Terahertz far-field imaging method and system based on time sequence classification

Abstract

The invention belongs to the technical field of terahertz far-field imaging, and relates to a terahertz far-field imaging method and system based on time sequence classification. According to the specific technical scheme, a sample is automatically scanned through an S-shaped path automatic planning algorithm, and terahertz time domain data of the sample is collected; performing time zero point positioning on each piece of sequence data of the terahertz time domain data, extracting an effective time domain window, and obtaining effective terahertz time domain data; carrying out supervised learning by adopting a one-dimensional convolutional neural network, taking the effective terahertz time domain data as model input, taking the classification label with the maximum probability in the classification label probability distribution as model output, and constructing and training a prediction model; and carrying out color assignment on a classification prediction result of the prediction model, and carrying out image construction based on a scanning point position to obtain a color classification graph of background segmentation. According to the method, stable imaging detection of the reflector and related metal samples can be realized in a complex environment, and accurate background segmentation is realized.

Description

Technical Field

[0001] This invention belongs to the field of terahertz far-field imaging technology, specifically relating to a terahertz far-field imaging method and system based on time-series classification. Background Technology

[0002] Terahertz waves are electromagnetic waves with frequencies between microwaves and infrared waves. They possess characteristics such as non-ionization, strong penetration, and transparency to most non-polar materials, thus showing broad application prospects in materials testing, structural analysis, safety screening, biomedicine, and non-destructive testing. Terahertz far-field imaging, with its large imaging range and non-contact nature, is particularly useful for identifying and reconstructing sample surfaces.

[0003] Terahertz imaging systems often rely on manually preset scanning paths, lacking intelligent planning and feedback control mechanisms. They also fail to combine automatic path planning with deep learning prediction, resulting in poor adaptability, high data redundancy, and low reconstruction accuracy. Furthermore, terahertz time-domain waveforms are prone to time delay drift, signal distortion, and energy attenuation during far-field propagation, making it difficult for traditional imaging methods that rely on direct signal strength or peak value judgments to accurately reconstruct the true distribution characteristics of the sample. In contrast, deep learning technology possesses superior feature extraction and pattern recognition capabilities in signal processing and image reconstruction.

[0004] To achieve accurate reflection of the true distribution characteristics of samples in terahertz far-field imaging and improve the reconstruction accuracy of imaging, an imaging control system that combines automatic path planning scanning with deep learning prediction is needed. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a terahertz far-field imaging method and system based on time-series classification. Terahertz far-field imaging can accurately reflect the true distribution characteristics of a sample, thereby improving the reconstruction accuracy of the imaging.

[0006] On one hand, this invention discloses a terahertz far-field imaging method based on temporal classification, comprising the following process:

[0007] Step S1: Automatically scan the sample and collect the terahertz time domain data of the sample using the S-shaped path automatic planning algorithm, and store the terahertz time domain data and its coordinates of each scanning point.

[0008] Step S2: Locate the zero point of time for each sequence of terahertz time domain data, extract the effective time domain window, and extract the data from the effective time domain window in the sequence data to obtain the effective terahertz time domain data;

[0009] Step S3: Supervised learning is performed using a one-dimensional convolutional neural network. The effective terahertz time-domain data is used as the model input, and the classification label with the highest probability in the classification label probability distribution is used as the model output. The prediction model is then constructed and trained.

[0010] Step S4: Assign color values ​​to the classification prediction results of the prediction model, construct an image based on the scan point positions, and obtain a color classification map of background segmentation.

[0011] Preferably, in step S1, the S-shaped path automatic planning algorithm includes:

[0012] The top left corner of the scanning area is taken as the starting point (x0, y0), and the coordinates of each scanning point (x0, y0) are... i,j ,y i,j )as follows:

[0013]

[0014] y i,j =y0+iΔ x

[0015] i and j are the i-th row and j-th column, respectively; N x Δ is the number of columns in the grid size of the scanned area. x x0 and y0 represent the scanning step size; x0 and y0 are the horizontal and vertical coordinates of the scanning start point, respectively.

