A method and system for determining the end point of fried potato chips

Abstract

The present application belongs to the technical field of food processing and quality detection, and particularly relates to a method and system for determining the end point of fried potato chips. The method controls multiple different wavelength light sources to illuminate potato chip samples in turn, collects image responses of the same sample under different wavelengths, constructs multi-band response characteristics in the frying process of potato chips, establishes a frying state and end point determination model, realizes the determination of the states of unfinished potato chips, near end point, reaching end point and over frying, and can more fully reflect the surface color change, structure change, browning change and ripening state change of potato chips in the frying process.

Description

Technical Field

[0001] This invention belongs to the field of food processing and quality testing technology, specifically relating to a method and system for determining the endpoint of fried French fries. Background Technology

[0002] Fried potato products are common snacks and fast foods, with French fries being widely popular due to their appealing color, flavor, and texture. In recent years, with the continuous improvement of industrialization, standardization, and intelligent production levels in fried food manufacturing, the market has placed higher demands on French fries in terms of appearance, color, degree of cooking, uniformity of browning, and processing stability. During the frying process, the quality of French fries is influenced by various factors, including frying temperature, frying time, raw material condition and size, and heat and mass transfer processes. Accurately determining the frying endpoint is crucial for ensuring product quality stability, reducing the risk of over-frying, improving production efficiency, and reducing energy consumption.

[0003] Currently, judging the endpoint of deep-fried French fries usually relies on human experience, timed control, or simple observation based on surface color. This approach suffers from high subjectivity, poor repeatability, and difficulty in adapting to differences in raw materials and process fluctuations. Especially under continuous production conditions, relying solely on fixed times or a single appearance indicator often fails to ensure consistency in the color, browning degree, and ripeness of the fries, easily leading to quality problems such as under-frying, excessive browning, and uneven coloring.

[0004] Existing methods for inspecting the quality of French fries mostly rely on manual sampling, offline photo analysis, or single-band visual inspection, which typically only acquire limited appearance information and are insufficient for rapid, comprehensive, and accurate assessment of the frying state. Therefore, current technology lacks a method that can directly characterize the processing state of French fries and determine the optimal frying endpoint by utilizing changes in multi-wavelength image response during the frying process. In particular, existing methods often focus on single appearance indicators or offline quality evaluation, failing to reflect the continuous state changes of French fries during frying and thus unable to meet the requirements for online, process-oriented, and intelligent control. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and system for determining the endpoint of deep-fried French fries.

[0006] A method for determining the endpoint of deep-fried French fries, comprising the following steps: Multiple spectral images of fried French fries samples under different frying conditions were acquired at multiple wavelengths to obtain multiple spectral images; Multiple acquired spectral images are preprocessed to eliminate interference from non-target factors, and the target region is extracted from each spectral image. Image features and cross-band response features are extracted from the target region. The cross-band response features include the average reflectance, reflectance standard deviation, and channel combination feature parameters corresponding to each wavelength of the target region. The image features include color features, texture features, edge features, and morphological features. The neural network convolutional model is trained by inputting spectral images at multiple wavelengths, the cross-band response features, and the image features into the neural network convolutional model to obtain a neural network convolutional model for outputting the fried state response. The spectral images of the French fry sample to be tested at multiple wavelengths are acquired, along with the corresponding image features and cross-band response features. The fried state response of the French fry sample to be tested is output using a trained neural network convolutional model.

[0007] Preferably, the plurality of wavelengths are 450 nm, 470 nm, 490 nm, 510 nm, 530 nm, 550 nm, 570 nm, 590 nm, 610 nm, 630 nm, 650 nm, 700 nm, 750 nm, 800 nm, 850 nm and 900 nm.

[0008] Preferably, the preprocessing includes whiteboard correction, dark current correction, background removal, and target segmentation.

[0009] Preferably, the color features include the average value, standard deviation, maximum value, and minimum value of each RGB channel in the target area, the average grayscale value, the standard deviation of grayscale value, the R / G color channel ratio, the R / B color channel ratio, the G / B color channel ratio, and the proportion of dark pixels; The texture features include the gray-level variance, gray-level standard deviation, number of dark spots, area ratio of dark spots, local gray-level variation amplitude, texture contrast, texture uniformity, and texture entropy of the target area. The edge features include the number of edge pixels, edge pixel ratio, contour length, average edge gray value, average edge reflectance, and color difference or reflectance difference between the edge region and the center region of the French fry target area. The morphological features include the area, length, width, aspect ratio, perimeter, area of ​​the circumscribed rectangle, rectangularity, circularity, and number of targets in the target region.

