Nondestructive testing method for identifying fish parasitic infection and system

Abstract

The present invention relates to the field of nondestructive testing. Disclosed is a nondestructive testing method for identifying fish parasitic infection. The average raw spectrum of fish belly is collected by means of visible light / near-infrared hyperspectral imaging technology, noise reduction or removal and feature extraction are carried out on the average raw spectrum, and the extracted features are added into a pre-established model for recognition, to obtain a recognition result. In the present invention, by using topmouth gudgeon for verification, it is found that the testing accuracy of the fish belly position is higher, with fewer image pixels collected, lower data analysis workload, and faster computation speed. Moreover, the present invention also provides a system for implementing the method.

Description

Non-destructive detection methods and systems for identifying parasitic infections in fish Technical Field

[0001] This invention relates to the field of nondestructive testing technology, specifically to a nondestructive testing method and system for identifying parasitic infections in fish. Background Technology

[0002] Fish, prized for its delicious taste and rich nutritional value, is a popular food enjoyed by consumers worldwide. However, the presence of parasites in fish severely impacts the fish industry. Firstly, countries refuse to import fish infected with parasites, resulting in significant economic losses. Secondly, parasites in raw fish (sashimi) pose risks to human health, causing problems such as abdominal pain, diarrhea, vomiting, and organ damage. This issue is particularly concerning to consumers. While chemical prevention and control have proven effective against fish parasites, they can also lead to drug resistance. Non-chemical methods are currently underdeveloped, leaving the industry lacking appropriate strategies to mitigate the effects of parasites.

[0003] Non-destructive testing (NDT) technologies have been widely applied to food quality grading, providing new solutions to quality problems in agricultural products that cannot be addressed during cultivation. In the fish industry, developing NDT methods for fish parasite detection is crucial, as it plays a key role in ensuring export and food safety. Currently, NDT technologies used to assess food quality include visible / near-infrared (VIS / NIR) spectroscopy, VIS / NIR hyperspectral imaging, machine vision, and electronic noses. However, electronic nose technology requires at least one minute per sample due to the need for volatile sampling and sensor cleaning. On the other hand, machine vision technology focuses only on the appearance of the sample and cannot capture the specific characteristics of fish parasites.

[0004] In contrast, visible / near-infrared spectroscopy can collect a wealth of information from the surface of the target, showing great potential for non-destructive detection of fish parasites. While visible / near-infrared spectroscopy focuses on a single point of the target, visible / near-infrared hyperspectral imaging can acquire comprehensive information about the entire fish body. Therefore, visible / near-infrared (VIS / NIR) hyperspectral imaging can provide a comprehensive understanding of fish and enable the development of a method for detecting fish parasites.

[0005] The applicant's earlier application CN114359539B discloses a method for intelligent identification of parasites in sashimi using hyperspectral images.

[0006] This method is mainly used for the detection of parasites or parasite eggs in fish fillets.

[0007] The above methods have many shortcomings when it comes to whole fish testing.

[0008] The most pressing question is: how to determine the image region to be detected?

[0009] If we examine the spectrum of a whole fish image, the data processing volume would be too large; if we examine a local area, can it accurately pinpoint the presence of parasites? These are all problems that need to be solved.

[0010] Summary of the Invention

[0011] The purpose of this invention is to provide a non-destructive detection method for identifying parasitic infections in fish. This invention was verified using minnows and found that the detection accuracy is higher when the fish's abdomen is located, and the acquired images have fewer pixels, require less data analysis, and have a faster calculation speed.

[0012] In addition, the present invention also provides a system for implementing the method.

[0013] To achieve the above objectives, the present invention provides the following technical solution: a non-destructive detection method for identifying parasitic infections in fish, which acquires the average original spectrum of the fish belly using visible light / near-infrared hyperspectral imaging technology, performs noise reduction or removal and feature extraction on the average original spectrum, and adds the extracted features to a pre-established model for identification to obtain the identification result.

[0014] In the above-mentioned non-destructive detection method for identifying parasitic infections in fish, the average original spectral acquisition area is not less than an area of ​​15 pixels * 15 pixels.

