and the distribution of the colony body color histogram.

[0052] Furthermore, the specific form of the aforementioned cross-entropy loss function is as follows: Suppose the training sample set contains a total of M labeled colony samples, and for the j-th sample, its true class label is... ( Indicates Listeria monocytogenes, (representing other Listeria species), the class probability output by the gradient boosting decision tree classifier is... Then the cross-entropy loss function Defined as:

[0053] Where M is the total number of training samples, and j is the sample index. Let j be the true class label of the j-th sample. The gradient boosting decision tree classifier outputs the probability of belonging to Listeria monocytogenes for the j-th sample, with a value range of (0,1). The value is the natural logarithm. The above loss function is the average of the sample-by-sample cross-entropy of all M samples in the training set, which measures the average classification error of the gradient boosting decision tree classifier on the entire training set. During the gradient boosting process, each newly added decision tree aims to reduce the cross-entropy loss function.

[0054] In this embodiment, to accommodate the use of different brands of chromogenic culture media by different disease control centers, colony samples cultured on various brands of culture media are used for training during the gradient boosting decision tree classifier training phase. This ensures that the discrimination rules of the gradient boosting decision tree classifier can cover the distribution range of halo quantification features on different culture media. When deployed to a specific disease control center, the output threshold of the gradient boosting decision tree classifier is fine-tuned based on a small number of labeled samples from that center, thereby adapting it to the specific testing environment.

[0055] Technical effects of this embodiment This implementation uses adaptive color normalization based on the culture medium background region in step 2 to correct the culture medium background color of each image to a uniform standard reference value, ensuring that images acquired under different batches and culture conditions have a consistent baseline in the color space. Therefore, the colony body color features and saturation variation features extracted in subsequent steps are no longer affected by global color shifts caused by batch differences in culture medium, culture time deviations, and culture temperature fluctuations, overcoming the problem of classifier discrimination boundary drift caused by unstable color features between different batches.

[0056] This embodiment transforms the detection of lecithinase reaction halos from a comparison of global color mean to the detection of changes in saturation spatial gradient by extracting saturation channel values ​​pixel-by-pixel along the radial direction and performing first-order difference operations in step 4. Since first-order difference operations are sensitive to the rate of change of the signal rather than its absolute value, even if the absolute difference in saturation between the lecithinase reaction halo region and the colony body is weak, the rate of change in saturation corresponding to the lecithinase reaction halo can still be characterized through difference operations. Therefore, the image recognition-based Listeria monocytogenes identification method can extract the abrupt change pattern of saturation—first decreasing and then increasing—from the radial color distribution profile, even when the lecithinase reaction halo is extremely weak due to early culture conditions or variations in culture medium composition. This overcomes the problem of insensitivity to weak lecithinase reaction halos when using the overall color mean as a feature.

[0057] Furthermore, this embodiment quantifies the halo angle coverage and average jump amplitude into numerical features in step 5, transforming the discrimination criterion of the gradient boosting decision tree classifier from fuzzy color similarity to quantitative indicators of the optical characterization of lecithinase activity products with clear physical meaning. The halo angle coverage reflects the distribution range of the lecithinase reaction halo around the colony, and the average jump amplitude reflects the color intensity of the lecithinase reaction halo. Combined with the colony body color histogram distribution characteristics, these two indicators provide the gradient boosting decision tree classifier with a direct correlation to Listeria monocytogenes phospholipase C activity, thereby improving the specificity of identification.

[0058] I need to carefully check for consistency issues between the application examples and the specific implementation steps.

[0059] Check each item one by one: Table 1: Pixel sampling data. Check if the color space conversion values ​​are reasonable - RGB(98,142,112) to Lab(55.4,-18.7,9.3) and HSV(0.389,0.310,0.557). The values ​​are reasonable and there are no obvious problems.

[0060] Table 2 Normalization Formulas: The formulas in the examples are as follows: That is, 55.4 - 78.2 + 76.5 = 53.7. It needs to be confirmed that this is consistent with the implementation method description—this is standard translation normalization, which is reasonable.