[0010] Preferably, the cross-band response characteristics include a browning characteristic index, a comprehensive color difference characteristic index, and a frying endpoint determination characteristic index: Browning Characteristic Index B: ; Comprehensive color difference characteristic index C: ; Deep-frying endpoint determination characteristic index F: ; , , , , , , , , , , and This represents the average reflectance under the corresponding spectrum.

[0011] Preferably, the average reflectance, reflectance standard deviation, and channel combination feature parameters corresponding to each wavelength are obtained based on the color features, texture features, edge features, and morphological features; the channel combination feature parameters include the difference, ratio, and normalized difference between the average reflectance of two or more wavelength channels, as well as the difference or ratio between the standard deviations of different wavelength channels; the cross-band response features are features based on image features, average reflectance, reflectance standard deviation, and channel combination feature parameters under multiple wavelength channels.

[0012] Preferably, the fried French fry sample is obtained by frying French fries at 150~190℃ for 3~7 minutes.

[0013] A system for implementing the aforementioned determination method includes: Image acquisition module: includes multiple wavelength light sources and imaging units, used to acquire spectral images of French fry samples under illumination conditions of various wavelengths; Image preprocessing module: used for denoising, correcting, and segmenting target regions in the acquired images; Feature extraction module: used to extract color, texture, edge, shape and cross-band response features; Model building and training module: Equipped with a neural network convolutional model, used for learning extracted features and training the model; State determination module: Used to determine the state of the French fry sample under test using a trained neural network convolutional model and obtain the output result; Display module: Used to display the output results.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention acquires image responses of French fry samples at different wavelengths using multi-wavelength LED active imaging. Compared to traditional RGB images or single-color judgment, it can more fully reflect the surface color changes, structural changes, browning changes, and cooking state changes of the French fries during frying. By establishing a frying process state determination model, this invention can achieve non-contact, rapid, and stable identification of the frying process state and optimal frying endpoint of French fries, making it suitable for online monitoring and precise control of the French fry processing. Attached Figure Description

[0015] Figure 1 The diagram and photograph are of the equipment structure, where A is the structural diagram and B is the photograph of the equipment.

[0016] Figure 2 This is the original image of deep-fried French fries.

[0017] Figure 3 Feature values ​​under different processing conditions were extracted for the neural network convolutional model. A represents the change in feature values ​​of 16 channels under 3 minutes of frying time, B represents the change under 5 minutes of frying time, C represents the change under 7 minutes of frying time, D represents the change under 150℃ frying temperature, E represents the change under 170℃ frying temperature, and F represents the change under 190℃ frying temperature. The lines represent different image feature channels; these feature channels change with frying temperature and time; therefore, they can be used as a basis for identifying the frying state. Detailed Implementation

[0018] The specific embodiments of the present invention are described in detail below, but it should be understood that the scope of protection of the present invention is not limited to the specific embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention. Unless otherwise specified, the experimental methods described in the embodiments of the present invention are conventional methods.

[0019] As mentioned earlier, in the existing French fry processing, the frying endpoint is usually determined by manually observing color changes, judging by experience, or controlling a fixed time. For indicators such as the overall color difference, browning degree, and endpoint state of the French fries, further manual judgment or offline analysis is usually required. It is not possible to make an accurate and comprehensive judgment at one time, which affects the stability of product quality and the efficient control of the frying process.

[0020] Multispectral image analysis technology can simultaneously acquire image information of target samples in multiple specific wavelength bands. Compared with traditional color images that only contain red, green and blue channel information, multispectral images can reflect the reflectance characteristics of samples at different wavelengths. Therefore, it has the advantages of large information content, strong selectivity and suitability for multi-feature analysis, and can be used for the identification and determination of quality status in food processing.

[0021] Based on this, the present invention proposes a method for determining the process and endpoint of frying French fries. The present invention acquires multispectral images of French fries samples under 16 LED wavelength illumination conditions, extracts color, texture, edge, shape, and cross-band response features, and employs a convolutional neural network (such as the YOLO algorithm) as a feature learning model to extract features and determine the state of the image response of the French fries target area under different wavelengths.