[0015] In the above-mentioned non-destructive detection method for identifying parasitic infections in fish, the original spectra are collected from the position corresponding to the stomach on the abdomen of the fish.

[0016] In the above-mentioned non-destructive detection method for identifying parasitic infections in fish, the standard normal variation method is used to reduce the noise of the average original spectrum.

[0017] In the above-mentioned non-destructive detection method for identifying parasitic infections in fish, a competitive adaptive reweighted sampling method is used to extract features of the average original spectrum after noise reduction processing.

[0018] In the above-mentioned non-destructive detection method for identifying parasitic infections in fish, the model is established using partial least squares discriminant analysis.

[0019] In the above-mentioned non-destructive detection method for identifying fish parasite infections, an RGB image containing the fish is acquired using visible light / near-infrared hyperspectral imaging technology; the location of the fish is determined using the RGB image; and the average raw spectrum is collected from the location of the stomach in the fish's abdomen using visible light / near-infrared hyperspectral imaging technology.

[0020] In the above-mentioned non-destructive testing method for identifying parasitic infections in fish, the fish is the minnow; and the parasite is Clonorchis sinensis.

[0021] Meanwhile, the present invention also discloses a system for implementing the above method, which includes a visible light / near-infrared hyperspectral imaging system and a server;

[0022] The visible / near-infrared hyperspectral imaging system is used to acquire the average raw spectrum;

[0023] The server is used to reduce or remove noise and extract features from the average raw spectrum. The extracted features are then added to a pre-established model for recognition to obtain the recognition result.

[0024] Compared with the prior art, the beneficial effects of the present invention are:

[0025] This invention is the first to utilize visible / near-infrared (VIS / NIR) hyperspectral imaging technology for non-destructive detection of parasites in fish.

[0026] The results show that visible / near-infrared hyperspectral imaging, combined with appropriate feature extraction, spectral preprocessing, and modeling methods, can serve as a reliable tool for detecting parasites in fish.

[0027] The optimal preprocessing and modeling method for fish parasites is raw spectrum + SNV + CARS + PLS_DA. Using this method, the detection accuracy on the calibration set and validation set based on the average spectrum of the whole fish is 99.99% and 90.90%, respectively.

[0028] When detecting parasites in different parts of the fish, the detection accuracy, from highest to lowest, was: stomach, head, tail / fin, and back. Specifically, the detection accuracy based on the average spectrum of the fish abdomen was 99.24% and 93.93% for the calibration and validation sets, respectively. These results indicate that the smallest region for non-destructive detection of gastric parasites in fish is a single spectral extraction unit (15×15 pixels), with detection accuracies of 99.99% and 90.90% for the calibration and validation sets, respectively.

[0029] Therefore, the recommended optimal method for non-destructive detection of fish parasites is to use the average raw spectrum of the stomach region, combined with SNV, CARS, and PLS_DA. If detection speed is a primary consideration, it is also recommended to use the average raw spectrum of the smallest region of the stomach, again combined with SNV, CARS, and PLS_DA. These results confirm that visible / near-infrared (VIS / NIR) hyperspectral imaging provides a smart, non-destructive method for detecting parasites in fish, which is of great significance to global fisheries. Attached Figure Description

[0030] Figure 1 is a structural schematic diagram of Embodiment 1 of the present invention;

[0031] Figure 2a is an RGB image of the fish shown in Example 2;

[0032] Figure 2b is a schematic diagram of the parts of the fish shown in Example 2;

[0033] Figure 3a shows the average original spectrum of the whole fish in Example 3;

[0034] Figure 3b shows the spectrum of the whole fish after SNV treatment in Example 3;

[0035] Figure 4a shows the average original spectrum of the fish head in Example 4;

[0036] Figure 4b shows the spectrum of the fish head after SNV treatment in Example 4;

[0037] Figure 4c shows the average original spectrum of the fish back in Example 4;

[0038] Figure 4d shows the spectrum of the fish back after SNV treatment in Example 4;

[0039] Figure 4e shows the average original spectrum of the fish belly in Example 4;