[0061] Table 4 Gradient Calculation: Looking up the table, S(44) = 0.231, but S(45) is not listed in the table (the next row in the table is r = 47, S = 0.198). S(45) = 0.187 does not appear in the table. The calculated value itself is not wrong. It is just that the interpolation point is not in the table, which is acceptable.

[0062] Table 5 Calculation of Jump Amplitude: The text states that "the jump amplitude is the sum of the absolute value of the minimum gradient in the negative segment and the maximum gradient value in the positive segment." However, the "average jump amplitude (original)" in Table 5 is 0.079, which is the average of 24 rays. The 0-degree ray is 0.088, and the average of 0.079 is reasonable and there is no contradiction.

[0063] Table 5. Standard deviation of normalized halo width: Original value 2.1 pixels, normalized value 0.134 — This needs to be confirmed as range-mean normalization. The description should state "processed by range-mean normalization," consistent with the step description.

[0064] The 36-dimensional feature vector consists of: halo angle coverage (1) + normalized average jump amplitude (1) + a-channel histogram (16) + b-channel histogram (16) + normalized halo width mean (1) + normalized halo width standard deviation (1) = 36 dimensions, which is calculated correctly.

[0065] Table 6 Classification Thresholds: The text states "judged by the threshold (0.5)", C03 The diagnosis was Listeria monocytogenes, consistent with the previous diagnosis.

[0066] Centroid coordinates of colony C03: The centroid of C03 in Table 3 is (row 634, column 701), which is consistent with "colony body pixels (row 634, column 701)" in Table 1 and "its centroid coordinates are (row 634, column 701)" in step 3.

[0067] Calculation of the side length of the extended region of interest: equivalent radius 42 pixels, extended distance = 2.0 × 42 = 84 pixels, side length = 2 × 84 = 168 pixels. The side length of C03 in Table 3 is 168 pixels, which is consistent.

[0068] Number of radial rays: The text states "24 radial rays are generated evenly at 15-degree angle intervals (covering 360 degrees)," and 360 / 15=24, which is consistent.

[0069] Halo angle coverage: 20 / 24≈0.833, consistent.

[0070] No obvious inconsistencies were found, and the original data is generally consistent. The complete application scenario and application examples are directly output.

[0071] In 20XX, the Food Safety Inspection Department of a Disease Control Center conducted a special monitoring program on foodborne pathogens in a batch of frozen food samples. After pretreatment, the samples were inoculated onto Listeria monocytogenes chromogenic medium and incubated at 37°C for 48 hours. Images of the agar plates were then captured using an industrial camera mounted on a fixed stand. This batch involved three plates, numbered 001, 002, and 003, each with several blue-green colonies. The inspectors needed to use an image recognition system to automatically identify each colony of Listeria monocytogenes from other Listeria species. Due to a temperature fluctuation of approximately 0.8°C during incubation, the inspectors anticipated potential batch-specific color variations between different plates and therefore relied on the system's adaptive color normalization function to eliminate this effect.

[0072] Taking colony identification on plate 001 as an example, the complete data flow process is demonstrated. After image acquisition, the system reads the original RGB image (2048×2048 pixels resolution) from plate 001 and then initiates a multi-color space conversion process. The system converts the RGB image to Lab color space representation and simultaneously to HSV color space representation. Both representations correspond to different encoding forms of the same original image, which are used in subsequent steps respectively.

[0073] Table 1. Color space conversion results of the original image of board 001 (typical pixel sampling)

[0074] The system performed K-means clustering (with 3 clusters) on the pixel values ​​of the a and b channels of the Lab image. Based on the number of pixels in each category, the category with the most pixels was determined to be the culture medium background area. The background area of ​​plate 001 had 3,187,420 pixels, accounting for 76.1% of the total number of pixels in the entire image, and was confirmed as the culture medium background area. Subsequently, the mean pixel value of the background area in the L, a, and b channels was calculated to obtain the color reference value for this batch.

[0075] Table 2 Color Normalization Parameters for Board 001

[0076] Taking the colony body pixel (634 rows, 701 columns) as an example, its original Lab value is Substitute the values ​​into the normalization formula to calculate the normalized pixel values:

[0077]

[0078]

[0079] After normalization, the Lab mean of the background region was uniformly corrected to the standard reference value. This eliminated the global color shift caused by fluctuations in culture temperature in this batch. The system then converted the normalized Lab image to the HSV color space to obtain the normalized HSV image, which was used for subsequent radial analysis.