[0022] In one embodiment of the present invention, a method for establishing a feature model under 16 LED wavelength illumination conditions for determining the endpoint of frying French fries is provided, comprising the following steps: French fry samples under different frying times and temperatures were placed on a multispectral image acquisition platform for image acquisition, obtaining image data of the French fry samples in 16 spectral channels; the center wavelengths of the 16 spectral channels were 450 nm, 470 nm, 490 nm, 510 nm, 530 nm, 550 nm, 570 nm, 590 nm, 610 nm, 630 nm, 650 nm, 700 nm, 750 nm, 800 nm, 850 nm and 900 nm.

[0023] After image acquisition, the original multispectral images were subjected to whiteboard correction, dark current correction, background removal, and target segmentation to obtain image data of the French fry target area under 16 LED wavelength illumination conditions. The average reflectance, standard deviation, and channel combination feature parameters for each channel were extracted. These feature parameters are used to characterize the surface color changes, browning degree, texture changes, and differences in frying state of the French fries at different frying stages.

[0024] During the frying process, the color, surface structure, edge morphology, and degree of browning of French fry samples change with time and temperature, resulting in differences in their image response under different LED wavelength illumination conditions. This invention utilizes these multi-wavelength response differences to construct frying process state features and establishes a correspondence between the frying process state and the optimal frying endpoint through a feature learning model.

[0025] This invention presents a convolutional neural network-based frying process state determination model based on image response features across 16 LED wavelengths. The model takes image features, color features, texture features, edge features, and morphological features at 16 wavelengths, along with three manually added cross-band response indices (B, C, and F), as input. After multi-layer feature filtering and fusion by the convolutional neural network, it outputs response values ​​or probability values ​​indicating the different frying process states of the French fries. B, C, and F are auxiliary feature indices added to the neural network to enhance the model's ability to identify browning changes, overall color changes, and endpoint trends. The final determination result is determined by the state response value output by the neural network.

[0026] Average reflectance across 16 spectral channels , , , , , , , , , , , , , , and Basic characteristics; Optionally, the cross-band response characteristics are determined by a combination of one or more of the following band characteristics: Constructing a browning characteristic index : ; Construct a comprehensive color difference characteristic index: ; Constructing a feature index for determining the frying endpoint: ; In this scheme, the convolutional neural network model for determining the frying process state employs supervised training. Spectral images of fried French fries samples under multiple frying states are acquired at multiple wavelengths. After spectral image preprocessing, image features and cross-band response features of the fried French fries samples are extracted. These spectral images at multiple wavelengths, the cross-band response features, and the image features are then input into the neural network convolutional model for training, resulting in a neural network convolutional model for outputting the frying state response, as detailed below:

[0027] The training dataset consists of 1000 samples of French fries in known frying states. Each sample represents one of four states: incomplete, nearing completion, reached completion, or over-fried. During sample acquisition, imaging responses at 16 LED wavelengths were simultaneously acquired. Feature engineering was used to extract color, texture, edge, and morphological features from the images at each wavelength. Cross-band response indices B, C, and F were also added. All features were concatenated to form a fixed-dimensional structured input vector. The dataset was divided into training and validation sets in a 7:3 ratio. The input was the aforementioned multi-source feature vector, and the output was the one-hot encoded label of the sample's true state. Multi-class cross-entropy loss was used as the training loss function. The model was trained on the training set and tested on the validation set to obtain the final trained model.

[0028] Under different frying temperatures and frying times, French fry samples exhibit three different frying states: "undercooked," "just right," and "overcooked." The YOLO-based image recognition method can effectively determine the frying state of the French fry, and the judgment results are highly consistent with the manually labeled results.

[0029] Table 3 shows the confusion matrix obtained by comparing the manual results with the model results, which is used to evaluate the model's recognition accuracy.

[0030] Table 1: Percentage of Manually Annotated Text / % Table 1 shows the results of manual classification and statistical analysis of samples based on the appearance of French fries under different combinations of frying temperatures and times.

[0031] French fry samples were fried at 150℃, 170℃, and 190℃ for 3 minutes, 5 minutes, and 7 minutes, respectively. Then, based on the color, appearance, and degree of over-frying, the samples were manually categorized into three types: "undercooked," "just right," and "over-fried." The values ​​in the table represent the proportion of each category of samples under these temperature and time conditions.

[0032] For example, at 150℃ for 3 minutes, 90% of the samples were uncooked, 10% were just cooked, and 0% were overcooked, indicating that most of the fries were still uncooked when fried at low temperature for a short time. As the frying time increased or the frying temperature rose, the proportion of uncooked samples gradually decreased, while the proportions of just cooked and overcooked samples gradually increased.

