Foreign matter detection method based on hyperspectral imaging, computer device and detection system

Abstract

The application discloses a foreign matter detection method based on hyperspectral imaging, which comprises the following steps: carrying out slice and wavelet transform pretreatment on a hyperspectral image of a sample to be detected, and carrying out double-path parallel detection after the pretreatment; on one hand, straight line segments in the image slice are extracted through a first detection branch, and broken line segments are merged and reconstructed; on the other hand, the sensitivity to small foreign matters is improved through a second detection branch, the network is focused on small suspicious areas, and the detection rate of small foreign matters is improved; after the double-path parallel detection, whether the corresponding foreign matter candidate frame corresponds to the same foreign matter is judged according to the intersection and union ratio, and unified threshold judgment is carried out; finally, in combination with a preset filtering rule, effective foreign matters can be accurately screened out, and the types, positions and confidence of the foreign matters are determined, and the false positive rate of the detection result is low. On this basis, the application further discloses corresponding computer equipment and a foreign matter detection system. The application is especially suitable for foreign matter detection of kelp seedlings, seaweed and other algal products with a flexible and thin structure.

Description

Technical Field

[0001] This invention relates to the field of foreign object detection technology, and in particular to a foreign object detection method, computer equipment and detection system based on hyperspectral imaging. Background Technology

[0002] Seaweed products such as kelp seedlings, laver, and wakame are soft and extremely thin, making them susceptible to contamination with foreign objects such as cotton thread, foam, fishing net lines, straw, hair, shrimp shells, and woven bags during harvesting, drying, processing, and packaging. To ensure food safety, it is necessary to identify, detect, and remove these foreign objects. Currently, using hyperspectral imaging technology to obtain hyperspectral images is one of the mainstream methods for non-destructive and rapid detection of foreign objects. Hyperspectral imaging technology is characterized by its continuous, narrow band structure (typically far more bands than multispectral imaging, and spectral resolution exceeding 10 nm), and includes two main modules: spectral analysis and image analysis. It can penetrate thin layers of biological tissue, simultaneously acquiring spatial and spectral information, and has unique advantages in detecting internal or interlayer foreign objects.

[0003] In spectral analysis, due to the flexible, thin structure of algal products, the algae are prone to layering, wrinkling, and adhesion during processing. This physical stacking can create overlapping and masking effects, covering foreign objects or even curling them up inside. Different numbers of overlapping algal layers can lead to drastic changes in spectral baseline drift, transmission peak attenuation, and noise distribution, making it difficult for existing spectral analysis methods to uniformly set preprocessing and feature selection parameters. Furthermore, fresh algae have a high water content and strong absorption and scattering of near-infrared light, causing the spectral characteristics of foreign objects to be masked by the strong absorption and scattering signals of the upper algal layers during transmission or reflection. This makes it difficult for traditional detection methods to extract sufficient discriminative information from the average spectrum, resulting in missed detections and misjudgments in these overlapping areas.

[0004] Furthermore, most existing spectral analysis methods are traditional machine learning or chemometric methods, with typical examples including partial least squares discriminant analysis (PLS-DA) and support vector machine (SVM). These are essentially shallow models that cannot automatically learn the nonlinear distortions introduced by changes in the number of algal stacks. Once the number of stacks exceeds the training range, the performance drops sharply.

[0005] In image analysis, existing methods are mainly divided into two categories: traditional image processing techniques such as edge detection (Canny, Sobel), line detection (Hough transform), and morphological operations; and deep learning techniques such as convolutional neural networks (CNN) and object detection networks (YOLO, SSD). Directly applying these two types of image analysis methods to foreign object detection in algae products presents the following technical bottlenecks: One challenge is extracting image features from linear foreign objects. Linear foreign objects such as hair strands, cotton threads, and fishing net lines appear in images as thin, curved, low-contrast linear structures, with some areas highly similar to the algal background texture. While traditional edge detection algorithms can respond to local gradient changes in the foreign object, the uneven grayscale caused by the folds and translucency of the algal surface, coupled with the local weakening of the foreign object's contrast, often results in multiple broken short line segments rather than complete continuous contours, exhibiting a significant problem of broken lines. Straight-line detection methods such as Hough transform have limited fitting ability for nonlinear curved structures and are prone to misidentifying false edges of algal fiber textures as foreign objects. Although the Line Segment Detector (LSD) can extract line segments, it also suffers from the problem of broken lines.

[0006] The consequences of broken lines are that the resulting short segments may be discarded by subsequent screening rules due to insufficient length, or they may cause positioning errors because they cannot accurately depict the complete spatial shape of the foreign object, easily leading to missed or false detections of linear foreign objects. At the same time, broken line segments and short fibers randomly distributed in the background may also be misjudged as multiple independent small foreign objects, generating a large number of false responses and causing a high false alarm rate.

[0007] Secondly, there is the challenge of detecting tiny foreign objects in images. Small objects such as tiny bubbles and broken threads are small in area, nearly isotropic, and occupy very few pixels in an image (usually less than 5×5 pixels), making them typical small targets. Existing deep learning object detection networks typically employ multi-layer downsampling (pooling or strided convolution) to reduce computational cost. This causes high-frequency detail features of small targets to rapidly decay or even disappear in deep feature maps, significantly reducing detection recall. While image slicing (such as sliding windows) can increase the spatial sampling density of small targets, it leads to a quadratic increase in computational cost, making it difficult to meet the real-time requirements of production lines. Furthermore, the attention mechanisms of general detection networks are mostly globally distributed, lacking selective focusing on tiny regions, resulting in a persistently high false negative rate for small targets.

[0008] Thirdly, there is the challenge of joint image analysis of multiple types of foreign objects. In actual production lines, linear foreign objects and tiny foreign objects differ significantly in morphology, size, and frequency domain features. Traditional image analysis methods are effective for linear structures but ineffective for small point targets; deep learning methods have some detection capability for small targets but do not fit well for slender, curved linear structures. It is evident that a single image analysis architecture cannot simultaneously cover the feature space of both types of targets, resulting in detection blind spots in existing detection methods. Summary of the Invention

[0009] The purpose of this invention is to provide a foreign object detection method based on hyperspectral imaging, which can detect both linear and minute foreign objects, improve the accuracy of foreign object detection, and reduce the false alarm rate.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A foreign object detection method based on hyperspectral imaging includes acquiring a hyperspectral image of the sample to be tested and performing image analysis processing on the hyperspectral image, including the following steps: S1. Preprocessing: The hyperspectral image is sliced ​​and each slice is subjected to wavelet transform. S2. Dual-path parallel detection: The preprocessed slice is input into the first detection branch. In the first detection branch, the line segment detector extracts the straight line segments in the slice, and the extracted broken line segments are merged by the line segment reconstruction algorithm to generate linear foreign object candidate boxes. At the same time, the original slice or the normalized slice is input into the second detection branch. In the second detection branch, the target detection network with an embedded attention module generates small foreign object candidate boxes. S3. Multi-source result fusion: Project the linear foreign object candidate set output by the first detection branch and the minute foreign object candidate set output by the second detection branch onto a unified coordinate system, and calculate the cross-union ratio (CUI) between any two candidate boxes from the first and second detection branches respectively; when the CUI is greater than the threshold, perform weighted fusion of the confidence scores of the two candidate boxes to obtain the comprehensive confidence score; when the CUI is less than the threshold, retain the original confidence scores of the two candidate boxes. S4. Classification Decision: Eliminate candidate boxes according to preset filtering rules, and determine valid foreign objects based on the remaining candidate boxes. The filtering rules include eliminating the corresponding candidate boxes when the overall confidence or the original confidence retained is less than the threshold.

[0011] Furthermore, in step S1: during the slicing process, the hyperspectral image is divided into multiple overlapping slices of fixed size, with an overlap rate of 10%-20% between the slices; during the wavelet transform process, a two-dimensional discrete wavelet transform is performed on each slice to separate the high-frequency components and the low-frequency components, and after amplifying the high-frequency components, they are reconstructed with the low-frequency components to obtain an enhanced slice.

[0012] Furthermore, in step S2, the calculation steps of the line segment reconstruction algorithm include: S211. Calculate the endpoint coordinates, direction angles, and lengths of all line segments; S212. Perform a connection judgment on any two line segments. When the spatial distance between the endpoints of the two line segments is less than the threshold T_d and the difference in direction angle is less than the threshold T_θ, the two line segments are determined to be connectable broken line segments. S213. Merge two connectable line segments into a new line segment, and determine the endpoint coordinates and direction angle of the new line segment; S214. Repeat steps S212 and S213 until all connectable line segments are merged. After the line segment reconstruction is completed, the line segments with a length exceeding the threshold L_min are marked as linear foreign object candidate boxes, and the confidence level is assigned to the linear foreign object candidate boxes according to the length, contrast and number of merging of the line segments.

[0013] Furthermore, in step S2, the processing steps of the second detection branch are as follows: S221. Input the original slice or the normalized slice into the object detection network to extract multi-layer feature maps; S222. Input the multi-layer feature map into the CBAM attention module to obtain the enhanced feature map; S223. Input the enhanced feature map into the detection head and output the candidate bounding boxes for small foreign objects and their confidence scores; The CBAM attention module includes a cascaded channel attention submodule and a spatial attention submodule. The channel attention submodule is used to adaptively assign weights to each channel of the input feature map to enhance the foreign object-related channel and suppress the background channel. The spatial attention submodule is used to generate a spatial weight map and apply the spatial weight map to the input feature map to enhance the foreign object-related location region in the input feature map.