[0040] Figure 4f shows the spectrum of the fish belly after SNV treatment in Example 4;

[0041] Figure 4g shows the average original spectrum of the fish tail in Example 4;

[0042] Figure 4h shows the spectrum of the fish tail after SNV treatment in Example 4;

[0043] Figure 4i shows the average original spectrum of the fish fins in Example 4;

[0044] Figure 4j shows the spectrum of the fish fin after SNV treatment in Example 4;

[0045] Figure 5a is the average original spectrum of the smallest region of the fish head in Example 5;

[0046] Figure 5b shows the spectrum of the smallest region of the fish head in Example 5 after SNV treatment;

[0047] Figure 5c shows the average original spectrum of the smallest region on the back of the fish in Example 5;

[0048] Figure 5d shows the spectrum of the smallest region on the back of the fish in Example 5 after SNV treatment;

[0049] Figure 5e is the average original spectrum of the smallest region of the fish belly in Example 5;

[0050] Figure 5f shows the spectrum of the smallest region of the fish belly after SNV treatment in Example 5;

[0051] Figure 5g is the average original spectrum of the smallest region of the fish tail in Example 5;

[0052] Figure 5h shows the spectrum of the smallest region of the fish tail after SNV treatment in Example 5;

[0053] Figure 5i is the average original spectrum of the smallest region of the fin in Example 5;

[0054] Figure 5j shows the spectrum of the smallest region of the fin in Example 5 after SNV treatment. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Example 1

[0057] Visible / Near Infrared (VIS / NIR) Hyperspectral Imaging Platform

[0058] Figure 1 illustrates the visible / near-infrared (VIS / NIR) hyperspectral imaging platform used in this invention, which specifically includes a dark box 1, a support 2 located inside the dark box, light sources 3 fixed on the left and right sides of the support, and a hyperspectral imaging spectrometer 4 (Pika XC2 spectrometer, Resonon Inc.) mounted on the top of the support. The hyperspectral imaging spectrometer employs a linear array CCD push-broom imaging camera. To ensure the acquisition of comprehensive hyperspectral images, a sample converter is used to synchronize the scanning speed of the imaging camera and the spectrometer.

[0059] The bottom of the support is the stage 5, and there is a placement platform 6 on the stage. The fish 8 is placed on the placement platform, and the hyperspectral imaging spectrometer is connected to the server 7 through the data cable 9.

[0060] The distance between the hyperspectral imaging spectrometer and the stage is 39 cm, and the distance between the light source and the stage is 22 cm.

[0061] Each hyperspectral raw image comprises 1200 rows and 1600 columns, capturing the visible / near-infrared spectrum of each pixel within the wavelength range of 386.37 nm to 1102.86 nm (including 462 nm data).

[0062] The hyperspectral image is in *.bil file format and the data type is uint 16.

[0063] Each time a hyperspectral image was acquired, six fish samples were placed on a white A4 sheet of paper and photographed simultaneously. The optimization of sampling parameters required repeated testing and adjustment. The final parameters are as follows: integration time of 11.87 ms, frame rate of 83.19 fps, line width of 52, scanning speed of 15.87 cm / s, movement speed of 14.99 cm / s, and jogging speed of 1.587 cm / s.

[0064] Example 2

[0065] Biochemical methods for detecting parasites

[0066] Wild topmouth gudgeon were sampled using visible / near-infrared (VIS / NIR) hyperspectral imaging on the platform described in Example 1. The entire fish meat sample was then treated with melamine and subjected to pepsin digestion for 20 hours, with a fish meat to pepsin weight ratio of 1:10. The pepsin solution was artificially prepared by mixing 15 g of pepsin with 8 ml of 36 wt% hydrochloric acid, followed by the addition of 1000 ml of 0.9 wt% physiological saline. After digestion, the filtrate was collected through a 60-mesh sieve, allowed to stand for 30 minutes, and then the supernatant was removed. The remaining residue was thoroughly washed with physiological saline, filtered through a 200-mesh sieve, and examined under an optical microscope for the presence of Clonorchis sinensis.