[0033] As shown in Table 1, the frying condition of French fries is closely related to frying temperature and frying time. At 150℃, the fries cook relatively slowly; at 170℃ for 5 minutes, the proportion of fries that are cooked just right is the highest, reaching 80%; at 190℃, the proportion of over-fried fries increases significantly with the extension of frying time, especially at 190℃ for 7 minutes, where the proportion of over-fried fries reaches 95%. This indicates that high-temperature, long-term frying easily leads to over-frying of French fries.

[0034] Table 2: Model Prediction Results Table 2 shows the prediction results obtained after using the YOLO image recognition model of this invention to identify the same French fry samples. First, the YOLO model is used to detect the target region in the French fry image to obtain the region where the French fry are located; then, based on the detected features of the French fry image, its frying state is judged, and three prediction results are output: "undercooked", "just right", and "overcooked". Finally, the proportion of samples predicted by the model to each category is statistically analyzed under different frying temperatures and times.

[0035] Table 2 is used to verify whether the model can identify the frying state of the French fries based on image features. A comparison with the manually labeled results in Table 1 shows that the overall trend of the model's predictions and the human judgments is consistent.

[0036] For example, at 150℃ for 3 minutes, the manually labeled uncooked proportion was 90%, while the model predicted an uncooked proportion of 88%, which is very close. At 170℃ for 5 minutes, the manually labeled proportion was exactly 80%, while the model predicted exactly 78%, indicating that the model can accurately identify the appropriate frying state. At 190℃ for 5 minutes, the manually labeled over-fried proportion was 67%, and the model also predicted an over-fried proportion of 67%, indicating that the model has a good ability to identify over-fried conditions.

[0037] Therefore, Table 2 shows that the image recognition model of the present invention can accurately reflect the changing pattern of French fries under different frying conditions.

[0038] Table 3: Confusion Matrix Table 3 is the confusion matrix, used to further evaluate the consistency between the model's predictions and the actual results obtained manually.

[0039] The confusion matrix shows the actual classification results from human intervention on the left and the model's predicted classification results on the top. The values ​​in the table represent the number of samples of a particular actual class that were predicted by the model to belong to each class. The values ​​on the diagonal represent the number of samples correctly predicted by the model, and the values ​​off-diagonal represent the number of samples incorrectly predicted by the model.

[0040] For example, among the samples that were actually "unripe", 280 were correctly predicted as "unripe", 18 were predicted as "just right", and 2 were predicted as "over-exploded"; among the samples that were actually "just right", 250 were correctly predicted as "just right"; and among the samples that were actually "over-exploded", 275 were correctly predicted as "over-exploded".

[0041] According to Table 3, the total number of samples is 900, of which the number of correctly identified samples is 280 + 250 + 275 = 805. Therefore, the overall recognition accuracy of the model is: 805 ÷ 900 = 89.44%.

[0042] This demonstrates that the identification method of the present invention has high accuracy in judging the frying state of French fries.

[0043] Based on the misclassifications, the model's errors mainly occurred between adjacent states. For example, "undercooked" was misclassified as "just right," "just right" as "overcooked" or "undercooked," and "overcooked" as "just right." This aligns with reality, as the transition from undercooked to just right and from just right to overcooked in French fries is a continuous process, and the boundary samples are inherently quite similar in appearance. Conversely, the number of instances where "undercooked" was misclassified as "overcooked" and vice versa was very small, indicating that the model can effectively distinguish between significantly different frying states.

[0044] Example 1 S01: Select French fry samples with basically consistent specifications and no obvious surface damage. Place the samples in a frying device and fry at 150℃ for 3 minutes. After frying, remove the samples and lay them flat on a black low-reflection background paper. Maintain a identifiable gap between adjacent fries to facilitate subsequent target segmentation.

[0045] It should be noted that the black low-reflection background paper and the French fry sample have significant differences in image brightness and color. Therefore, in subsequent image processing, the black background area can be identified as a non-target area and is not included in the calculation of color, texture, edge, shape and cross-band response features.

[0046] S02: Place the prepared French fry sample on a multispectral image acquisition platform, and sequentially turn on LED light sources with different center wavelengths to acquire multispectral images of the French fry sample under multiple LED wavelength illumination conditions. The center wavelengths of the multiple LED wavelengths are 450nm, 470nm, 490nm, 510nm, 530nm, 550nm, 570nm, 590nm, 610nm, 630nm, 650nm, 700nm, 750nm, 800nm, 850nm, and 900nm.