[0016] j = k mod N x

[0017] k mod N represents rounding down; mod represents the remainder operation. x This means k divided by N x The remainder after; k represents the index, the row and column index in row-major order, k∈{0,…,N-1}; N is the total number of grids, N=N x ×N y N y The number of rows in the grid of the scanned area;

[0018] Then the start time of data collection at point k With end time for:

[0019]

[0020] t move The time required for each point to move and stabilize; t meas t0 represents the measurement acquisition time; t0 represents the start time.

[0021] Preferably, step S2 includes the following process:

[0022] Zero-point positioning is achieved using peak detection, as shown in the following formula:

[0023] t 0,k =argmax t |S k (t)|

[0024] t 0,k The reference time point where the main peak is located; argmax t S represents the time point t at which the objective function reaches its maximum value; k (t) is a sampling sequence, where t is time and k is the scan point number;

[0025] From time reference point t 0,k Iterate forward through the signal, finding the point in time when the signal first falls below a set amplitude threshold δ as a valid starting point:

[0026] t a,k =max{t|t <t 0,k ,|S k (t)|<δ}

[0027] Based on the zero point obtained from the positioning, the signal is aligned with the time reference:

[0028] S′ k (t)=S k (tt a,k )S k ′(t) is the alignment signal after time shift;

[0029] Extract the effective time domain window [t] a,k ,t b,k ], t b,k =t a,k +Δ t ,Δ t For a fixed time-domain window length, then:

[0030] S k "(t)=S′ k (t), t∈[t a,k ,t b,k ]

[0031] S′ k ′(t) is the normalized and aligned effective time-domain signal.

[0032] Preferably, the prediction model in step S3 includes: taking the terahertz time-domain data after time zero-point calibration and effective time window extraction as input, using a one-dimensional convolutional neural network for supervised learning, the output of the convolutional layer is processed by batch normalization, max pooling and random deactivation, and global features are extracted through a global average pooling layer to complete the classification.

[0033] Preferably, the color assignment based on the classification result in step S4 includes:

[0034] The category labels are obtained through the prediction model:

[0035]

[0036] x(k) is the time-domain sequence of the kth scan point; c represents the class number predicted by the model; c is the class index, with values ​​{0, 1, 2, ..., C-1}.

[0037] Category number According to the coordinate position of the scan point (x k ,y k Mapped to the corresponding pixel value:

[0038]

[0039] I(x k ,y k ) represents the corresponding pixel value in the imaging result; ColorMap(·) is the mapping function from category to color; (x k ,y k ) represents the spatial coordinates of the k-th point on the scanning path.

[0040] On the other hand, this invention discloses a terahertz far-field imaging system based on time-series classification, comprising:

[0041] The data acquisition unit automatically scans the samples and acquires the terahertz time-domain data of the samples through an S-shaped path automatic planning algorithm, and stores the terahertz time-domain data of each scanning point and its corresponding coordinates.

[0042] The data preprocessing unit performs time zero-point positioning processing on the terahertz time domain data to obtain an effective time domain window, and performs marking processing on the effective time domain window to obtain effective terahertz time domain data;

[0043] The model prediction unit takes effective terahertz time-domain data as input, uses a one-dimensional convolutional neural network for supervised learning, and the output of the convolutional layer is processed by batch normalization, max pooling and random deactivation. Global features are extracted and classification is completed through a global average pooling layer.

[0044] The image construction unit performs color classification based on the classification number predicted by the model prediction unit and constructs the image based on the scan point location to obtain the sample detection image.

[0045] Accordingly, the present invention also discloses a computer device / equipment / system, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above method.

[0046] Accordingly, the present invention also discloses a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0047] Accordingly, the present invention also discloses a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the above-described method.

[0048] The present invention has the following beneficial effects:

[0049] This method achieves terahertz time-domain spectral data acquisition by automatically planning an S-shaped path for scanning and coordinating with a compatible terahertz host for control. In the data preprocessing stage, waveform alignment and effective sequence region extraction are performed on the acquired terahertz time-domain spectral data. The preprocessed dataset is then used for model training, achieving an accuracy rate exceeding 99%. Color classification is performed on the prediction results for each scanning point, and image reconstruction is achieved based on the scanning point location, ultimately completing image construction. Experiments show that this deep learning-based terahertz far-field time-series classification and prediction imaging method can achieve stable imaging detection of mirrors and related metal samples in complex environments, and achieve accurate background segmentation. Attached Figure Description

[0050] Figure 1 This is a flowchart of the imaging method of the present invention;

[0051] Figure 2 This is a structural block diagram of the imaging system of the present invention;

[0052] Figure 3 This is a peak intensity imaging diagram of the imaging method of the present invention;

[0053] Figure 4 This is a one-dimensional convolutional neural network prediction image of the imaging method of the present invention. Detailed Implementation

[0054] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Unless otherwise specified, the technical means used in the embodiments are conventional means well known to those skilled in the art.