[0067] As a result, this study observed that 21 out of 66 fish were infected with Clonorchis sinensis, with an infection rate of 31.82%.

[0068] In Example 2, visible / near-infrared (VIS / NIR) hyperspectral imaging sampling was performed, and visible / near-infrared hyperspectral imaging data were obtained.

[0069] RGB images were extracted from visible / near-infrared hyperspectral imaging data, including images at wavelengths of 615.98 nm, 549.62 nm, and 470.39 nm. RGB images visually display each sample to determine its contour, providing a localization basis for subsequent acquisition of the local average raw spectrum of the fish. See Figure 2a for an example of an RGB image.

[0070] Each pixel in the visible / near-infrared hyperspectral imaging data contains the spectrum in both the visible and near-infrared ranges. To further investigate the parasite detection effect in different parts of the fish, each fish sample was carefully divided into five parts: head, back, abdomen, tail, and fins (Fig. 2b).

[0071] Example 3

[0072] Data preprocessing and modeling based on whole-fish visible / near-infrared hyperspectral imaging data

[0073] Step 1: Obtain the average raw spectrum of the entire fish (from head to tail) from visible / near-infrared hyperspectral imaging data;

[0074] Specifically, the visible / near-infrared hyperspectral imaging data is the raw spectrum. The average raw spectrum can be obtained by averaging the raw spectrum.

[0075] A region contains many pixels, and each pixel stores a spectrum, namely the original spectrum. The average original spectrum is the average of the multiple original spectra of this region, forming a spectral curve that can represent this region. This spectral curve is the average original spectrum.

[0076] Step 2: Apply the Savitzky-Gorye filter (SG) (Jamshidi, Minaei, Mohajerani, & Ghassemian, 2012) to the average raw spectrum to minimize the jitter noise of the average raw spectrum.

[0077] Step 3: Based on the processing in Step 2, the standard normal variation (SNV) method (Barnes, Dhanoa and Lister, 1989) is used to reduce scattering noise;

[0078] Figure 3a shows the average original spectrum of the whole fish. In Figure 3a, the horizontal axis represents wavelength and the vertical axis represents spectral energy value.

[0079] The spectra shown in Figure 3a fail to detect waveform differences between infected and uninfected samples. Inputting these spectra into a model built using the PLS-DA method yields unsatisfactory results.

[0080] Figure 3b shows the spectrum after SNV processing. As can be seen from Figure 3b, the clustering performance of the spectrum is significantly improved after SNV processing, thereby improving the accuracy of parasite detection.

[0081] Figures 3a and 3b above show the spectrum of the entire fish;

[0082] The Savitzky-Gore filter (SG) and the standard normal variation (SNV) method do not require parameter settings and can use the default parameters.

[0083] Step 4: Feature extraction. The CARS method is used to extract features from the spectrum processed in Step 3.

[0084] In the preliminary experiments, Principal Component Analysis (PCA) (Wold, Esbensen, and Gelabi, 1987) and Continuous Projection Algorithm (SPA) (Soares, Gomes, Araujo, ...) were used respectively. Filho and Feature extraction was performed using the competitive adaptive reweighted sampling method (CARS) (Li, Liang, Xu and Cao, 2009);

[0085] After verification, CARS outperforms PCA and SPA. This can be attributed to the fact that CARS is a supervised feature selection method that selects features based on the relationship between input and output data (Yao et al., 2022), while PCA and SPA are unsupervised methods that rely only on the differences in input data (Sauwen et al., 2016; Jiang et al., 2018).

[0086] CARS (Competitive Adaptive Reweighted Sampling) is a feature selection method for spectral data analysis. It selects the bands with the highest information content through an iterative adaptive reweighting sampling process. In each step, competitive screening and weight adjustment are performed based on the contribution of each band to the model to obtain the feature bands, thereby improving the model's predictive performance.

[0087] Step 5: Model building, using PLS-DA to build the model;

[0088] In this step, three methods were used to verify the optimal modeling method: Partial Least Squares Discriminant Analysis (PLS-DA) (Pérez-Enciso & Tenenhaus, 2003), Support Vector Machine (SVM) (Suthaharan & Suthaharan, 2016), and Random Forest (RF) (Rigatti, 2017).