[0080] Threshold segmentation and connected component labeling were performed on the color-normalized Lab image. The system detected five independent colony instances on board 001, numbered C01 to C05. For each colony, the centroid coordinates and equivalent radius were calculated. An extended region of interest (ROI) was extracted with an extension distance of 2.0 times the equivalent radius (the preset multiplier is 2.0, falling within the range of 1.5 to 3.0). Taking colony C03 as an example, its centroid coordinates are (row 634, column 701), its equivalent radius is 42 pixels, and its extension distance is 84 pixels. A square image block centered on the centroid with a side length of 168 pixels was extracted as the extended ROI. Simultaneously, the corresponding spatial location was extracted from the normalized HSV image to obtain the HSV extended ROI.

[0081] Table 3. Segmentation results of colony instances from plate 001

[0082] Using the HSV extended region of interest of colony CO3 as the operating object, the system uniformly generates 24 radial rays (covering 360 degrees) at 15-degree intervals centered on the centroid. For each ray, the S and V channel values ​​are extracted pixel by pixel along the radial direction to generate radial distribution sequences of saturation and brightness, forming a radial color distribution profile. Before performing first-order differencing, a sliding window mean smoothing process with a width of 5 pixels is performed on each radial saturation distribution sequence.

[0083] Taking the 0-degree ray (directly to the right) of colony CO3 as an example, this ray extends from the centroid to the boundary of the extended region of interest, with a total of 84 sampling points. The smoothed radial distribution sequence of saturation is shown below. The sequence exhibits a decreasing-then-increasing trend within the radial distance range of r=38 to r=55 pixels, with the corresponding position falling precisely on the outer edge of the colony. Performing a first-order difference on this sequence, taking r=44 as an example:

[0084]

[0085] Table 4. 0-degree radial ray saturation distribution and gradient sequence of CO3 colonies (partial sampling points)

[0086] The system sets the gradient magnitude threshold. =0.025, within the interval r=41 to r=47 (The negative value segment holds true) within the interval r=50 to r=53. (Positive value segment established), a valid transition pattern was detected. Simultaneously, the brightness change trend was verified: the average brightness of the first half (r=41 to r=47) of the transition interval (r=41 to r=53) was 0.621, and the average brightness of the second half (r=48 to r=53) was 0.684. The average brightness of the first half was less than that of the second half, indicating that the brightness first decreased and then increased trend was established, confirming a valid halo transition.

[0087] Statistical analysis of the hopping detection results of all 24 radial rays of colony CO3 revealed that 20 rays showed effective hopping modes, with a halo angle coverage rate of [missing information]. Calculate the saturation jump amplitude of each effective ray and take the average value to obtain the average jump amplitude. Taking the 0-degree ray as an example, its jump amplitude is the sum of the absolute value of the minimum gradient in the negative segment and the maximum gradient value in the positive segment: .

[0088] Table 5. Statistical analysis of the quantitative characteristics of CO3 halo in bacterial colonies.

[0089] Simultaneously, the color histogram distribution features of the a-channel and b-channel of the colony CO3 body segmentation region were extracted from the color-normalized Lab image. The a-channel and b-channel were each divided into 16 uniform intervals, and the pixel frequency distribution was statistically analyzed to obtain 16-dimensional color histogram vectors for the a-channel and b-channel respectively. Finally, the halo angle coverage (1D), normalized average jump amplitude (1D), a-channel color histogram (16D), b-channel color histogram (16D), normalized halo width mean (1D), and normalized halo width standard deviation (1D) were concatenated into a 36-dimensional halo quantization feature vector, which was then input into the classifier.

[0090] The 36-dimensional halo quantized feature vector of colony C03 is input into a pre-trained gradient boosting decision tree classifier. The classifier performs joint discrimination based on the halo angle coverage (0.833), the normalized average jump amplitude (0.761), and the colony body a-channel color histogram (the main frequency is concentrated in the -22 to -14 interval, reflecting the color characteristics of blue-green colonies), and outputs the probability of belonging to the Listeria monocytogenes class. After being judged by the threshold (0.5), the output category label is "Listeria monocytogenes".