[0055] like Figure 1 As shown, this invention discloses a terahertz far-field imaging method based on temporal classification, comprising the following process:

[0056] Step S1: Automatically scan the samples and collect terahertz time-domain data using an S-shaped path automatic planning algorithm, and store the terahertz time-domain data and its coordinates for each scan point. The terahertz time-domain data is the raw terahertz time-domain spectrum data.

[0057] In some implementations, the S-shaped path automatic planning algorithm automatically scans the region samples point by point and acquires the corresponding terahertz time-domain data. The S-shaped path automatic planning algorithm includes the following process:

[0058] Based on the scanning area size and step parameters, an S-shaped motion trajectory is automatically generated. The scanning platform moves alternately in the horizontal direction. At each scanning position, the system automatically triggers the terahertz host to acquire signals, obtain waveform data of the corresponding scanning point, and store the acquired waveform data and scanning coordinates. The saved data serves as the terahertz time domain data of the entire sample.

[0059] The scan area can be gridded according to the set scan step Δ. x The horizontal length L of the scanned area x Longitudinal length L of the scanned area y (The scan step size must be divisible by the horizontal and vertical lengths), and the grid is calculated as follows: OK, Column, i.e., grid size (total number of scan points): N = N x ×N y The row and column indices in row-major order (represented by line indices k∈{1,…,N}) are as follows:

[0060] j = k mod N x

[0061] i and j represent the i-th row and j-th column, respectively; k mod N represents rounding down; mod represents the remainder operation. x This means k divided by N x The remainder after;

[0062] Coordinates of each point in the S-shaped scan:

[0063] y i,j =y0+iΔ x

[0064] Triggering the data collection sequence: The time required for each point to move and stabilize is t. move The measurement and acquisition time is t. meas The starting time is t0. Then, the start time of data collection at point k is... With end time for:

[0065]

[0066] Compared with the traditional "line-back-line" raster scanning, the core advantage of S-shaped scanning is that it eliminates the idle return time and merges the scanning process from a series of discontinuous short line segments into a continuous, long-distance motion.

[0067] Step S2: Locate the zero point of time for each sequence of terahertz time-domain data and extract the effective time-domain window to obtain effective terahertz time-domain data. Specifically, perform signal filtering on each sequence of data, perform signal detection within a preset time region to determine whether a valid signal exists. If a valid signal exists, find the point where the signal value is less than the set amplitude threshold δ based on the forward judgment of the sequence, determine this point as the starting point of the time-domain signal of the sequence data, and use this starting point as a reference to extract the time-domain sequence trend of a preset length, and determine this sequence region as the processed time-domain waveform data.

[0068] Specifically, by sampling each sequence S k (t) (where t is time and k is the scan point number), perform time zero-point positioning (i.e., the time point where the peak value is located in the terahertz time domain data of the kth scan point is t), and determine the reference time point t where the main peak is located. 0,k Then, using this point as a reference, we traverse forward to find the starting point of the main peak, thus achieving alignment of the time-domain signal.

[0069] Zero-point localization can be achieved using peak detection, as shown in the following formula:

[0070] t 0,k =argmax t |S k (t)|

[0071] argmax t This represents the independent variable (i.e., the time point) that causes the objective function to reach its maximum value; the range of t is the acquisition interval of terahertz time-domain data [3950, 4050]ps.

[0072] At the peak time point t 0,kPreviously, the signal was iterated forward to find the time point when the signal first fell below the set amplitude threshold δ as a valid starting point:

[0073] t a,k =max{t|t <t 0,k ,|S k (t)|<δ}

[0074] δ is the set amplitude threshold, with a value range of (0,1); max{t|t <t 0,k ,|S k (t)|<δ} represents t <t 0,k There exists a region within the interval that satisfies |S k The time t when (t)|<δ is defined as the starting time point t.