[0089] In parasite detection modeling, PLS-DA outperforms SVM and RF. PLS-DA primarily focuses on detecting linear correlations between input and output data (Chevallier, Bertrand, Kohler, & Courcoux, 2006), while SVM and RF emphasize detecting nonlinear correlations (Dong, Li, & Xie, 2014). Therefore, it can be inferred that linear correlations contribute more than nonlinear correlations in fish parasite detection modeling.

[0090] During the modeling process, the data was divided into two parts, a calibration set and a validation set, using a random partitioning method with a ratio of 8:2, where the calibration set consisted of 8 parts and the validation set consisted of 2 parts.

[0091] In whole-fish detection, the modeling data is the average spectral value corresponding to the region where each fish is located.

[0092] During the testing phase, infected and uninfected fish were randomly divided into three groups. Each test performed three-fold cross-validation, with each cross-validation set consisting of alternating sets of infected and uninfected samples (22 samples) and two cross-validations consisting of a calibration set (44 samples).

[0093] In the PLS-DA model, the latent variables (LVs) refer to the latent variables selected by the model, extracted by maximizing the covariance between the explanatory and response variables. Properly setting LVs helps achieve effective data dimensionality reduction and balances the model's interpretability and predictability. The optimal number is typically determined by examining the proportion of variance explained by each component or by using cross-validation to avoid overfitting and improve model performance.

[0094] Step 6: After the model is built, the features of the fish to be detected are obtained through the methods in steps 1 to 4. The features are then input into the model trained in step 5 for detection to obtain the detection results.

[0095] The parasite detection results based on the visible / near-infrared spectra of the whole fish are shown in Table 1;

[0096] Table 1: Parasite detection results based on different processing and modeling methods of whole fish VIS / NIR spectra

[0097] Note: The 1-fold, 2-fold, and 3-fold divisions in the table above represent different ways of dividing the dataset. Due to the small number of samples, in order to obtain more convincing results, each group of randomly grouped samples was used as a validation set independently, and the other two groups were used as calibration sets. The above three-fold divisions were performed three times to obtain their respective accuracies, thereby obtaining the average accuracy and verifying the applicability of the algorithm.

[0098] As can be seen from the results in Table 1 above, its accuracy is 90.90%.

[0099] Example 4

[0100] Data preprocessing and modeling based on visible / near-infrared hyperspectral imaging data of fish parts

[0101] Step 1: Obtain the average raw spectrum of different parts of the fish from visible / near-infrared hyperspectral imaging data;

[0102] The parts described in this step are the fish head, back, belly, tail, and fins;

[0103] Step 2: Apply a Savitzky-Gore filter (SG) to the averaged raw spectrum to minimize jitter noise;

[0104] Step 3: Based on the processing in Step 2, the standard normal variation (SNV) method is used to reduce scattering noise;

[0105] Figures 4a-4i show the original average spectra and the spectra processed by the SNV method for the fish head, back, belly, tail, and fins (hereinafter referred to as SNV spectra), respectively. In Figures 4a-4i, the sampling area is the entire part, such as the original average spectrum of the fish head, which refers to the original average spectrum of the entire fish head area.

[0106] It is noteworthy that the clustering performance of the original spectrum on the back side (Fig. 4c) is relatively worse than that of other regions. This difference may be due to the greater curvature of the back region, which introduces more scattering noise into the spectrum. Therefore, compared to other regions, the SNV spectrum on the back side (Fig. 4d) exhibits a greater degree of jitter noise.

[0107] Furthermore, compared to the original average spectra of the head (Fig. 4a) and abdomen (Fig. 4e), the original average spectra of the tail (Fig. 4g) and fins (Fig. 4i) appear to be flatter. From this observation, it can be inferred that the spectra of the head and abdomen capture more information about the fish's characteristics.

[0108] Step 4: Feature extraction. The CARS method is used to extract features from the spectrum processed in Step 3.