[0091] Table 6 Summary of all colony identification results for plate 001

[0092] In this identification process, the data started with the original RGB image of board 001. Step 1 involved multi-color space conversion to obtain Lab and HSV dual-channel representations, providing a foundation for color information separation in subsequent processing. In Step 2, the system used K-means clustering to identify the background region from the Lab image and calculated the batch color baseline value. Compared with standard reference value To correct the discrepancies between the samples, translation correction is performed on each pixel of the entire image, eliminating color shifts introduced by temperature fluctuations in this batch and ensuring cross-batch comparability of subsequently extracted features. In step 3, the normalized Lab image supports accurate colony segmentation. The centroid and equivalent radius parameters of the five colony instances directly determine the truncation range of the expanded region of interest and the extension length of the radial rays in step 4. In step 4, the normalized HSV image provides input for radial saturation profile analysis. First-order difference transforms the saturation change of colony CO3 from r=38 to r=55 pixels, which is first decreased and then increased, into a quantifiable jump pattern. Combined with brightness verification, false jump interference is eliminated. In step 5, the jump detection results are aggregated into quantitative indicators such as halo angle coverage, average jump amplitude, and halo width statistics. These are concatenated with the colony body color histogram to form a 36-dimensional feature vector, compressing spatial distribution information into a structured numerical expression. In step 6, the classifier completes the final discrimination based on the feature vector. Among the five colonies on plate 001, C03 and C04 are identified as Listeria monocytogenes, while C01, C02, and C05 are identified as other Listeria species. The results are directly output to the testing personnel's work interface to support subsequent epidemic risk assessment.

[0093] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for identifying Listeria monocytogenes based on image recognition, characterized in that, Includes the following steps: Images of chromogenic culture medium plates were acquired, and the original plate images were converted from the RGB color space to the Lab color space and the HSV color space, respectively. In the Lab color space, pixels are divided into culture medium background region and non-background region by clustering. The color reference value of the culture medium background region is calculated. Based on the difference between the color reference value and the preset standard reference value, color normalization operation is performed on each pixel in the image to obtain the color-normalized Lab image. The color-normalized Lab image is then converted to the HSV color space to obtain the color-normalized HSV image. Colony instance segmentation is performed on the color-normalized Lab image to obtain the segmentation region of each colony. The centroid coordinates and equivalent radius of each colony are calculated. An extended region of interest is extracted with the centroid as the center and the equivalent radius as a preset multiple as the extension distance. The HSV extended region of interest corresponding to the spatial position is extracted from the color-normalized HSV image. Within the extended region of interest of HSV corresponding to each colony, the saturation channel values ​​are extracted pixel by pixel along multiple radial ray directions starting from the centroid of the colony to generate a saturation radial distribution sequence. A first-order difference operation is performed on the saturation radial distribution sequence to obtain a saturation gradient sequence. A jump mode in the saturation gradient sequence, where the saturation first decreases and then increases, is detected to obtain the halo jump detection result. Based on the halo jump detection results, the halo angle coverage and average jump amplitude are calculated, and combined with the colony body color histogram distribution features extracted from the color-normalized Lab image, they are combined into a halo quantification feature vector. The halo quantization feature vector is input into a pre-trained classifier, which outputs the identification result of each colony as Listeria monocytogenes or other Listeria species.

2. The method for identifying Listeria monocytogenes based on image recognition according to claim 1, characterized in that, The step of dividing pixels into culture medium background regions and non-background regions through clustering, calculating the color reference value of the culture medium background region, and performing color normalization operation on each pixel in the image based on the difference between the color reference value and a preset standard reference value includes: Pixel values ​​of the a and b channels are extracted in the Lab color space. K-means clustering is performed on the pixel values ​​using the two-dimensional color vector formed by the a and b channels as input. The number of clusters is set to at least two. The category with the most pixels in the clustering results is determined as the culture medium background region. Calculate the average value of all pixels in the culture medium background area on the L channel, a channel, and b channel to obtain the color reference value for the current batch; For each pixel in the image, the average value of the background area of ​​the current batch of culture medium in the corresponding channel is subtracted from the value of the pixel in the L channel, a channel, and b channel respectively, and then the value of the preset standard reference value in the corresponding channel is added to obtain the normalized pixel value; the standard reference value is determined based on the Lab average value of the background area of ​​the culture medium in the reference batch image collected under standard culture conditions.