[0047] Specifically, each LED wavelength corresponds to the acquisition of one image of French fries, thus obtaining image data of the same French fries sample under 16 wavelength channels. Images at different wavelengths can reflect the differences in reflectance of the French fries sample under different lighting conditions, thereby providing multi-band image information for subsequent judgment of the frying status.

[0048] S03: Before or during the acquisition of the French fry sample image, acquire the white board image and the dark field image respectively to correct the white board response and dark current response.

[0049] During whiteboard calibration, a standard whiteboard is placed at the same imaging position as the French fry sample, and images of the whiteboard are acquired under the same camera parameters, exposure time, and LED wavelength illumination conditions. This whiteboard image serves as a high-reflectance reference for the corresponding wavelength channel. During image processing, the image response corresponding to the whiteboard area can be normalized to its maximum value, such as 255 or 1, to reduce the impact of differences in brightness from different wavelength LED light sources, camera response differences, and uneven brightness in the shooting area on subsequent analysis.

[0050] When performing dark current correction, dark field images are acquired under conditions of turning off the light source, blocking the lens, or using a black, low-reflection background as a dark field reference. This dark field image serves as the camera's background signal. During image processing, the image response of the dark field or black background area can be normalized to 0 or close to 0 to reduce the impact of camera dark current, ambient stray light, and black level offset on reflectivity calculation.

[0051] The corrected relative reflectance can be expressed as: ; Where I represents the pixel response value of the French fry sample image, D represents the pixel response value of the dark field image, W represents the pixel response value of the white board image, and R represents the corrected relative reflectance.

[0052] It should be explained that the pixel response value can be understood as the digital image value formed after the camera receives the reflected light from the sample; after whiteboard correction and dark current correction, the image value can be converted into relative reflectance for subsequent feature extraction and model calculation.

[0053] S04: Perform background removal and target segmentation on the corrected multispectral image.

[0054] When removing the background, the difference in brightness, color, or reflectivity between the fries area and the black background area is used to obtain a mask for the fries target area through grayscale thresholding, color thresholding, or binarization. The black background area is marked as a non-target area, and the area containing the fries is marked as the target area, so that the background area is not involved in subsequent modeling.

[0055] During target segmentation, based on the binary mask image obtained after background removal, contour detection or connected component analysis is performed using Python and OpenCV image processing methods. Since there are black background gaps between adjacent fries, multiple fries in the image can be separated into individual fries target regions based on these gaps. Furthermore, excessively small noise regions are removed to obtain an independent target region corresponding to each fries.

[0056] S05: Extract image features for each French fry target area.

[0057] The image features include color features, texture features, edge features, and morphological features.

[0058] The color features include the average, standard deviation, maximum, and minimum values ​​of each RGB channel within the target area of ​​the French fries; the average and standard deviation of grayscale; the R / G, R / B, and G / B color channel ratios; and the proportion of dark pixels. These color features are used to characterize the color changes of the French fries during the frying process, from light to yellow, golden, brown, or dark brown.

[0059] Texture features include the grayscale variance, grayscale standard deviation, number of dark spots, area ratio of dark spots, local grayscale variation amplitude, texture contrast, texture uniformity, and texture entropy of the target area of ​​the French fries. These texture features are used to characterize the changes in surface spots, browning areas, charring points, and surface roughness during the frying process of the French fries.

[0060] Edge features include the number of edge pixels, edge pixel ratio, contour length, average edge gray value, average edge reflectance, and color or reflectance difference between the edge and center regions of the French fry target area. These edge features are used to characterize edge browning, edge deepening, changes in contour clarity, and outer edge charring trends during the frying process.

[0061] Morphological features include the area, length, width, aspect ratio, perimeter, area of ​​the circumscribed rectangle, rectangularity, circularity, and number of targets in the target region of the French fry. These morphological features are used to characterize the size, shape, contour regularity of the French fry sample, and the morphological differences between different French fry targets.

[0062] It should be noted that the above features can all be calculated from the target area of ​​the French fries by an image processing program. For each LED wavelength channel, the above image features can be extracted separately to obtain the image response data of the French fries sample under different wavelength conditions.