[0109] Step 5: Model building, using PLS-DA to build the model;

[0110] In the partial detection, the modeling data is the average spectral data of the head, stomach, back, tail, and fin regions corresponding to each fish.

[0111] Step 6: After the model is built, the features of the fish to be detected are obtained through the methods in steps 1 to 4. The features are then input into the model trained in step 5 for detection to obtain the detection results.

[0112] The results of parasite detection based on visible / near-infrared spectroscopy of fish parts are shown in Table 2;

[0113] Table 2: Parasite detection results based on different processing and modeling methods of VIS / NIR spectra from different parts of the fish.

[0114] The modeling and detection results based on the spectra of different fish body parts (using the original average spectrum + SNV + CARS + PLS-DA) (Table 2) are consistent with the spectral characteristics shown in Figure 4. Regarding parasite detection accuracy, the accuracy of different fish body parts, from highest to lowest, is: abdomen > head > tail / fin > back. Notably, the validation set accuracy using the stomach spectrum (93.93%) exceeded that using the whole fish spectrum. This indicates that removing appropriate redundant information can further improve detection accuracy.

[0115] Example 5

[0116] Data preprocessing and modeling based on the minimum region of visible / near-infrared hyperspectral imaging data of fish parts

[0117] Step 1: Obtain the average raw spectrum of the smallest region from different parts of the fish from the visible / near-infrared hyperspectral imaging data;

[0118] The parts described in this step are the fish head, back, belly, tail, and fins;

[0119] The smallest region refers to a 15x15 pixel area of ​​that part.

[0120] Step 2: Apply a Savitzky-Gore filter (SG) to the averaged raw spectrum to minimize jitter noise;

[0121] Step 3: Based on the processing in Step 2, the standard normal variation (SNV) method is used to reduce scattering noise;

[0122] Figures 5a to 5j show the original and SNV spectra of the minimum regions of the head, back, abdomen, tail, and fins. Notably, the clustering performance of the original spectrum from the minimum region of the head (Figure 5a) is worse than that of the original spectrum of the entire head region (Figure 4a), due to the heterogeneity of the head, including the face, mouth, and eyes. Conversely, the clustering performance of the original spectrum from the minimum region of the back (Figure 5c) is better than that of the entire back region (Figure 4c), because the smaller region effectively reduces some scattering noise caused by the non-planar nature of the back. However, the original spectral waveforms of the minimum regions of the back, as well as the minimum regions of the tail (Figure 5g) and fins (Figure 5i), are not particularly pronounced. On the other hand, the original spectral waveform quality of the minimum region of the abdomen (Figure 5e) remains good, and its SNV spectrum (Figure 5f) has less noise compared to the head (Figure 5b) and back (Figure 5d). Furthermore, the waveform of the minimum region of the stomach is also better than that of the tail (Figure 5h) and fin (Figure 5j) regions.

[0123] Step 4: Feature extraction. The spectral data of the smallest region of the stomach after processing in Step 3 are used to extract features using the CARS method.

[0124] Step 5: Model building, using PLS-DA to build the model;

[0125] This step extracted the spectra of one, two, three, four, and five minimal regions (15*15 pixels) in the stomach. This data was used to build a parasite detection model.

[0126] In the minimum region detection (15x15 pixels), the modeling data is the average raw spectral data extracted from the 15x15 rectangle of the stomach region of each fish;

[0127] For the stomach, which yields the best results in site-specific detection, a minimum spectral region (15*15 pixels) is selected to reduce the amount of spectral data and computational complexity, thereby improving detection speed while maintaining high accuracy. By selecting spectral data from one, two, three, four, and five minimum regions in the stomach, and then processing this data through model building, the accuracy rate corresponding to selecting a single minimum region was found to be optimal. This confirms that selecting the spectral information of a 15*15 pixel region in the stomach is feasible, enabling faster detection of parasites in fish.

[0128] Step 6: After the model is built, the features of the fish to be detected are obtained through the methods in steps 1 to 4. The features are then input into the model trained in step 5 for detection to obtain the detection results.