3. The method for identifying Listeria monocytogenes based on image recognition according to claim 1, characterized in that, The step of performing colony instance segmentation on the color-normalized Lab image to obtain the segmentation region of each colony includes: In the a and b channels of the color-normalized Lab image, the mean and standard deviation of the pixel values ​​in the culture medium background area are calculated respectively. Pixels outside the range determined by the mean plus or minus a preset multiple of the standard deviation are identified as candidate colony pixels. Connectivity region labeling is performed on candidate pixels of colonies. Connectivity regions with an area greater than a preset minimum area threshold are identified as independent colony instances. The pixel mask of each connected region is used as the segmentation result of the colony.

4. The method for identifying Listeria monocytogenes based on image recognition according to claim 1, characterized in that, The extended region of interest is extracted by taking the centroid as the center and using an equivalent radius that is a preset multiple as the extension distance, wherein the value of the preset multiple ranges from 1.5 to 3.0; When extracting the extended region of interest, the pixels in each extended region of interest that belong to other colony segmentation regions are masked, and the values ​​of the pixels are replaced with the average color value of the culture medium background region to obtain the extended region of interest that removes interference from adjacent colonies.

5. The method for identifying Listeria monocytogenes based on image recognition according to claim 1, characterized in that, The multiple radial rays are evenly distributed within a 360-degree range with the centroid of the colony as the center and at fixed angular intervals. The angular intervals range from 5 degrees to 15 degrees. Each ray extends from the centroid to the boundary of the extended region of interest. The pixel-by-pixel extraction of saturation channel values ​​includes: for each radial ray, starting from the centroid pixel coordinates of the colony, calculating the pixel coordinates at each radial distance in sequence along the ray direction with a single pixel step size, and using a bilinear interpolation algorithm to obtain the saturation channel values ​​at non-integer coordinates. Before performing the first-order difference operation, a sliding window smoothing process is performed on the saturation radial distribution sequence of each radial ray. The width of the sliding window is 3 to 7 pixels, and the smoothing method is the mean operation of the pixel values ​​within the window. The first-order difference operation is then performed based on the smoothed saturation radial distribution sequence.

6. The method for identifying Listeria monocytogenes based on image recognition according to claim 1, characterized in that, The detection of a transition pattern in the saturation gradient sequence where saturation first decreases and then increases includes: The intervals in the saturation gradient sequence where the gradient value is continuously less than a preset gradient magnitude threshold are defined as negative value segments, and the intervals where the gradient value is continuously greater than the gradient magnitude threshold are defined as positive value segments; when a negative value segment appears first and then a positive value segment appears in the interval outside the edge of the colony body, it is determined that a jump mode exists; The gradient magnitude threshold is determined based on the sum of the mean and standard deviation of the absolute values ​​of the saturation gradient values ​​on the radial rays in the background region of the culture medium in the training samples. After detecting the abrupt change mode, the brightness channel values ​​are extracted pixel by pixel along the radial ray to generate a brightness radial distribution sequence. Within the radial distance range corresponding to the saturation abrupt change interval, the values ​​of that interval in the brightness radial distribution sequence are extracted. The relationship between the average brightness of the pixels in the first half and the average brightness of the pixels in the second half of the interval is used as the criterion. If the average value of the first half is less than the average value of the second half, the abrupt change mode is determined to be a valid halo abrupt change; if the average value of the first half is greater than the average value of the second half, the abrupt change mode is determined to be a pseudo abrupt change.