[0075] Based on the zero point obtained from the positioning, the signal is aligned with the time reference:

[0076] S k ′(t)=S k (tt a,k )S k ′(t) is the alignment signal after time shift, where t is the time variable and the unit is ps.

[0077] Based on this, extract the effective time domain window [t]. a,k ,t b,k ], t b,k =t a,k +Δ t ,Δ t The set fixed time-domain window length (including the main peak and its oscillation tail):

[0078] S k "(t)=S k ′(t),t∈[t a,k ,t b,k ]

[0079] S′ k ′(t) is the normalized and aligned effective time-domain signal, i.e., the effective terahertz time-domain data.

[0080] By performing the above processing on the terahertz time-domain data, the terahertz time delay can be eliminated and the impact of environmental noise can be reduced.

[0081] Step S3: The terahertz time-domain waveform after time zero-point calibration and effective time window extraction is used as input. Supervised learning is performed using a one-dimensional convolutional neural network. The output of the convolutional layer is processed by batch normalization, max pooling and random deactivation. Finally, global features are extracted through a global average pooling layer and classification is completed.

[0082] Specifically, the input signal first undergoes multiple layers of one-dimensional convolution calculations to extract local temporal features. The convolution calculation process of the l-th layer is as follows:

[0083] H l =f(H (l-1) *W l +b l )

[0084] Among them, H l H is the output feature map of the l-th layer. (l-1) W is the output feature map of the previous layer (layer l-1). l For dimension (K) l C l-1 C l The convolution kernel parameters, where K l C is the kernel size. l-1 C represents the number of channels in the previous layer. l b is the number of output channels for the current layer. l The term '*' represents a one-dimensional convolution operation, and f is the ReLU non-linear activation function. This convolution can capture local patterns of the signal in the time dimension, such as peak structure and energy distribution.

[0085] The output of the convolutional layer undergoes batch normalization to improve training stability and accelerate convergence. The calculation formula is as follows:

[0086]

[0087] Where BN(x) is the normalized output, μ B and Let represent the mean and variance of the current batch, γ and β be learnable parameters, and ∈ be a small constant to prevent division by zero. Then, downsampling is performed using max pooling.

[0088]

[0089] in, Ω represents the m-th output feature value after pooling in layer l, and max represents taking the maximum value of the activation values. m H represents m pooling regions. (l) (t) represents the activation value of the feature map output by the l-th convolutional layer at time position t.

[0090] To avoid overfitting, a random deactivation mechanism is used during the training phase to mask the outputs of some neurons. The calculation process is as follows:

[0091]

[0092] in, Indicates the first The output after the layer is randomly deactivated. Indicates the first Layered pooling output, Given a Bernoulli random matrix of the same dimension as the input, To preserve the probability of a neuron, the symbol "⊙" represents the element-wise multiplication operation.

[0093] After multiple convolutions and pooling, the resulting feature sequence is mapped to a fixed-length feature vector by a global average pooling layer, which is used to represent the global features of the input signal.

[0094]

[0095] Among them, F k,ch To represent the ch-th channel value of the feature vector of the k-th sample, ch∈{1,…,Ch} (Ch is the number of channels in the last convolutional layer), the final F k It is a (Ch,1) dimensional vector. is the feature sequence of the ch-th channel after the last layer is deactivated, L is the encoding of the last convolutional layer, T is the length of the output sequence of the last layer, and t is the index time step.

[0096] Global features are processed through a fully connected layer and a softmax classification layer to output the class probability distribution:

[0097]

[0098]

[0099] in, For the first One sample input signal belong The probability of each category (here are the possible probabilities of each category, i.e., category 1 has a confidence level of 0.895, category 2 has a confidence level of 0.911, and so on). Let k be the true class of the k-th sample; The input of the kth sample (preprocessed, aligned, and truncated into a fixed-length vector) is the original input of the network; The first output of the fully connected layer Unnormalized score; The sum of the exponents of all categories is used to normalize the probabilities so that the sum is 1; For the first Each sample is a fully connected layer output vector of length C; This is a category index, where C is the total number of categories. and These are the weights and bias parameters of the fully connected layer, respectively.