[0014] Furthermore, the weighted fusion calculation formula in step S3 is: conf_final=α×conf_A+β×conf_B, where conf_final is the overall confidence level, α and β are weight coefficients and α is greater than β, conf_A is the confidence level of the linear foreign object candidate box output by the first detection branch, and conf_B is the confidence level of the small foreign object candidate box output by the second detection branch.

[0015] Furthermore, in step S1, the hyperspectral image is acquired by a hyperspectral camera in a full-transmission manner; in step S4, the filtering rules also include removing candidate boxes with an area smaller than a threshold and removing candidate boxes with a length smaller than a threshold.

[0016] Furthermore, it also includes spectral analysis processing of the hyperspectral data of the sample to be tested. The steps of spectral analysis processing are as follows: a. Constructing a sample dataset: Collect hyperspectral data of samples carrying foreign objects of different materials and with different numbers of overlapping layers, and perform black-and-white correction on the hyperspectral data; b. Extracting the region of interest: Selecting the region of interest from the corrected hyperspectral data and obtaining the average spectral data of the region of interest; c. Preprocessing: Perform Savitzky-Golay smoothing preprocessing on the average spectral data; d. Dataset partitioning: The average spectral data is partitioned using a sample set partitioning algorithm based on the XY joint distance, and divided into training set, validation set and test set according to a preset ratio; e. Selecting characteristic wavelengths: Based on the competitive adaptive reweighted sampling algorithm and / or the continuous projection algorithm, extract characteristic wavelengths from the preprocessed average spectral data; f. Using the average spectral data and the types of foreign materials corresponding to the spectral data in the training set, train a one-dimensional convolutional neural network to establish a one-dimensional convolutional neural network model for identifying the types of foreign materials. g. Validate and test the one-dimensional convolutional neural network model using the average spectral data and the types of foreign materials corresponding to the validation set and the test set, and obtain the final one-dimensional convolutional neural network model; h. By using the final one-dimensional convolutional neural network model to judge the hyperspectral data of the sample under test, the foreign material type can be determined.

[0017] Furthermore, a competitive adaptive reweighted sampling algorithm was used to select 72 characteristic wavelengths, which are in the range of 449-955nm; a continuous projection algorithm was used to select 7 characteristic wavelengths, namely 759.80nm, 891.94nm, 967.85nm, 638.62nm, 1015.50nm, 1013.00nm, and 395.52nm.

[0018] Another objective of this invention is to provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor is used to perform the foreign object detection method based on hyperspectral imaging as described above.

[0019] The present invention also aims to provide a foreign object detection system, comprising a dark box, a conveying device, a hyperspectral camera, at least two independent light sources, at least two calibrators, and the aforementioned computer equipment; the dark box contains at least two relatively independent illumination areas, each illumination area containing a light source with a different light source frequency and / or a different light source intensity; the conveying device is used to sequentially convey samples to the light sources in each illumination area; the hyperspectral camera is used to acquire hyperspectral images of the samples in each illumination area and transmit the hyperspectral images to the computer equipment; the calibrators are fixedly disposed within the illumination areas, and the transmittance of the calibrators in different illumination areas corresponds to foreign object-free samples with different numbers of overlapping layers.

[0020] The present invention has the following beneficial effects: 1. In terms of image analysis and processing, this application first performs slicing and wavelet transform preprocessing on the hyperspectral image to avoid the loss of small foreign objects, highlight the edge response of linear foreign objects, and suppress background noise. After preprocessing, dual-path parallel detection is performed. The first detection branch extracts straight line segments from the image slices and merges and reconstructs broken line segments. At the same time, the second detection branch improves the sensitivity to small foreign objects, enabling the network to focus on small suspicious areas and improve the detection rate of small foreign objects. After dual-path parallel detection, multi-source results are fused. Based on the intersection-union ratio of the candidate boxes of foreign objects generated by the two detection branches, it is determined whether they correspond to the same foreign object, and the comprehensive confidence score is calculated accordingly by weighted fusion or the original confidence score is retained. Finally, based on the preset filtering rules, including the unified threshold judgment of the confidence score, the effective foreign objects and their types, locations, and confidence scores can be determined. It can be seen that this method overcomes the problem that it is difficult to detect both linear and small foreign objects at the same time, and can effectively improve the accuracy of foreign object detection and reduce the false alarm rate.

[0021] 2. In terms of spectral analysis and processing, the one-dimensional convolutional neural network model constructed by this invention based on SmSG smoothing, SPA feature selection and 1D-CNN model can overcome the spectral degradation caused by the overlapping and occlusion of algae products, especially kelp seedlings, and achieve high-precision and lightweight foreign object detection. Its test accuracy reaches 98.96%, and its precision, recall and F1 score are all close to 99%. The model size is only 0.53MB, and the training time is short. It can provide a feasible technical solution for online sorting of algae products.

[0022] 3. The computer equipment and foreign object detection system provided by the present invention apply the above-mentioned foreign object detection method based on hyperspectral imaging. It also has many characteristics such as high detection accuracy, low false alarm rate, and good real-time performance. The foreign object detection system also has at least two independent illumination areas and matching light sources, which can provide sequential illumination to obtain multiple hyperspectral images of the sample under different spectral conditions. This is beneficial for distinguishing foreign objects from the background and improving detection accuracy. In addition, the system also has at least two calibrators with transmittance matching samples with different numbers of overlapping layers, which is beneficial for correcting nonlinear attenuation introduced by changes in sample layer thickness during detection and analysis. Attached Figure Description

[0023] Figure 1 This is a flowchart of the image analysis and processing of the foreign object detection method based on hyperspectral imaging according to the present invention.

[0024] Figure 2 This is a diagram of the target detection network architecture with an embedded attention module in this invention.

[0025] Figure 3 This is a flowchart of the multi-source result fusion and classification decision-making process of the present invention.

[0026] Figure 4This is an image of laver with cotton thread-like foreign matter, as presented in this invention.

[0027] Figure 5 for Figure 4 The diagram after wavelet transform processing.

[0028] Figure 6 This is a schematic diagram of line segment extraction for the first detection branch of the present invention.

[0029] Figure 7a This is an image of the broken line segment before the merging of elements in this invention.

[0030] Figure 7b This is an image of the merged broken line segment from the present invention.

[0031] Figure 8a This is a schematic diagram of the linear foreign object candidate frame before line segment reconstruction in this invention.

[0032] Figure 8b This is a schematic diagram of the linear foreign object candidate frame after the line segment reconstruction of the present invention.

[0033] Figure 9 This is a schematic diagram illustrating the construction of kelp seedling samples with foreign matter of different overlapping layers according to the present invention.

[0034] Figure 10 These are pseudo-color images of kelp seedling samples before and after noise reduction, as presented in this invention.

[0035] Figure 10 a is a pseudo-color image before the noise reduction process of this invention.

[0036] Figure 10 b is the pseudo-color image after noise reduction processing according to the present invention.

[0037] Figure 11 This is a schematic diagram of the average spectral curves of various types of samples in this invention.

[0038] Figure 11 a represents the average spectral curve of the samples corresponding to the seven types of foreign substances in this invention.

[0039] Figure 11 b represents the average spectral curve of single-layer to four-layer samples containing only kelp seedlings in this invention.

[0040] Figure 11 c represents the average spectral curves of single-layer to four-layer samples when the foreign object is cotton thread.

[0041] Figure 11 d represents the average spectral curves of single-layer to four-layer samples when the foreign material in this invention is plastic foam.

[0042] Figure 11 e represents the average spectral curve of single-layer to four-layer samples when the foreign object is fishing net line.

[0043] Figure 11 f represents the average spectral curve of single-layer to four-layer samples when the foreign object is a woven bag.

[0044] Figure 11 g represents the average spectral curve of single-layer to four-layer samples when the foreign object is hair.

[0045] Figure 11 h represents the average spectral curve of single-layer to four-layer samples when the foreign object is shrimp shell.

[0046] Figure 11 i represents the average spectral curves of single-layer to four-layer samples when the foreign material is straw.

[0047] Figure 12a This is a graph showing the change in the number of variables during the iteration process of the Competitive Adaptive Reweighted Sampling (CARS) algorithm of this invention.

[0048] Figure 12b This is a graph showing the relationship between the number of iterations and the root mean square error (RMSECV) of the competitive adaptive reweighted sampling (CARS) algorithm of this invention.

[0049] Figure 12c This is a path diagram of the regression coefficients of the Competitive Adaptive Reweighted Sampling (CARS) algorithm of this invention.

[0050] Figure 13 This is a distribution diagram of the characteristic wavelengths selected by the competitive adaptive reweighted sampling algorithm (CARS) of this invention.

[0051] Figure 14 This is a distribution map of the characteristic wavelengths selected by the continuous projection algorithm (SPA) of this invention.

[0052] Figure 15 This is a schematic diagram of the one-dimensional convolutional neural network (1D-CNN) model of the present invention.

[0053] Figure 16 This is a schematic diagram comparing the training of a one-dimensional convolutional neural network and a one-dimensional residual neural network according to the present invention.

[0054] Figure 16 a represents the accuracy change curve of the one-dimensional convolutional neural network (1D-CNN) of this invention during training.

[0055] Figure 16 b is the loss change curve of the one-dimensional convolutional neural network (1D-CNN) training process of this invention.