[0063] S06: Based on the French fry target area image under each wavelength channel, calculate the average reflectance, reflectance standard deviation, and channel combination characteristic parameters.

[0064] Specifically, for the i-th wavelength channel, the average relative reflectance of all pixels in the target area of ​​the French fry is calculated to obtain the average reflectance Ri under this wavelength channel; the dispersion of the relative reflectance of all pixels in the target area of ​​the French fry is calculated to obtain the standard deviation Si of the reflectance under this wavelength channel.

[0065] The average reflectance is used to characterize the overall intensity of the reflected light signal of the French fry sample under corresponding wavelength illumination conditions; the standard deviation of reflectance is used to characterize the uniformity of the reflectance distribution on the surface of the French fry. When spots, browning, edge charring, or uneven color appear on the surface of the French fry, the standard deviation of reflectance at the corresponding wavelength will change.

[0066] Based on this, two different wavelength channels are selected for combined calculation to obtain channel combination characteristic parameters. Taking the i-th wavelength channel and the j-th wavelength channel as an example, the channel combination characteristic parameters include:

[0067] Average reflectance difference: Ri - Rj; Average reflectance ratio: Ri / (Rj+ε); Normalized difference in average reflectance: (Ri-Rj) / (Ri+Rj+ε); Standard deviation: Si - Sj; Standard deviation ratio: Si / (Sj+ε).

[0068] Where Ri and Rj represent the average reflectance of the target area of ​​the French fry under the i-th and j-th wavelength channels, respectively, Si and Sj represent the standard deviation of the reflectance of the target area of ​​the French fry under the corresponding wavelength channels, and ε is a minimal constant set to prevent the denominator from being zero.

[0069] It should be explained that the difference in average reflectance between different wavelength channels reflects the difference in image response strength; the sign of the difference reflects the direction of response change between different wavelengths; the absolute value, ratio, and normalized difference reflect the magnitude of response change between different wavelengths. The difference or ratio of standard deviations between different wavelength channels reflects the degree of difference in uneven color, spots, browning, and texture changes on the surface of the fries at different wavelengths.

[0070] Based on the image response, average reflectance, reflectance standard deviation, and channel combination feature parameters under multiple wavelength channels, cross-band response features are obtained. These features are used to characterize the differences in multi-band image response caused by color yellowing, surface browning, edge deepening, texture changes, and degree of cooking during the frying process of French fries.

[0071] S07: Input the French fry images, color features, texture features, edge features, morphological features, average reflectance, reflectance standard deviation, and channel combination feature parameters from each wavelength channel into the neural network convolutional model for state determination.

[0072] Specifically, the neural network convolutional model extracts local color, texture, edge, and morphological changes in images at different wavelengths through convolution operations. Simultaneously, manually calculated average reflectance, standard deviation, and channel combination feature parameters are used as auxiliary features input to the model. During training, the model continuously adjusts its internal weight parameters based on manually labeled frying status results, thereby learning the correspondence between different wavelength channels and different image features and the frying status of French fries, resulting in the trained neural network convolutional model.

[0073] After training, the images and feature data of the French fries sample fried at 150℃ for 3 minutes in this embodiment were input into the trained neural network convolutional model. The model outputs the response value of the sample under different frying conditions. In this embodiment, the model output value is 0.21, and the judgment result is incomplete. The manual observation result is also incomplete, and the surface condition is characterized by a lighter color.

[0074] The frying process status and endpoint determination results of the French fry samples obtained in this embodiment are shown in Table 4.

[0075] Example 2 The difference between Example 2 and Example 1 is that the frying conditions for the French fry sample were 150°C for 5 minutes, while the remaining steps and conditions were the same as in Example 1.

[0076] Following the method described in Example 1, color features, texture features, edge features, morphological features, average reflectance, reflectance standard deviation, and channel combination feature parameters of the target area for the French fries were extracted at each wavelength channel. The above image data and feature data were input into a trained neural network convolutional model, which outputs a frying state response value. In this example, the model output value was 0.36, indicating an incomplete result, a manually observed result of incompleteness, and a slightly yellow surface condition.

[0077] It should be noted that, compared with Example 1, the frying time in this example was extended from 3 minutes to 5 minutes, and the surface color of the French fries changed from light to slightly yellow. However, the model output still did not reach the frying endpoint, indicating that the French fries under this condition had not reached the optimal frying endpoint.

[0078] The frying process status and endpoint determination results of the French fry samples obtained in this embodiment are shown in Table 4.