[0100] The predicted category is the one with the highest probability:

[0101]

[0102] For the category determined by the model, take the index of the category with the highest probability (here, based on the confidence level of the determination, such as category 1 with a confidence level of 0.895 and category 2 with a confidence level of 0.911, take the category 2 with the highest confidence level of 0.911); Given an input signal Under the given conditions, the category c that has the highest probability.

[0103] The cross-entropy loss function is used as the training objective:

[0104]

[0105] in, The average cross-entropy loss is used to measure the difference between the model's predicted probabilities and the true labels; y k,c For one-hot encoding, if the true label of sample k is c * ,but Otherwise y k,c =0; N0 is the total number of samples. The network parameters are updated using the Adam optimization algorithm to minimize the loss function and improve classification accuracy.

[0106] Step S4: Assign color values ​​to the classification prediction results of the prediction model, construct an image based on the scan point positions, and obtain a color classification map of background segmentation.

[0107] Based on the weight parameters trained by the model, real-time detection samples are predicted, and the real-time acquired terahertz time-domain data is classified to obtain a category number label for each scan point. This category number is then mapped to the corresponding pseudo-color pixel value, thereby achieving two-dimensional imaging reconstruction of the sample surface, such as... Figure 3-4 As shown.

[0108] Specifically, the scan point index is k∈{1,…,N}, where k represents the k-th scan point, and the corresponding time-domain waveform input is x(k). This is achieved through a deep learning model f. θ (·) Predicted category labels:

[0109]

[0110] Among them, f θ (x(k))=argmax c P(y k =c|x(k)), which is the trained deep learning model (i.e., a one-dimensional convolutional neural network), x(k) is the temporal sequence of the k-th scan point, and c is the class index, which takes the value {0, 1, 2, ..., C-1}.

[0111] Category number According to the coordinate position of the scan point (x k ,y k Mapped to the corresponding pixel value:

[0112]

[0113] Where I(x) k ,y k ) represents the corresponding pixel value in the imaging result, ColorMap(·) is the category-to-color mapping function, and is a fixed color table, (x k ,y k ) represents the spatial coordinates of the k-th point on the scanning path, i.e., the path determined by the S-shaped path planning algorithm.

[0114] The two-dimensional imaging reconstruction method in this invention adopts existing technology without improvement, so its detailed process will not be described here.

[0115] like Figure 2 As shown, this invention also discloses a terahertz far-field imaging system based on temporal classification, including a data acquisition unit, a data preprocessing unit, a model prediction unit, and an image construction unit.

[0116] The data acquisition unit automatically scans the samples and acquires the terahertz time-domain data of the samples through the S-shaped path automatic planning algorithm. It stores the terahertz time-domain data of each scanning point and its corresponding coordinates, and transmits it to the data preprocessing unit for preprocessing.

[0117] The data preprocessing unit performs time zero-point positioning processing on the terahertz time domain data to obtain an effective time domain window, and performs marking processing on the effective time domain window to obtain effective terahertz time domain data;

[0118] The model prediction unit takes the terahertz time-domain data after time zero-point calibration and effective time window extraction as input, uses a one-dimensional convolutional neural network for supervised learning, and the output of the convolutional layer is processed by batch normalization, max pooling and random deactivation. Global features are extracted through the global average pooling layer and classification is completed.

[0119] The image construction unit performs color classification based on the classification number predicted by the model prediction unit and constructs the image based on the scan point location to obtain the sample detection image.

[0120] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, alterations, substitutions, or variations made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A method of terahertz far-field imaging based on temporal classification, characterized in that: The method comprises the following steps: Step S1, automatically scanning the sample and collecting terahertz time domain data of the sample by an S-shaped path automatic planning algorithm, and storing the terahertz time domain data of each scanning point and the coordinates thereof correspondingly; Step S2, performing time zero point positioning on each sequence data of the terahertz time domain data, extracting an effective time domain window, and obtaining effective terahertz time domain data by intercepting data of the effective time domain window in the sequence data; Step S3, performing supervised learning by using a one-dimensional convolutional neural network, taking the effective terahertz time domain data as model input, taking a classification label with the maximum probability in a classification label probability distribution as model output, constructing and training a prediction model; Step S4, performing color assignment on a classification prediction result of the prediction model, and constructing an image based on a scanning point position to obtain a color classification image of background segmentation.