[0056] Figure 16 c represents the accuracy change curve of the one-dimensional residual neural network (1D-ResNet) of this invention during training.

[0057] Figure 16 d represents the loss change curve during the training process of the one-dimensional residual neural network (1D-ResNet) of this invention.

[0058] Figure 17 This is the confusion matrix of the one-dimensional convolutional neural network (1D-CNN) model using the continuous projection algorithm (SPA) selected for the feature wavelengths in this invention.

[0059] Figure 18 This is a schematic diagram of the foreign object detection system of the present invention.

[0060] Explanation of key component symbols: 1. Dark box; 2. Conveying device; 3. Hyperspectral camera; 4. Computer equipment; 5. Sample; 6. Inlet; 7. Outlet; 8. Light source; 9. Calibration material; 10a-10c. One-way mirror; I-III. Illumination area. Detailed Implementation

[0061] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0062] Example 1 This invention provides a foreign object detection method based on hyperspectral imaging, applicable to the detection of foreign objects in algae products with flexible, thin structures such as kelp seedlings, nori, and wakame. The method includes acquiring hyperspectral images of the sample to be tested and performing image analysis processing. This image analysis processing method is particularly suitable for image feature extraction, target segmentation, and automatic identification of foreign objects such as hair strands, cotton threads, fishing net lines, and small bubbles. Figure 1 As shown, image analysis and processing specifically includes the following steps: S1. Preprocessing: The hyperspectral image of the acquired sample to be tested is sliced ​​and wavelet transform is performed on each slice.

[0063] The hyperspectral image should ideally be a high-resolution color image with a resolution of at least 1920×1080. Furthermore, it is best to acquire the hyperspectral image using a hyperspectral camera in a fully transmissive manner, allowing light to penetrate the sample during acquisition. For semi-transparent samples such as kelp seedlings and laver, this method helps improve the contrast and detection rate of foreign matter spectral characteristics, reducing the difficulty of distinguishing foreign matter from the background.

[0064] During the slicing process, the SAHI strategy is adopted to divide the hyperspectral image into multiple overlapping fixed-size slices to ensure that tiny foreign objects are not lost during subsequent downsampling, while reducing the computational burden of a single processing step. In this embodiment, the image slice size is fixed at 512×512 pixels, and the overlap rate between slices is 10%-20%.

[0065] During wavelet transform processing, a two-dimensional discrete wavelet transform (DWT) is performed on each slice to separate the high-frequency components (HH, HL, LH) and the low-frequency components (LL). After amplifying the high-frequency components, the amplified high-frequency components are reconstructed with the low-frequency components to obtain the enhanced slice.

[0066] Since linear foreign objects such as hair strands, cotton threads, and fishing net lines exhibit high-frequency edge features in images, while the background texture of algae such as laver and kelp seedlings is relatively low-frequency, amplifying the high-frequency components and then reconstructing them with the low-frequency components can effectively highlight the edge response of linear foreign objects, suppress background noise, and provide clearer input for subsequent line segment detection. In this embodiment, the gain coefficient is set to 1.2-1.5. Taking laver as an example, its original image is as follows: Figure 4 As shown, the effect after wavelet transform is as follows: Figure 5 As shown.

[0067] S2. Dual-path parallel detection: The preprocessed slice is input into the first detection branch, and at the same time, the original slice or the slice after only simple normalization preprocessing is input into the second detection branch for dual-path parallel detection.

[0068] In the first detection branch, straight line segments are extracted from the image slices by a line segment detector, and the extracted broken line segments are merged by a line segment reconstruction algorithm. After line segment reconstruction, a linear foreign object candidate box is generated for each linear foreign object candidate, and the set of all linear foreign object candidate boxes is the linear foreign object candidate set.

[0069] In this embodiment, the line segment detector used is LSD, EDLine, or Hough transform. The image after line segment extraction processing is as follows: Figure 6 As shown. The computational steps of the line segment reconstruction algorithm used include: S211. Calculate the endpoint coordinates, direction angle (0°-180°), and length of all line segments.

[0070] S212. Perform a connection judgment on any two line segments. If the spatial distance between the endpoints of the two line segments is less than the threshold T_d and the directional angle difference is less than the threshold T_θ, the two line segments are determined to be connectable broken line segments; otherwise, they are determined to be unconnectable. In this embodiment, the spatial distance threshold T_d is set to 8-12 pixels, and the directional angle threshold T_θ is set to 10°-15°.

[0071] S213. Based on the determination result of step S212, merge the two connectable line segments into a new line segment, and determine the endpoint coordinates and direction angle of the new line segment. The endpoints of the new line segment are the outermost endpoints of the original two line segments, and the direction angle is recalculated.

[0072] S214. Repeat steps S212 and S213 until all connectable straight line segments are merged, thus completing the merging and reconstruction of broken line segments.

[0073] After processing in steps S211-S214, the original broken line segment (such as...) Figure 7a , 8a (As shown) are connected as follows Figure 7b , 8b The complete outline of the foreign object is shown, with the connection points of the broken line segments compared front and back. Figure 7a , 7b As indicated by the middle arrow, it can be seen that line segment reconstruction can significantly improve the recall rate of linear foreign objects and reduce pseudo-line segments generated by background textures. After the line segment reconstruction process is completed, straight line segments with a length exceeding the threshold L_min (e.g., 30 pixels) are marked as candidate boxes for linear foreign objects, and the confidence level of the candidate boxes for linear foreign objects is assigned according to the length, contrast, and number of merging of the straight line segments.

[0074] In the parallel second detection branch, a target detection network with an embedded attention module generates candidate boxes for small foreign objects. The set of all candidate boxes for small foreign objects is the candidate set for small foreign objects. By using the embedded attention module, the network's sensitivity to small foreign objects (typically with a target pixel area of ​​less than 5×5) such as small bubbles and broken wires can be effectively improved.

[0075] In this embodiment, a lightweight object detection network, such as YOLO-NAS or SSD-MobileNet, is selected, and it can be accelerated using tools such as TensorRT. The embedded attention module is the CBAM attention module. Figure 2 As shown, the CBAM attention module is embedded after the multi-layer feature map of the object detection network. That is, the processing steps of the second detection branch are mainly as follows: S221. Input the original slice or the normalized slice into the object detection network to extract the multi-layer feature map; S222. Input the multi-layer feature map into the CBAM attention module to obtain the enhanced feature map; S223. Input the enhanced feature map into the detection head and output the candidate box of the small foreign object and its confidence score.

[0076] The CBAM attention module includes cascaded channel attention submodules and spatial attention submodules. The channel attention submodule adaptively assigns weights to each channel of the input feature map to enhance foreign object-related channels and suppress irrelevant background channels. The spatial attention submodule generates a spatial weight map and applies it to the input feature map to enhance the foreign object-related location regions within the feature map.

[0077] By cascading the channel attention and spatial attention submodules, the object detection network can focus on small-scale suspicious regions, effectively avoiding the loss of small object features caused by deep downsampling. During training, data augmentation strategies such as Mosaic, random cropping, and brightness jitter are employed, ultimately outputting candidate boxes for small foreign objects along with their locations and confidence scores.

[0078] S3. Multi-source result fusion: The linear foreign object candidate set output by the first detection branch and the small foreign object candidate set output by the second detection branch are fused for decision-making. Each linear foreign object candidate box contains position, length and confidence information, and each small foreign object candidate box contains bounding box and confidence information.

[0079] like Figure 3 As shown, the linear foreign object candidate set output by the first detection branch and the minute foreign object candidate set output by the second detection branch are first projected onto a unified coordinate system, such as mapping back to the original image coordinate system. Next, the cross-union ratio (CUP) between any two candidate boxes from the first and second detection branches is calculated, and further judgment is made based on the CUP threshold.

[0080] When the intersection-union ratio (CIU) is greater than a threshold, the two corresponding candidate boxes can be determined to correspond to the same foreign object. In this case, the confidence scores of the two candidate boxes are weighted and fused to obtain a comprehensive confidence score. The formula for weighted fusion is: conf_final = α × conf_A + β × conf_B, where conf_final is the comprehensive confidence score, α and β are weight coefficients, conf_A is the confidence score of the linear foreign object candidate box output by the first detection branch, and conf_B is the confidence score of the tiny foreign object candidate box output by the second detection branch. Simultaneously, since the line segment detector is more sensitive to linear targets, it is given a higher weight; that is, α is set to be greater than β to enhance the sensitivity and recall rate of linear foreign objects such as hair strands and cotton threads, preventing them from being falsely filtered out due to low confidence scores.

[0081] When the crossover ratio is less than the threshold, it can be determined that the two candidate boxes correspond to different foreign objects. In this case, the original confidence scores of the two candidate boxes are retained.

[0082] Preferably, the cross-union ratio (IoU) threshold is set to 0.3. That is, when IoU > 0.3, weighted fusion calculation is performed, and when IoU ≤ 0.3, the original confidence scores of the two candidate boxes are retained.

[0083] S4. Classification Decision: Candidate boxes are eliminated according to preset filtering rules, and valid foreign objects are determined based on the remaining candidate boxes. The filtering rules include eliminating the corresponding candidate box when the overall confidence level or the retained original confidence level is less than a threshold, which can reduce false detections and improve the accuracy and reliability of detection. In this embodiment, the confidence level threshold is set to 0.4, that is, the corresponding candidate box is eliminated when the overall confidence level or the retained original confidence level is less than 0.4.