[0079] Example 3 The difference between Example 3 and Example 1 is that the frying conditions for the French fry sample were 150°C for 7 minutes, while the remaining steps and conditions were the same as in Example 1.

[0080] Following the method described in Example 1, color features, texture features, edge features, morphological features, average reflectance, reflectance standard deviation, and channel combination feature parameters are extracted for each wavelength channel. The above image data and feature data are input into a trained neural network convolutional model to obtain the frying state determination result. In this example, the model output value is 0.48, indicating a near-endpoint determination result. Manual observation also indicates a near-endpoint determination result, with the surface state showing uneven coloring.

[0081] It should be explained that after frying at 150℃ for 7 minutes, the French fry sample gradually approached the endpoint from the incomplete state, but its surface color uniformity was still insufficient. Therefore, the model judged it as near the endpoint rather than reaching the endpoint.

[0082] The frying process status and endpoint determination results of the French fry samples obtained in this embodiment are shown in Table 4.

[0083] Example 4 The difference between Example 4 and Example 1 is that the frying conditions for the French fry sample were 170°C for 3 minutes.

[0084] Following the method described in Example 1, color features, texture features, edge features, morphological features, average reflectance, reflectance standard deviation, and channel combination feature parameters were extracted from each wavelength channel. These features, along with 16 wavelength images, were then input into a trained neural network convolutional model. The model outputs the frying state response value of the French fry sample in this example. In this example, the model output value was 0.42, indicating an incomplete judgment result, an incomplete manual observation result, and a surface condition characterized by yellowing.

[0085] It should be explained that, compared with the 150℃, 3min condition, the surface of the French fries turned more yellow under the 170℃, 3min condition, but its comprehensive image features and cross-band response features still did not reach the endpoint state, so it was judged as incomplete.

[0086] The frying process status and endpoint determination results of the French fry samples obtained in this embodiment are shown in Table 4.

[0087] Example 5 The difference between Example 5 and Example 1 is that the frying conditions for the French fry sample were 170°C for 5 minutes.

[0088] Color, texture, edge, and morphological features are extracted from the target area of ​​the French fries, and the average reflectance, standard deviation of reflectance, and channel combination feature parameters are calculated for each wavelength channel. The above image and features are input into a trained neural network convolutional model to obtain the frying state determination result. In this embodiment, the model output value is 0.71, indicating that the frying state has been reached. Both the manual observation result and the result indicate that the frying state has been reached, with the surface exhibiting a golden and crispy appearance.

[0089] It should be explained that, under the conditions of 170℃ and 5min, the color characteristics of the French fry sample were golden yellow, and the texture and edge features did not show obvious over-charring. The cross-band response features matched the features of the samples that reached the endpoint in the training set. Therefore, the model determined that the sample had reached the optimal frying endpoint.

[0090] The frying process status and endpoint determination results of the French fry samples obtained in this embodiment are shown in Table 4.

[0091] Example 6 The difference between Example 6 and Example 1 is that the frying conditions for the French fry sample were 170°C for 7 minutes.

[0092] Color features, texture features, edge features, morphological features, average reflectance, reflectance standard deviation, and channel combination feature parameters are extracted for each wavelength channel. The above image data and feature data are then input into a trained neural network convolutional model. In this embodiment, the model output value is 0.85, indicating an overexposure trend. Both manual observation and surface observation indicate an overexposure trend, with the surface condition showing edge deepening.

[0093] It should be explained that when the frying time was extended to 7 minutes at 170℃, the color of the edge area of ​​the French fries darkened, and the color difference or reflectance difference between the edge area and the center area increased. At the same time, the standard deviation of reflectance under some wavelength channels changed, indicating that the surface uniformity of the French fries decreased and an over-frying trend appeared. Therefore, the model judged it as an over-frying trend. The over-frying trend can be used as an early manifestation of the over-frying state.

[0094] The frying process status and endpoint determination results of the French fry samples obtained in this embodiment are shown in Table 4.

[0095] Example 7 The difference between Example 7 and Example 1 is that the frying conditions for the French fry sample were 190°C for 3 minutes.

[0096] The color, texture, edge, and morphological features, average reflectance, standard deviation of reflectance, and channel combination feature parameters of the target area for the French fries are calculated for each wavelength channel. This data is then input into a trained neural network convolutional model, which outputs a frying state determination result. In this embodiment, the model output value is 0.67, indicating that the frying state has been reached. Both the manual observation and the actual frying state indicate that the frying state has been reached, and the surface condition is characterized by good color.