7. The method for identifying Listeria monocytogenes based on image recognition according to claim 1, characterized in that, The calculation of halo angle coverage and average jump amplitude based on the halo jump detection results includes: The halo angle coverage rate is the ratio of the number of radial rays that detect effective transition modes to the total number of all radial rays. The average jump amplitude is the mean of the saturation jump amplitudes on all radial rays that have detected valid jump modes, and the saturation jump amplitude is the sum of the absolute value of the minimum gradient value of the negative segment and the maximum gradient value of the positive segment in the saturation gradient sequence on a single radial ray. The average jump amplitude is subjected to mean normalization based on the range, which scales the average jump amplitude to the range of 0 to 1. For colonies for which no effective hopping pattern was detected, the halo angle coverage was set to zero, and the average hopping amplitude was set to zero.

8. The method for identifying Listeria monocytogenes based on image recognition according to claim 7, characterized in that, The combination of halo quantization feature vectors also includes: For each radial ray that detects an effective transition mode, the difference between the starting radial distance of the saturation decrease segment and the ending radial distance of the increase segment is recorded as the halo width estimate in that direction. The mean and standard deviation of the halo width estimates on all effective radial rays are calculated. The mean and standard deviation of the halo width estimate are scaled to the range of 0 to 1 using mean normalization based on the range, and then concatenated into the halo quantization feature vector. The extraction method of the colony body color histogram distribution features is as follows: In the color-normalized Lab image, extract the color values ​​of all pixels in the colony body segmentation area in the a channel and b channel, divide the value range of the a channel and b channel into a preset number of uniform intervals, count the frequency of pixels falling into each interval in the colony body segmentation area, and obtain the normalized frequency distribution vector, which are respectively used as the color histogram distribution features of the a channel and the color histogram distribution features of the b channel.

9. The method for identifying Listeria monocytogenes based on image recognition according to claim 1, characterized in that, The pre-trained classifier is a gradient boosting decision tree classifier. The gradient boosting decision tree classifier uses the cross-entropy loss function to measure the classification error. Weak learners of decision trees are added step by step in the gradient boosting manner. Each decision tree fits the negative gradient residual of the current ensemble model on the training samples. The outputs of all decision trees are weighted and summed and then mapped to class probabilities by the Sigmoid function. The class probabilities are converted into class labels of Listeria monocytogenes or other Listeria species after threshold judgment. The cross-entropy loss function takes the true class label of each sample in the training sample set and the class probability output by the gradient boosting decision tree classifier as input, and takes the average of the cross-entropy of all training samples as the objective function to measure the classification error. During the training phase, halo quantization feature vectors of labeled colony samples cultured on various brands of culture media are used as training data; during deployment, the output threshold of the gradient boosting decision tree classifier is fine-tuned based on labeled samples in the target detection environment.

10. A Listeria monocytogenes colony identification system based on radial color distribution analysis, used to perform the Listeria monocytogenes identification method based on image recognition as described in any one of claims 1 to 9, characterized in that, include: The image acquisition and color space conversion module is used to acquire images of chromogenic culture medium plates and convert the original plate images from the RGB color space to the Lab color space and HSV color space, respectively. An adaptive color normalization module is used to divide pixels into culture medium background regions and non-background regions in the Lab color space by clustering, calculate the color reference value of the culture medium background region, and perform color normalization operation on each pixel in the image based on the difference between the color reference value and the preset standard reference value to obtain the color-normalized Lab image and the color-normalized HSV image. The colony segmentation and extended region extraction module is used to perform colony instance segmentation on the color-normalized Lab image to obtain the segmentation region of each colony, calculate the centroid coordinates and equivalent radius of each colony, and extract the extended region of interest and HSV extended region of interest. The radial color distribution analysis and halo detection module is used to extract saturation channel values ​​pixel by pixel along multiple radial ray directions within the HSV extended region of interest corresponding to each colony to generate a saturation radial distribution sequence. The module performs a first-order difference operation on the saturation radial distribution sequence to obtain a saturation gradient sequence. The module detects a jump pattern in the saturation gradient sequence where the saturation first decreases and then increases, and obtains the halo jump detection result. The feature quantization module is used to calculate the halo angle coverage and average jump amplitude based on the halo jump detection results, and combine them with the colony body color histogram distribution features to form a halo quantization feature vector. The classification and identification module is used to input the halo quantification feature vector into a pre-trained classifier and output the identification result of each colony as Listeria monocytogenes or other Listeria species.

Images