[0084] Building upon this, the filtering rule can further include removing candidate boxes with an area smaller than a threshold and candidate boxes with a length smaller than a threshold, in order to eliminate non-foreign object interference caused by minor wrinkles, debris, etc. on the sample surface, eliminate false detections, further improve the accuracy and reliability of foreign object detection, and reduce false alarms. Preferably, the area threshold in this filtering rule is set to 10 square pixels, and the length threshold is set to 20 pixels.

[0085] After filtering, valid foreign objects can be identified based on the remaining candidate boxes, which are the finally detected foreign objects. The final output information includes the foreign object type (linear / small), location coordinates (bounding box), and confidence level. The results can be overlaid on the original image for display or sent to the PLC to trigger the rejection device.

[0086] In summary, this method constructs a dual-path parallel detection architecture from an image analysis perspective. On one hand, it designs a line segment reconstruction algorithm based on endpoint spatial distance and orientation angle constraints to reconstruct spatially adjacent and directionally similar broken line segments into complete linear foreign object contours, effectively solving the feature fragmentation problem of linear foreign objects, restoring the true shape of linear foreign objects, and improving the recall rate and localization accuracy of linear foreign object detection. On the other hand, it embeds a CBAM attention module into a lightweight detection network to enhance the feature response of small foreign objects from both channel and spatial dimensions, solving the problem of feature loss of small targets. Finally, through multi-source result fusion and classification decision, the two detection results are overlap-determined, confidence-weighted fused, and classified and filtered under a unified coordinate system, realizing collaborative image analysis of linear and small foreign objects, achieving the goal of high accuracy, low false alarms, and real-time detection.

[0087] Experimental results show that this method improves the recall rate for both linear and minute foreign objects and significantly reduces the false negative rate. Compared to existing single detection methods, the overall false alarm rate of this method can be controlled below 3.4%. Furthermore, while ensuring high accuracy, this method also fully considers computational efficiency, meeting the real-time processing needs of production lines.

[0088] Therefore, this method can effectively identify both linear and minute foreign objects in thin, translucent algae such as laver, kelp seedlings, and wakame, meeting the online, automated, and continuous foreign object detection needs of algae product production lines. It has good versatility and industrial application value.

[0089] Example 2 Based on the above embodiment one, the invention also includes spectral analysis processing of the hyperspectral data of the sample to be tested, which includes the following steps a to h.

[0090] Step a. Construct a sample dataset: Collect hyperspectral data of samples carrying foreign objects of different materials and with different numbers of overlapping layers, and perform black and white correction on the hyperspectral data.

[0091] Taking foreign object detection in kelp seedlings as an example, this embodiment utilizes transmission hyperspectral imaging to construct a sample dataset. During construction, hyperspectral data of samples containing seven types of foreign object materials and four different overlap layers were collected. The kelp seedlings used were kept refrigerated and fresh. Before collection, all kelp seedlings underwent preliminary screening, removing individuals with obvious surface damage, attached substances, or visible foreign objects to ensure sample consistency and representativeness. To simulate impurities that may be introduced during actual production, seven typical foreign objects of different materials were selected: cotton thread (CT), plastic foam (PF), nettingtwine (NT), rice straw (RS), human hair (HH), tiny shrimp (TS), and woven bag (WS). Furthermore, pure kelp seedlings (KS) were used as a background control.

[0092] To investigate the effect of the number of overlapping layers of kelp seedlings on the detection effect of foreign objects under different overlapping conditions, such as Figure 9 As shown, this embodiment designed experimental samples with different numbers of overlapping layers. From top to bottom, the following four types of overlapping layers are shown: 1. Single-layer sample: The foreign object is placed alone above the central area of ​​the kelp seedling leaf, leaving the foreign object uncovered; 2. Double-layer sample: Based on the single-layer sample, a layer of kelp seedling is added, completely covering the kelp seedling containing the foreign object, completely encasing the foreign object between the two layers of kelp seedlings, forming a double-layer sandwich structure of kelp seedling-foreign object-kelp seedling; 3. Triple-layer sample: Based on the double-layer sample, another layer of kelp seedling is added, covering the foreign object within the three layers of kelp seedlings; 4. Quadruple-layer sample: Based on the triple-layer sample, a fourth kelp seedling is added, covering the foreign object within the four layers of kelp seedlings, to create more complex concealment conditions. For the various combinations of the above 7 types of foreign matter and 4 overlapping layers, as well as the 4 overlapping situations of pure kelp seedlings, 50 replicate samples were prepared for each situation to eliminate random errors and ensure statistical significance, and a total of 1600 samples were prepared.

[0093] Before acquiring hyperspectral data, a 30-minute warm-up is required to ensure the stability of the light source used for acquisition. The parameter settings during acquisition are as follows: integration time 1000μs, averaging times 10, hyperspectral camera frame rate 100f / s, exposure time 5ms, and lens distance from the sample surface 35-45cm.

[0094] Before formal data acquisition, reference spectra are acquired by continuously recording 100 spectra without samples as reference spectra. The software then automatically converts the reference spectra into absorbance. During formal data acquisition, the constructed sample is placed flat on the sample conveyor belt, and the detection light source beam passes perpendicularly through its center for full transmission. After the light penetrates the sample and carries information about internal foreign matter, it is received by a hyperspectral camera on the opposite side of the sample. Each sample is acquired three times at the same position with the same orientation, and the average value is taken as the raw hyperspectral data for that sample.

[0095] During the hyperspectral data acquisition process, the raw hyperspectral data will contain background noise due to the influence of dark current in the instrument detector and stray light from the outside. In order to eliminate these interferences and obtain spectral data that is closer to the optical properties of the real sample, the raw hyperspectral data needs to be corrected for black and white. The calculation formula for black and white correction is shown in the following formula (1):

[0096] In the formula, I raw For the raw hyperspectral data, I dark For the completely black spectral data (dark current noise) acquired under conditions of turned-off light source, I white The data represents the full-white spectral data (incident light intensity) collected under sample-free conditions, with T representing the corrected relative transmittance. The corrected hyperspectral data is used for subsequent modeling and analysis.

[0097] Furthermore, during hyperspectral data acquisition, factors such as vibration of the sample conveyor belt motor and unbalanced operation of mechanical transmission components can introduce periodic noise into the hyperspectral image. This type of noise manifests as regular stripes or grid-like artifacts in the spatial domain. Although its intensity is not significant, it has a fixed spatial frequency, interfering with the extraction of foreign object edge features. This is especially true when the foreign object is obscured by multiple layers of seaweed seedlings, where periodic noise may further weaken the already weak foreign object signal. To suppress this type of noise, this embodiment employs wavelet transform for spatial domain denoising of the hyperspectral image. Considering that periodic noise exhibits discrete peaks in the frequency domain, while the true sample signal (such as seaweed seedling texture and foreign object edges) is concentrated in the low-frequency region and has continuous spectrum characteristics, the multi-resolution characteristics of wavelet transform can effectively separate noise from the valid signal.

[0098] The processing first involves performing a two-dimensional discrete wavelet transform on the hyperspectral grayscale image of each band. The db4 wavelet, with its good orthogonality and tight support, is selected as the basis function, and the decomposition layer is set to four, resulting in a 4×4 subband matrix. Each layer decomposes to generate a low-frequency approximate subband (LL) and high-frequency detail subbands in three directions: horizontal (HL), vertical (LH), and diagonal (HH). Energy spectrum analysis of the decomposed subbands reveals that periodic noise is mainly concentrated in the HL subbands of the first and second layers. This is because the sample conveyor belt's motion direction is perpendicular to the hyperspectral camera's scan line, and the introduced periodic changes are particularly pronounced in the horizontal gradient, thus being concentrated in the horizontal detail subband. Based on this analysis, the db4 wavelet decomposition is performed on the spatial grayscale image of each band, and (1,2) and (2,2) (i.e., the HL subbands of the first and second layers) are removed. Then, an inverse wavelet transform is performed on the remaining coefficients to reconstruct the hyperspectral image after removing periodic noise. Figure 10 As shown in a and 10b, after denoising, the sample pseudocolor image effectively reduces the stripe artifacts caused by the vibration of the sample conveyor belt while maintaining the clarity of the foreign object edge. Therefore, the hyperspectral image processed by denoising and reconstruction can provide a cleaner data foundation for subsequent region of interest extraction and spectral analysis.

[0099] Step b. Extracting the region of interest: Select the region of interest from the corrected and denoised hyperspectral data, and obtain the average spectral data of the region of interest.

[0100] In a hyperspectral image, each pixel corresponds to a continuous spectrum. Since the contrast between foreign objects and the background is higher in infrared single-band images and is not affected by visible light color, it is easier to identify and accurately select with the naked eye. Therefore, in this embodiment, based on the differences in the spectral characteristics of the materials, the infrared single-band image with the most obvious distinction between the foreign object and the kelp seedling background is selected as the base map. Using ENVI 5.3 software, the Region of Interest (ROI) is manually selected for the hyperspectral data of the above 1600 samples.

[0101] Considering that the occlusion of kelp seedlings may prevent the complete acquisition of spectral information for the entire foreign object region, to ensure the accuracy of the extracted spectral information, 50 pixels were uniformly selected within the ROI for each type of foreign object sample, and their spectral data were extracted. During extraction, the average spectral data of all pixels within the ROI was taken as the representative spectral data for that sample. After removing bands with poor signal-to-noise ratios at both ends of the spectrum, spectral data in the 400-1000nm range were finally used for subsequent modeling and analysis. The average spectral data of all samples together constitute the average spectral feature set of kelp seedling foreign objects, which can provide a clear spectral stratification benchmark for subsequent algorithm selection.