2. The time-gated terahertz far-field imaging method based on time-sequential classification according to claim 1, characterized in that: In the step S1, the S-shaped path automatic planning algorithm comprises the following steps: The upper left corner of the scanning area is taken as the scanning starting point (x0, y0), and the coordinates (x i,j ,y i,j ) of each scanning point are as follows: y i,j = y0+ iΔ x i and j are the i-th row and j-th column, respectively; N x is the number of columns of the scanning area grid; Δ x is the scanning step; x0, y0are the horizontal and vertical coordinates of the starting point of the scan, respectively. denotes rounding down; mod denotes a modulo operation, k mod N x denotes k divided by N x ; k denotes an index, a row-column index in row-major order, k e {1,..., N}; N is the total number of grids, N = N x x N y , N y is the number of rows of the scan region grid size; the acquisition start time of the kth point and the end time are: t move time required to move and stabilize each point; t meas tmeasuredis the time of acquisition of the measurement; t0is the initial time.

3. The time-gated terahertz far-field imaging method based on time-sequential classification according to claim 1, characterized in that: The step S2 comprises the following steps: The zero point positioning is achieved by using a main peak detection, and a formula is as follows: t 0,k = argmax t |S k (t)| t 0,k reference time point where the main peak is located; argmax t denotes the time point t corresponding to the maximum value of the objective function; S k (t) is a sampling sequence, t is time, and k is the scan point number; From the time reference point t 0,k Forward traversing the signal, looking for the time point at which the signal first falls below a set amplitude threshold δ as the valid starting point: t a,k = max{t | t < t 0,k , | S k (t) | < δ} According to the time zero point obtained by positioning, the signal is aligned with a time reference: S' k (t) = S k (t - t a,k ) S' k (t) is the time-shifted aligned signal; extracting the effective time window [t a,k ,t b,k ] , t b,k = t a,k + Δ t , Δ t is a set fixed time window length, then: S" k (t) = S' k (t), t e [t a,k ,t b,k ] S" k (t) is the normalized, aligned, effective time domain signal.

4. The time-gated terahertz far-field imaging method based on time-sequential classification according to claim 1, characterized in that: In the step S3, the prediction model comprises the following steps: taking the terahertz time domain data after the time zero point calibration and the effective time window extraction as input, performing supervised learning by using a one-dimensional convolutional neural network, performing batch normalization, maximum pooling and random inactivation processing on a convolutional layer output, and extracting global features through a global average pooling layer and completing classification.

5. The time-gated terahertz far-field imaging method based on time-sequential classification according to claim 1, characterized in that: In the step S4, the color assignment based on the classification result comprises the following steps: The class label is obtained by the prediction model: x(k) is the time-domain sequence of the kth scanning point; Predict the category number for the model; c is the category index, taking values {0, 1, 2, …, C-1}. Classify the numbers According to the coordinate position (x k ,y k ) of the scanning point, map to the corresponding pixel value: I(x k ,y k ) is the corresponding pixel value in the imaging result; ColorMap(·) is a category-to-color mapping function; (x k ,y k ) is the spatial coordinate of the kth point on the scanning path.

6. A terahertz far-field imaging system based on temporal classification, characterized by: The method comprises the following steps: A data acquisition unit automatically scans the sample and collects terahertz time domain data of the sample by an S-shaped path automatic planning algorithm, and stores the terahertz time domain data of each scanning point and the coordinates thereof correspondingly; A data preprocessing unit performs time zero point positioning processing on the terahertz time domain data to obtain an effective time domain window, and performs marking processing on the effective time domain window to obtain effective terahertz time domain data; A model prediction unit takes the effective terahertz time domain data as input, performs supervised learning by using a one-dimensional convolutional neural network, performs batch normalization, maximum pooling and random inactivation processing on a convolutional layer output, extracts global features through a global average pooling layer, and completes classification; An image construction unit performs color classification according to a classification number of a prediction result of the model prediction unit, and constructs an image based on a scanning point position to obtain a sample detection image.

7. A computer apparatus / device / system, characterized by: The computer program comprises a memory, a processor and a computer program stored in the memory, and the processor executes the computer program to realize the steps of the method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that: The computer program / instructions stored thereon are executed by the processor to realize the steps of the method according to any one of claims 1-5.

9. A computer program product, characterised in that: The computer program / instructions are executed by the processor to realize the steps of the method according to any one of claims 1-5.

Images