[0129] The results of parasite detection based on the minimum visible / near-infrared spectra of fish parts are shown in Table 3;

[0130] Table 3. Parasite detection results based on fish minimum region spectroscopy.

[0131] After the ideal region size was determined to be the minimum region, the detection accuracy on the validation set reached 90.90%.

[0132] Combining the spectra of the smallest regions of different parts of the fish during the modeling process (mixed regions in Table 3) significantly reduces the detection accuracy of the validation set, dropping it to 79.39%.

[0133] Surprisingly, the detection performance using the raw average spectrum of the entire fish was comparable to that using the smallest region of the stomach. This indicates that as the spectral extraction area increases, the detection accuracy initially decreases and then begins to increase, while the ratio of useful information to noise also increases. Based on the dual requirements of efficiency and accuracy, this embodiment uses the average raw spectrum of the smallest region of the stomach, preprocesses it using SNV and CARS, and performs detection based on a model built using the PLS_DA method.

[0134] Results analysis:

[0135] The experimental results of Examples 3 to 5 of the present invention determine that the most effective method for fish parasite pretreatment and modeling is raw spectrum + standard normal variable (SNV) + competitive adaptive reweighted sampling (CARS) + partial least squares discriminant analysis (PLS_DA).

[0136] The accuracy of parasite detection in different parts of a fish, from highest to lowest, is as follows: stomach, head, tail / fin, and back.

[0137] The smallest area for non-destructive parasite detection is a single spectral extraction unit (15×15 pixels) in the stomach.

[0138] The optimal non-destructive detection method for fish parasites is to use the average raw spectrum of the stomach in conjunction with SNV, CARS, and PLS_DA. This method achieved detection accuracies of 99.24% and 93.93% on the calibration and validation sets, respectively. For applications requiring faster detection, it is recommended to consider the average raw spectrum of the smallest region of the stomach, combined with SNV, CARS, and PLS_DA. Using this method, the detection accuracies on the calibration and validation sets are 99.99% and 90.90%, respectively. These results demonstrate the significant potential of VIS / NIR hyperspectral imaging as a smart non-destructive detection method for fish parasites.

[0139] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A non-destructive detection method for identifying parasitic infections in fish, characterized in that, The average raw spectrum of the fish belly is acquired by visible light / near infrared hyperspectral imaging technology. Noise reduction or removal and feature extraction are performed on the average raw spectrum. The extracted features are added to a pre-established model for recognition to obtain the recognition result.

2. The non-destructive detection method for identifying parasitic infections in fish according to claim 1, characterized in that, The average original spectral acquisition area is not less than an area of ​​15 pixels * 15 pixels.

3. The non-destructive detection method for identifying parasitic infections in fish according to claim 1, characterized in that, The original spectra were collected from the abdomen of the fish, corresponding to the stomach.

4. The non-destructive detection method for identifying fish parasite infections according to claim 1, characterized in that, The noise of the average raw spectrum is reduced by using the standard normal variation method.

5. The non-destructive detection method for identifying fish parasite infections according to claim 4, characterized in that, A competitive adaptive reweighted sampling method is used to extract features from the average original spectrum after noise reduction processing.

6. The non-destructive detection method for identifying parasitic infections in fish according to claim 1, characterized in that, The model was established using partial least squares discriminant analysis.

7. The non-destructive detection method for identifying parasitic infections in fish according to claim 1, characterized in that, RGB images containing fish were acquired using visible / near-infrared hyperspectral imaging technology; the location of the fish was determined using the RGB images; and the average raw spectrum was collected from the location of the stomach in the fish's abdomen using visible / near-infrared hyperspectral imaging technology.

8. The non-destructive detection method for identifying parasitic infections in fish according to claim 1, characterized in that, The fish in question is the minnow; the parasite is Clonorchis sinensis.

9. A system for implementing the method of any one of claims 1-8, characterized in that: Includes visible / near-infrared hyperspectral imaging systems and servers; The visible / near-infrared hyperspectral imaging system is used to acquire the average raw spectrum; The server is used to reduce or remove noise and extract features from the average raw spectrum. The extracted features are then added to a pre-established model for recognition to obtain the recognition result.

Images