[0102] like Figure 11As shown in Figure a, based on the average spectral data of samples corresponding to any one of the seven foreign object types (cotton thread, foam, fishing net line, woven bag, hair, shrimp shells, and straw) collected by transmission hyperspectral imaging, it can be seen that hair has the highest transmission intensity across the entire wavelength range, while foam has the lowest, and the difference between the two is significant, which can be distinguished by the intensity threshold. Straw shows a clear absorption valley around 500 nm, while cotton thread and shrimp shells have broad absorption characteristics in the 550-650 nm region. The differences in the position and depth of these characteristic peaks / valleys can be used as key indicators to distinguish the above-mentioned foreign objects.

[0103] like Figure 11 As shown in b-11i, from single-layer to four-layer samples, the transmission intensity of all samples continuously decreases as the number of kelp seedling coverage layers increases, with a rapid attenuation zone between two and three layers. This may be related to the multiplication of scattering paths after the kelp seedlings are stacked.

[0104] In single-layer samples, the spectra of various samples exhibited rich characteristic differences. Kelp seedlings showed a flat background spectrum with a transmittance intensity between 0.4 and 1.4, while cotton thread and shrimp shells showed absorption valleys around 580 nm, and straw showed an absorption valley around 500 nm. Hair and fishing net thread had high and flat spectra, while foam showed the lowest intensity. All samples showed good separability under single-layer conditions. In double-layer samples, the depth of the characteristic absorption valleys decreased and the valley bottoms widened, but the 500 nm valley of straw and the 580 nm valley of shrimp shells and cotton thread remained identifiable. In triple-layer samples, the spectra tended to be flatter, and the intensity differences between different foreign objects significantly decreased.

[0105] In the four-layer sample, most signals dropped below 0.10. Under four-layer conditions, the spectral curves of all samples tended to be uniform (intensity 0.05-0.15), and the original characteristic information was completely masked by the strong absorption and multiple scattering of the kelp seedling tissue. Any method relying on characteristic peaks or absolute intensity faced physical limits. At the same time, the standard deviation decreased significantly with the increase of the number of layers, reflecting the homogenization effect of multi-layer coverage, but it also led to a faster decrease in the signal mean and a continuous deterioration in the signal-to-noise ratio. It is worth noting that under four-layer conditions, the STD (standard deviation) of some samples (such as cotton thread and fishing net thread) slightly rebounded in the long-wavelength region, which may be related to the increased noise contribution when the signal is extremely low.

[0106] Step c. Preprocessing: Savitzky-Golay smoothing is performed on the average spectral data. Savitzky-Golay smoothing (SmSG) effectively suppresses random noise in the kelp seedling spectrum, overcomes spectral fluctuations caused by physical differences such as uneven sample thickness and surface wrinkles, and preserves local spectral features while smoothing, avoiding the loss of key information due to over-smoothing. Furthermore, this processing can partially reduce baseline drift caused by scattering and absorption by kelp seedlings, improve the signal-to-noise ratio of the spectral signal, and provide cleaner and more stable spectral data for subsequent modeling and analysis.

[0107] Step d. Data set partitioning: The average spectral data, namely the average spectral feature set of the above-mentioned kelp seedling foreign matter, is partitioned using the sample set partitioning based on joint XY distances (SPXY) algorithm, and divided into training set, validation set and test set according to a preset ratio.

[0108] The SPXY algorithm is an extension of the Kennard-Stone (KS) algorithm. When calculating the distance between samples, it considers both spectral features (X) and labels (Y), normalizes them, and then performs a weighted sum. This ensures that the selected training set is uniformly distributed in the spectral space while also considering the representativeness of the label space. It is particularly suitable for classification tasks where the class distribution is uneven or there is a non-linear relationship between the spectrum and the labels. In this embodiment, the average spectral data is divided into a training set, a validation set, and a test set in a 6:2:2 ratio, i.e., 60% is used for training, 20% for validation, and 20% for testing.

[0109] Step e. Selecting Feature Wavelengths: Extract feature wavelengths from the preprocessed average spectral data based on the Competitive Adaptive Reweighted Sampling (CARS) algorithm and / or the Successive Projection Algorithm (SPA).

[0110] Hyperspectral data is characterized by continuous bands and high dimensionality. Strong collinearity often exists between adjacent wavelengths, resulting in a large amount of redundant information. This not only increases computational complexity but also degrades model performance when the number of training samples is limited. This embodiment implements CARS and / or SPA in Python, which can filter out a subset of feature wavelengths containing key information and good separability from the full-band spectrum (400-1100nm). This effectively reduces data dimensionality and improves model accuracy and robustness, and is a crucial step in extracting effective information and achieving accurate foreign object identification.

[0111] Among them, the CARS algorithm, which combines Monte Carlo sampling and exponential decay function, can adaptively select wavelengths that contribute greatly to the target variable and has good variable selection stability; while the SPA algorithm selects the wavelength combination with the minimum collinearity and the lowest information redundancy through vector projection analysis, which can effectively reduce multicollinearity between wavelengths and is suitable for extracting representative wavelengths from high-dimensional spectra.

[0112] Specifically, the process of selecting feature wavelengths using the Competitive Adaptive Reweighted Sampling (CARS) algorithm is as follows: Figures 12a-12c As shown. In this embodiment, the Monte Carlo sampling is set to 500 times, and 5-fold cross-validation is used. The result with the smallest root mean square error (RMSECV) of cross-validation is selected as the optimal feature subset. Since each sampling is performed randomly, the CARS selection results have some fluctuations. Therefore, multiple runs and comparisons are needed to determine the optimal feature wavelength variables. Figure 12a It can be seen that during the first 20 sampling processes, a large number of wavelength variables were quickly eliminated, and then the elimination rate gradually slowed down, and the number of variables tended to stabilize. This trend reflects the screening characteristics of the CARS algorithm in the "coarse selection" and "fine selection" stages. Figure 12b The study demonstrates the variation of the root mean square error of cross-validation (RMSECV) with the number of samplings. Before the RMSECV reaches its minimum, the error generally shows a decreasing trend, indicating that a large number of wavelength variables unrelated to the kelp seedlings and their foreign objects are effectively eliminated, and the predictive performance of the model gradually improves. Figure 12c The regression coefficients of each wavelength variable were shown to change when RMSECV reached its minimum value (18th sampling). Ultimately, when the RMSECV value was 1.6354, 72 characteristic wavelength variables were selected, accounting for only 15% of the total original spectral variables. These selected characteristic wavelengths provided an important variable basis for the subsequent construction of an infrared spectroscopy-based model for detecting foreign objects in kelp seedling layers.

[0113] like Figure 13 As shown, the 72 characteristic wavelengths selected by the CARS algorithm are in the range of 449-955nm, spanning the visible light (400-700nm) and near-infrared regions (700-955nm). These bands are concentrated in several key spectral regions, which can comprehensively characterize the spectral differences between kelp seedlings and seven types of foreign substances.

[0114] The characteristic wavelength results selected by the SPA algorithm are as follows: Figure 14As shown in the figure, the locations of the feature wavelengths selected by the algorithm are marked with red dots. The figure reveals that SPA selected the following seven feature wavelengths: 759.80nm, 891.94nm, 967.85nm, 638.62nm, 1015.50nm, 1013.00nm, and 395.52nm. From the wavenumber distribution, the selected feature wavelengths cover the key spectral range from ultraviolet-visible to near-infrared.

[0115] The 395.52nm wavelength lies at the near-ultraviolet-visible boundary, and its absorption is associated with aromatic amino acids, carotenoids, and chlorophyll. Kelp seedlings, rich in chlorophyll and carotenoids, show significant absorption in this band, while synthetic materials such as cotton thread, foam, fishing net lines, and woven bags exhibit primarily scattering and weak absorption. Hair, containing keratin, and shrimp shells, containing astaxanthin, also produce characteristic responses. Therefore, this wavelength is useful for distinguishing between biomass and synthetic materials.

[0116] 638.62nm is in the red region of visible light, close to the absorption peak of chlorophyll (660nm). Kelp seedlings absorb strongly in this band, while white or light-colored foreign objects (such as foam, cotton thread, fishing net line) reflect more strongly and absorb less. At the same time, astaxanthin in shrimp shells also has some absorption. Therefore, this band can highlight the kelp seedling tissue.

[0117] 891.94nm belongs to the near-infrared band, mainly corresponding to the second harmonic vibration of CH bonds. Foreign matter such as cotton thread (cellulose), straw (lignocellulose), and woven bags (polyolefins) are rich in CH bonds. Although kelp seedlings also contain polysaccharides and proteins, their absorption intensities differ. This band can be used to distinguish between synthetic materials and natural tissues.

[0118] The 1013.00nm and 1015.50nm wavelengths are both located in the near-infrared region, belonging to the second-order harmonic region of the CH bond, and are sensitive to components such as fats, cellulose, and plastics. Shrimp shells (fat), straw (cellulose), woven bags and fishing net lines (polyolefins), and hair (keratin) all exhibit characteristic absorption in this region, which differs from the polysaccharide absorption of kelp seedlings, thus effectively distinguishing various foreign substances.

[0119] It is evident that the seven characteristic wavelengths selected by SPA highlight the spectral differences in composition between different biomass (kelp seedlings, shrimp shells, straw, and hair) and synthetic materials (such as cotton thread, foam, fishing net line, and woven bags). Even under the layered coverage of kelp seedlings, it can still capture the characteristic signals of small foreign objects in the gaps, while achieving data dimensionality reduction, providing key characteristic variables for the subsequent construction of a lightweight and high-precision foreign object detection model.

[0120] Step f. Using the average spectral data and the corresponding foreign material types in the training set, train a one-dimensional convolutional neural network (1D-CNN) to establish a one-dimensional convolutional neural network model for identifying foreign material types. 1D-CNN extracts local spectral features through one-dimensional convolution operations. For the input spectral sequence x, the output of the j-th feature map of the k-th convolutional layer at position t is shown in equation (2):

[0121] in, For convolution kernel weights, Here, S is the kernel size, F is the number of input feature maps, and f(·) is the ReLU activation function. The network alternately stacks convolutional layers, batch normalization layers, and dropout layers. The first two convolutional layers use a stride of 2 for downsampling, while the last two convolutional layers maintain their length. Finally, a global average pooling layer and a fully connected layer output the classification probability. A schematic diagram of the model structure is shown below. Figure 15 As shown in the figure, k represents the kernel size, s represents the stride, p represents padding, BN represents batch normalization, ReLU represents corrected linear unit, p (in Dropout) represents the dropout rate, and FC represents a fully connected layer.

[0122] Step g. Validate and test the one-dimensional convolutional neural network model using the average spectral data and the corresponding foreign material types in the validation and test sets, and obtain the final one-dimensional convolutional neural network model.

[0123] Step h. Use the final one-dimensional convolutional neural network model to judge the hyperspectral data of the sample to obtain the foreign material type judgment result.

[0124] To comprehensively evaluate model performance, this invention uses multiple quantitative indicators to evaluate independent test sets. For multi-class foreign object classification, the following indicators are calculated: Accuracy, which represents the proportion of correctly classified samples in all test samples; Precision, which measures the proportion of samples predicted as belonging to a certain class that actually belong to that class, reflecting the model's ability to avoid false positives; Recall, which represents the proportion of correctly identified real samples of a certain class, reflecting the model's sensitivity to detecting various types of foreign objects; and F1 score, which serves as the harmonic mean of precision and recall, providing a balanced evaluation, especially suitable for datasets that may have class imbalance. The calculation formulas for each indicator are shown in the following equations (3)-(6):

[0125]

[0126]

[0127]

[0128] Among them, TP indicates true positive, TN indicates true negative, FP indicates false positive, and FN indicates false negative.

[0129] All metrics were calculated individually for each category, and macro averages were reported to evaluate the model's overall performance across all categories. All modeling and evaluation processes were performed in the same computing environment to ensure comparability and reproducibility of results. Hardware configuration included: Intel(R) Core(TM) i7-14700KF CPU @ 3.40 GHz, NVIDIA GeForce RTX4060 GPU (8 GB VRAM), 64 GB RAM, and a 64-bit Windows 11 Professional operating system. Traditional machine learning models were implemented using Python 3.11 and libraries such as scikit-learn (1.0.2) and NumPy (1.21.0).

[0130] All deep learning models were built using the PyTorch (1.12.0) framework and trained on a GPU with accelerated performance using the AdamW optimizer. The initial learning rate was set to 1×10⁻⁶. -3 The model employs a ReduceLROnPlateau learning rate scheduling strategy. Hyperparameters (such as the number of convolutional kernels and the Dropout rate) are optimized using a grid search on the validation set. Early stopping (with a patience value of 30 epochs) is used during model training to prevent overfitting. These uniform settings ensure that performance differences between models stem solely from the network architecture itself, rather than the training strategy. All traditional models are implemented using Python's scikit-learn library, and five-fold cross-validation is performed on the training set to optimize hyperparameters.

[0131] First, the models were evaluated against different preprocessing techniques and dataset partitioning algorithms. This evaluation was based on a random sampling algorithm (RS) and a sample set partitioning algorithm based on the XY joint distance (SPXY). Spectral data processed by different preprocessing techniques and the original data were input into different classification models for comparison. The preprocessing techniques used included Multiplicative Scatter Correction (MSC), Standard Normal Variation (SNV), Gaussian Smoothing (SmGauss), Savitzky-Golay First Derivative (SG1D), Savitzky-Golay Second Derivative (SG2D), and the preferred Savitzky-Golay Smoothing (SmSG) of this invention. The corresponding classification models included Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), Backpropagation Neural Network (BPNN), and One-Dimensional Convolutional Neural Network (1D-CNN). The evaluation results are shown in Table 1 below: Table 1: Classification accuracy evaluation of full-band LDA, SVM, BPNN and 1D-CNN models on the test set under different preprocessing techniques.

[0132] As can be seen from the comparison, SPXY partitioning is significantly better than RS partitioning in terms of dataset splitting. Almost all models have higher accuracy under SPXY, which shows that SPXY can better guarantee the representativeness of the training set and the test set.

[0133] Regarding preprocessing techniques, Savitzky-Golay smoothing (SmSG) achieved optimal or near-optimal classification results in the vast majority of classifiers, demonstrating good versatility and stability. In Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), and One-Dimensional Convolutional Neural Network (1D-CNN) models, SmSG preprocessing achieved the highest classification accuracy. When using SPXY partitioning, the accuracy of SVM combined with SmSG was 97.71%, and the accuracy of 1D-CNN combined with SmSG was 81.25%. Overall, SmSG is the most robust preprocessing method. Other preprocessing methods performed reasonably well on some models, but were generally less stable than SmSG. Furthermore, second derivative (SG2D) processing caused a significant performance drop in all classifiers; the accuracy of SVM, BPNN, and 1D-CNN all fell below 70%, indicating that the second derivative excessively amplified spectral noise and lost effective feature information, making it unsuitable for the sample dataset of this invention. This invention uses SmSG preprocessed spectral data as input features in model construction and analysis, and combines it with SPXY to partition the dataset, which can improve the generalization of the model and the reproducibility of the results, and achieve better classification performance.

[0134] Secondly, the feature wavelength selection algorithm and classification model selection were evaluated. This evaluation used three feature selection algorithms: Competitive Adaptive Reweighted Sampling (CARS), Continuous Projection (SPA), and their combination. The algorithms were then compared with five classification models: Linear Discriminant Analysis (LDA), Support Vector Machine (SVM), Backpropagation Neural Network (BPNN), One-Dimensional Convolutional Neural Network (1D-CNN), and One-Dimensional Residual Neural Network (1D-ResNet). Among the five classification models, the One-Dimensional Residual Neural Network (1D-ResNet) model is constructed by introducing residual connections on top of 1D-CNN. The basic form of the residual block is defined as shown in equation (7):

[0135] In the formula, x is the input, F is the residual mapping to be learned (containing two convolutional layers), and W... i These are the weights for the convolutional layers. Residual connections allow gradients to be directly backpropagated, making it possible to train deeper networks. In this 1D-ResNet model, each residual block contains two sets of convolutional layers, batch normalization layers, and ReLU activation functions. The downsampling stride can be set as needed, with a uniform downsampling stride of 2, enabled only when using residual blocks for spatial dimensionality reduction. The stride of the remaining convolutional layers (conv3 and conv4) remains at 1.

[0136] The evaluation results are shown in Table 2 below: Table 2: Performance, size, and running time of different models using different wavelength selection algorithms

[0137] As can be seen, among the five classification models, 1D-CNN, combined with SPA feature selection, achieved the best test performance, with an accuracy of 98.96%, and precision, recall, and F1 score all approaching 99%, comprehensively outperforming other models. In comparison, SVM achieved an accuracy of 97.71% with SPA features, which, while also excellent, was slightly lower by 1.25 percentage points; 1D-ResNet only reached 96.88% with the same features, and its model size was larger (0.76MB) and training time was longer (9.70 seconds); LDA is sensitive to dimensionality reduction, and its accuracy with SPA features was only 65.83%, far lower than the results across all bands; BPNN had the weakest overall performance, with a highest accuracy of only 73.33%. Furthermore, the 1D-CNN model size remained constant at 0.53MB, and its training time was only 6.79 seconds, achieving the best balance between accuracy and efficiency. Therefore, 1D-CNN is the model with the best overall performance.

[0138] Among feature selection algorithms, SPA (7-dimensional) stands out, achieving an accuracy of 98.96% on 1D-CNN and an excellent result of 97.71% on SVM. Simultaneously, its model size is extremely small (SVM only 0.114MB, LDA only 0.0025MB), and its training time is short, proving that SPA can effectively extract the most discriminative minimal features. CARS (72-dimensional) achieves accuracies of 94.17% and 97.29% on SVM and 1D-CNN, respectively, outperforming the original full-band model but still lower than SPA. CARS-SPA (19-dimensional) achieves an accuracy of 98.13% on 1D-CNN, with performance between CARS and SPA, serving as a compromise between accuracy and dimensionality. While the original full-band model maintains a high accuracy of 97.71% on SVM, its large model size (9.17MB), long training time (2.07 seconds), and generally inferior performance on other models compared to the dimensionality-reduced features make SPA the optimal feature selection method. In summary, the combination of 1D-CNN and SPA feature selection achieves the best results in terms of test accuracy, model size, and training time, making it the optimal solution for this spectral classification task.

[0139] like Figure 16As shown in a-16d, the accuracy and loss curves during the training process of 1D-CNN and 1D-ResNet also reveal that the 1D-CNN model based on the original full-band (480 dimensions) converges slowly. Although the validation accuracy can reach 100% in the later stages, the loss curve fluctuates drastically, with the validation loss being much higher than that of the CARS and SPA feature selection schemes. This indicates that high-dimensional noise leads to difficulties in model training, poor generalization ability, and high computational cost. The 1D-CNN model based on CARS (72 dimensions) has the fastest convergence rate, achieving 100% validation accuracy in the 67th round, with the validation loss dropping to 0.008. The training curve is smooth and stable, demonstrating the best classification performance and generalization ability. The 1D-CNN model based on CARS-SPA (19 dimensions) achieves 100% validation accuracy with a validation loss of 0.0203 and a test accuracy of 98.13%. The 1D-CNN model based on SPA (7 dimensions) has the highest validation accuracy of 99.55% with a validation loss of only 0.0144, and the loss curve is smooth with no overfitting. In terms of convergence speed, the SPA feature set performs best, with a verification accuracy of 90% in 35 rounds and a verification loss of less than 0.1 after about 40 rounds; while the original full-band converges the slowest, requiring 110 rounds to achieve a verification accuracy of 90%, and the loss curve fluctuates wildly and has poor stability.

[0140] For the 1D-ResNet model, the validation accuracy based on the original full-band model consistently remained below 82.5%, with multiple sharp drops during training. The validation loss frequently spiked above 1.5. Although the training loss could be reduced to 0.07, the validation loss remained high (0.46), indicating severe overfitting and highly unstable curves. The CARS (72-dimensional) model maintained a stable validation accuracy of 88%-89%, with the validation loss decreasing to approximately 0.41. Convergence was smooth, and the training accuracy approached 100%, with a training loss of only 0.024, indicating relatively controllable overfitting. The CARS-SPA (19-dimensional) model achieved a validation accuracy of approximately 88% and a validation loss of approximately 0.40, with a smooth curve, but the accuracy improvement in the mid-to-late stages was slightly slower than with CARS. The SPA (7-dimensional) model achieved a maximum validation accuracy of approximately 88% and a validation loss of approximately 0.48, while the training accuracy was only 97.7%, with a training loss of 0.063, indicating slight underfitting. In terms of convergence speed, CARS performed the best, with a validation accuracy of about 80% after about 40 rounds and stabilizing above 85% after 80 rounds, while the validation loss decreased smoothly in sync. The original full-band was the worst, with slow and drastic accuracy improvement and multiple abnormal rebounds in loss, making it unusable.

[0141] As can be seen, 1D-CNN can achieve a validation accuracy of 98%-100% under the same features, with a validation loss as low as 0.01-0.02; while 1D-ResNet's highest validation accuracy is only 89%, and its validation loss is about 30 times higher. The residual structure did not play an advantage in this spectral classification task; instead, the high model complexity led to optimization difficulties and severe overfitting.

[0142] like Figure 17 As shown, the final one-dimensional convolutional neural network model obtained by using 1D-CNN and SPA feature selection in this application achieved excellent classification performance on samples containing 7 types of foreign objects and pure kelp seedlings. Among them, pure kelp seedlings (KS), woven bags (WS), hair (HH), and shrimp shells (TS) were completely correct (0 errors); straw (RS) had 57 correct classifications and 1 misclassified as cotton thread (CT); cotton thread (CT) had 72 correct classifications and 1 misclassified as straw (RS); foam (PF) had 49 correct classifications and 2 misclassified as straw (RS); and fishing net line (NT) had 86 correct classifications and 1 misclassified as woven bag (WS). The overall accuracy reached 98.96% (475 / 480), with only 5 samples misclassified, and the errors were mainly concentrated among straw (RS), cotton thread (CT), foam (PF), and fishing net line (NT). The results fully demonstrate the high precision and strong generalization ability of the combination of SPA feature selection and 1D-CNN in the spectral analysis processing for foreign matter detection in algal products.

[0143] Example 3 Based on the above embodiment one or embodiment two, the present invention also provides a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor is used to perform the foreign object detection method based on hyperspectral imaging as described above.

[0144] Thanks to the aforementioned foreign object detection methods, this computer equipment can accurately identify the type of foreign object (linear foreign object, micro foreign object) through image analysis and processing when handling foreign object detection tasks for flexible thin algae products such as kelp seedlings, nori, and wakame. It can also accurately identify the type of foreign object material (such as shrimp shells, straw, cotton thread, etc.) through spectral analysis and processing, thus accurately achieving dual determination of the foreign object's morphology and composition.

[0145] This computer equipment can be deployed at the production site of algae products and process the hyperspectral image data of the products in real time. It can accurately detect foreign objects and output the type, location and confidence level of the foreign objects, providing a reliable decision-making basis for automated rejection devices.

[0146] Example 4 Based on the above embodiment three, the present invention also provides a foreign object detection system, such as... Figure 18As shown, the system includes a dark box 1, a conveying device 2, a hyperspectral camera 3, at least two independent light sources 8, at least two calibration objects 9, and an industrial computer.

[0147] The dark chamber 1 is equipped with an outlet 7 and an inlet 6 for the sample 5 to enter and exit the internal cavity. The interior of the dark chamber 1 is physically divided into at least two relatively independent lighting areas by a partition structure. Each lighting area is equipped with a light source 8 of different frequencies and / or different intensities. A conveying device 2 is used to sequentially convey the sample 5 to each lighting area and to position the sample 5 directly above the corresponding light source 8. The conveying direction of the conveying device 2 is consistent with the arrangement direction of the lighting areas. The conveying device 2 can be an existing device such as a belt conveyor with a glass plate or an electrically controlled guide rail. During conveying, the sample 5 to be tested is placed on the glass plate. The specific structure and control method of the conveying device 2 are existing technologies, and can be found in the invention patent with publication number CN114577718A, which provides a device for detecting the oil content of tobacco leaves using laser-induced breakdown spectroscopy and hyperspectral imaging. These details will not be elaborated here. A hyperspectral camera 3 is used to acquire hyperspectral images of sample 5 in various illumination areas and transmit the hyperspectral images to computer device 4. The hyperspectral camera 3 can be positioned above each illumination area to acquire images directly above sample 5, or it can be used to guide light through a single-sided mirror. The industrial computer is the aforementioned computer device 4.

[0148] On the one hand, thanks to the aforementioned foreign object detection method and computer equipment 4, this detection system is suitable for foreign object detection tasks in flexible, thin-walled algae products such as kelp seedlings, nori, and wakame, and has high detection accuracy. On the other hand, the system supports full-transmission acquisition of hyperspectral images, effectively penetrating the flexible, thin structure of algae products during detection to detect foreign objects covered or encased within the sample 5, reducing surface reflection interference and improving the spectral contrast between foreign objects and the background, thus increasing the applicability of foreign object detection for algae products. Furthermore, by utilizing independently set illumination areas and matching light sources 8 within the system, sequential illumination can be provided to acquire multiple hyperspectral images of the sample 5 under different spectral conditions, which also helps distinguish foreign objects from the background, thereby improving detection accuracy.

[0149] In this embodiment, the hyperspectral camera 3 uses the ATH1010 product from Optoelectronics, employing a line scanning method with a spectral range of 400-1100nm and a spectral resolution of 1.5nm, acquiring a total of 480 bands. The light source 8 is a 50W halogen lamp (PHILIPS R7S, Netherlands). The darkroom 1 contains three illumination areas: illumination area I, illumination area II, and illumination area III. Each illumination area contains one light source 8 and one one-way mirror. The first light source 8 for illumination area I, the second light source 8 for illumination area II, and the third light source 8 for illumination area III focus on different spectral ranges. Through differentiated spectral configuration, targeted acquisition of spectral information for each band can be achieved. Each one-way mirror has the characteristics of total internal reflection and reverse transmission of light, and is positioned vertically to the light source 8 to guide the light passing through the sample 5.

[0150] In use, sample 5 can be sequentially transported to three locations—illumination area I, illumination area II, and illumination area III—using the conveying device 2 for data acquisition. When sample 5 moves to illumination area I, the first light source 8 is activated, and the light transmitted through sample 5 is guided to the hyperspectral camera 3 on the side of the dark box 1 by the positive total internal reflection of one-way mirror 10a and the transmission effect of one-way mirrors 10b and 10c, completing the first set of data acquisition. When sample 5 moves to illumination area II, the second light source 8 is activated, and the light transmitted through sample 5 is guided to the hyperspectral camera 3 by the positive total internal reflection of one-way mirror 10b and the transmission effect of one-way mirror 10c. Similarly, when sample 5 moves to illumination area III, the third light source 8 is activated, and the light is guided only by the positive total internal reflection of one-way mirror 10c.

[0151] During the acquisition process, the camera is simultaneously triggered to record images, and each image can be used for manual visual inspection. The acquired hyperspectral images are processed by removing bands with poor signal-to-noise ratios at both ends of the spectrum and performing black-and-white correction before being used for subsequent modeling and analysis.

[0152] In addition, the system includes at least two calibrators 9, which are fixedly positioned within the illumination area and located outside the horizontal projection area of ​​the conveyor 2's bearing surface. These calibrators 9 are acquired through at least one reflective optical element forming a common field of view with the hyperspectral camera 3. The calibrators 9 are physically separated from the sample 5 to be tested, preventing the overlapping or slippage of the flexible kelp sample 5 from covering the surface of the calibrators 9.

[0153] The absorption characteristics of each calibrator 9 (such as polysaccharides and proteins) are consistent with the spectral domain of the light source 8 in the corresponding illumination area. This can serve as a standard for spectral reflectance and a tool for verifying the consistency of light intensity. It can be used to correct subtle differences between the three light sources 8 caused by aging, temperature fluctuations, or uneven spatial distribution, ensuring the comparability of the spectral responses of the same foreign object under different illumination areas. Moreover, the characteristic absorption peaks of each calibrator 9 are located precisely within the segmented wavelength ranges of each light source 8, making each calibrator 9 most sensitive to its corresponding light source 8 and less sensitive to other light sources 8, further enhancing its regional specificity.

[0154] Furthermore, the transmittance of the calibrator 9 in different lighting areas corresponds to the foreign matter-free samples 5 with different numbers of overlapping layers. For example, the transmittance of one calibrator 9 corresponds to that of a single-layer kelp seedling sample 5, and the transmittance of another calibrator 9 corresponds to that of a three-layer stacked kelp seedling sample 5.

[0155] Since the absorption and scattering intensity of light increases non-linearly with the increase of the number of kelp seedling layers, by changing the ratio of solid powder raw materials (polysaccharide plus protein), tablet thickness or density of the three calibrators 9, their transmittance can be changed accordingly, so that their respective absorbance curves can accurately correspond to the pure kelp of the corresponding number of layers. The same applies to other algae samples 5.

[0156] By designing calibrators 9 with transmittance matching that of foreign matter-free samples 5 with different overlapping layers, the system can pre-establish reference responses for different degrees of obscuration before actual detection or during detection intervals. Based on this, the spectrum of sample 5 can be normalized to a reference framework of "equivalent layer number," significantly improving the accuracy of foreign matter signal extraction and enabling dynamic, real-time, and quantitative spectral correction for different layer thicknesses, thus correcting the nonlinear attenuation introduced by variations in the layer thickness of sample 5.

[0157] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art should understand that various changes in form and detail to the invention without departing from the spirit and scope of the invention as defined in the appended claims are within the scope of protection of the invention.

Claims

1. A foreign object detection method based on hyperspectral imaging, comprising acquiring a hyperspectral image of the sample to be tested and performing image analysis processing on the hyperspectral image, characterized in that, Includes the following steps: S1. Preprocessing: The hyperspectral image is sliced ​​and each slice is subjected to wavelet transform. S2. Dual-path parallel detection: The preprocessed slice is input into the first detection branch. In the first detection branch, a line segment detector extracts straight line segments from the slice, and the extracted broken line segments are merged by a line segment reconstruction algorithm to generate linear foreign object candidate boxes. The set of all linear foreign object candidate boxes is called the linear foreign object candidate set. At the same time, the original slice or the normalized slice is input into the second detection branch. In the second detection branch, a target detection network with an embedded attention module generates tiny foreign object candidate boxes. The set of all tiny foreign object candidate boxes is called the tiny foreign object candidate set. S3. Multi-source result fusion: Project the linear foreign object candidate set output by the first detection branch and the small foreign object candidate set output by the second detection branch onto a unified coordinate system, and calculate the cross-union ratio between any two candidate boxes from the first detection branch and the second detection branch respectively. When the crossover ratio (CUR) is greater than the threshold, the confidence scores of the two candidate boxes are weighted and fused to obtain the overall confidence score; when the CUR is less than the threshold, the original confidence scores of the two candidate boxes are retained. S4. Classification Decision: Eliminate candidate boxes according to preset filtering rules, and determine valid foreign objects based on the remaining candidate boxes. The filtering rules include eliminating the corresponding candidate boxes when the overall confidence or the original confidence retained is less than the threshold.

2. The foreign object detection method based on hyperspectral imaging as described in claim 1, characterized in that, In step S1: During the slicing process, the hyperspectral image is divided into multiple overlapping slices of fixed size, with an overlap rate of 10%-20%. During the wavelet transform process, a two-dimensional discrete wavelet transform is performed on each slice to separate the high-frequency and low-frequency components. After amplifying the high-frequency components, they are reconstructed with the low-frequency components to obtain an enhanced slice.

3. The foreign object detection method based on hyperspectral imaging as described in claim 1, characterized in that, In step S2: The calculation steps of the line segment reconstruction algorithm include: S211. Calculate the endpoint coordinates, direction angles, and lengths of all line segments; S212. Perform a connection judgment on any two line segments. When the spatial distance between the endpoints of the two line segments is less than the threshold T_d and the difference in direction angle is less than the threshold T_θ, the two line segments are determined to be connectable broken line segments. S213. Merge two connectable line segments into a new line segment, and determine the endpoint coordinates and direction angle of the new line segment; S214. Repeat steps S212 and S213 until all connectable line segments are merged. After the line segment reconstruction is completed, the line segments with a length exceeding the threshold L_min are marked as linear foreign object candidate boxes, and the confidence level is assigned to the linear foreign object candidate boxes according to the length, contrast and number of merging of the line segments.

4. The foreign object detection method based on hyperspectral imaging as described in claim 1, characterized in that, In step S2, the processing steps of the second detection branch are as follows: S221. Input the original slice or the normalized slice into the object detection network to extract multi-layer feature maps; S222. Input the multi-layer feature map into the CBAM attention module to obtain the enhanced feature map; S223. Input the enhanced feature map into the detection head and output the candidate bounding boxes for small foreign objects and their confidence scores; The CBAM attention module includes a cascaded channel attention submodule and a spatial attention submodule. The channel attention submodule is used to adaptively assign weights to each channel of the input feature map to enhance the foreign object-related channel and suppress the background channel. The spatial attention submodule is used to generate a spatial weight map and apply the spatial weight map to the input feature map to enhance the foreign object-related location region in the input feature map.

5. The foreign object detection method based on hyperspectral imaging as described in claim 1, characterized in that, The weighted fusion calculation formula in step S3 is: conf_final = α × conf_A + β × conf_B, where conf_final is the overall confidence level, α and β are weighting coefficients and α is greater than β, conf_A is the confidence level of the linear foreign object candidate box output by the first detection branch, and conf_B is the confidence level of the small foreign object candidate box output by the second detection branch.

6. The foreign object detection method based on hyperspectral imaging as described in claim 1, characterized in that: In step S1, the hyperspectral image is acquired by a hyperspectral camera in a fully transmissive manner; In step S4, the filtering rules also include removing candidate boxes with an area smaller than a threshold and removing candidate boxes with a length smaller than a threshold.

7. The foreign object detection method based on hyperspectral imaging as described in claim 1, characterized in that, It also includes spectral analysis processing of the hyperspectral data of the sample to be tested. The steps of spectral analysis processing are as follows: a. Constructing a sample dataset: Collect hyperspectral data of samples carrying foreign objects of different materials and with different numbers of overlapping layers, and perform black-and-white correction on the hyperspectral data; b. Extracting the region of interest: Selecting the region of interest from the corrected hyperspectral data and obtaining the average spectral data of the region of interest; c. Preprocessing: Perform Savitzky-Golay smoothing preprocessing on the average spectral data; d. Dataset partitioning: The average spectral data is partitioned using a sample set partitioning algorithm based on the XY joint distance, and divided into training set, validation set and test set according to a preset ratio; e. Selecting characteristic wavelengths: Based on the competitive adaptive reweighted sampling algorithm and / or the continuous projection algorithm, extract characteristic wavelengths from the preprocessed average spectral data; f. Using the average spectral data and the types of foreign materials corresponding to the spectral data in the training set, train a one-dimensional convolutional neural network to establish a one-dimensional convolutional neural network model for identifying the types of foreign materials. g. Validate and test the one-dimensional convolutional neural network model using the average spectral data and the types of foreign materials corresponding to the validation set and the test set, and obtain the final one-dimensional convolutional neural network model; h. By using the final one-dimensional convolutional neural network model to judge the hyperspectral data of the sample under test, the foreign material type can be determined.

8. The foreign object detection method based on hyperspectral imaging as described in claim 7, characterized in that, In step e: A competitive adaptive reweighted sampling algorithm was used to select 72 characteristic wavelengths, which are in the range of 449-955nm. A continuous projection algorithm was used to select 7 characteristic wavelengths, namely 759.80nm, 891.94nm, 967.85nm, 638.62nm, 1015.50nm, 1013.00nm, and 395.52nm.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the computer program is executed by the processor, the processor is used to perform the foreign object detection method based on hyperspectral imaging as described in any one of claims 1-8.

10. A foreign object detection system, characterized in that: It includes a darkroom, a transmission device, a hyperspectral camera, at least two independent light sources, at least two calibrators, and the computer equipment as described in claim 9; The dark chamber is divided into at least two relatively independent lighting areas, each of which is equipped with a light source with a different light source frequency and / or a different light source intensity; the conveying device is used to sequentially convey the sample to the light source in each lighting area; the hyperspectral camera is used to acquire hyperspectral images of the sample in each lighting area and transmit the hyperspectral images to the computer device; the calibrator is fixedly placed in the lighting area, and the transmittance of the calibrator in different lighting areas corresponds to foreign matter-free samples with different numbers of overlapping layers.

Images