[0097] It should be explained that, under the conditions of 190℃ and 3 minutes, although the frying time is relatively short, the high frying temperature allows the French fry samples to quickly develop a better color. After integrating the image response features at 16 wavelengths, the model determines that the French fry samples have reached the frying endpoint under these conditions.

[0098] The frying process status and endpoint determination results of the French fry samples obtained in this embodiment are shown in Table 4.

[0099] Example 8 The difference between Example 8 and Example 1 is that the frying conditions for the French fry sample were 190°C for 5 minutes.

[0100] Color features, texture features, edge features, morphological features, average reflectance, reflectance standard deviation, and channel combination feature parameters of the target area of ​​the French fries under different wavelength channels are extracted. The above images and features are then input into a trained neural network convolutional model. In this embodiment, the model output value is 0.91, indicating over-explosion. Manual observation also indicates over-explosion, with the surface showing obvious browning.

[0101] It should be noted that under conditions of 190℃ and 5 minutes, the browning degree on the surface of the French fry sample was significantly enhanced. The proportion of dark pixels, the proportion of dark spots, the color difference between the edge and center regions, and the standard deviation of some wavelength channels all showed over-explosion-related changes. After integrating the above image features and cross-band response features, the neural network convolutional model determined that the sample was in an over-exploded state.

[0102] The frying process status and endpoint determination results of the French fry samples obtained in this embodiment are shown in Table 4.

[0103] Example 9 The difference between Example 9 and Example 1 is that the frying conditions for the French fry sample were 190°C for 7 minutes.

[0104] Color, texture, edge, and morphological features are extracted from each target area of ​​the French fries, and the average reflectance, reflectance standard deviation, and channel combination feature parameters are calculated. The above image data and feature data are input into a trained neural network convolutional model, which outputs a frying state determination result. In this embodiment, the model output value is 0.97, indicating severe overfrying. Manual observation also confirms severe overfrying, with the surface showing obvious charring.

[0105] It should be noted that under the conditions of 190℃ and 7 min, the surface of the French fry sample showed obvious charring, with a further increase in the proportion of dark pixels and the area of ​​dark spots. The color of the edge region was significantly deepened, and the reflectance response and standard deviation under multiple wavelength channels were significantly different from the state at the endpoint. After comprehensive judgment by the neural network convolution model, the sample was determined to be in a severely over-exploded state. The term "severely over-exploded" is a further subdivision of the over-explosion state.

[0106] The frying process status and endpoint determination results of the French fry samples obtained in this embodiment are shown in Table 4.

[0107] Table 4: Results of judging the frying state of French fry samples under different frying conditions Note: The output value of the frying process state determination model includes at least the response values ​​of states such as incomplete, near the endpoint, reached the endpoint, and over-fried; the state with the largest response value is taken as the frying state determination result of the current sample. When the output response value of the "reached the endpoint" state is the largest, or when the output response value of the "reached the endpoint" state reaches the preset threshold, the sample is determined to have reached the optimal frying endpoint; when the output response value of the "incomplete," "near the endpoint," or "over-fried" states is the largest, it is determined to be incomplete, near the endpoint, or over-fried, respectively.

[0108] Examples 1-3 show the quality status assessment results of finished French fry samples after frying at 150℃ for 3 min, 5 min, and 7 min, respectively; Examples 4-6 show the quality status assessment results of finished French fry samples after frying at 170℃ for 3 min, 5 min, and 7 min, respectively; and Examples 7-9 show the quality status assessment results of finished French fry samples after frying at 190℃ for 3 min, 5 min, and 7 min, respectively. In Table 1, the manual observation results are reference assessment results obtained under the same frying conditions based on manual observation or manual annotation. The examples used an image feature-based frying endpoint discrimination model based on 16 LED wavelength illumination conditions for assessment. As shown in Table 1, the multi-wavelength image response model output values ​​of the French fry samples showed significant differences under different frying temperatures and times, which could distinguish between incomplete, near-endpoint, endpoint reached, over-fried, and severely over-fried frying process states. This result indicates that the present invention does not rely on single-color observation, but rather on process response characteristics obtained through multi-wavelength LED active imaging to comprehensively determine the frying state and optimal frying endpoint of the French fries.

[0109] It should be noted that when numerical ranges are mentioned in the claims of this invention, it should be understood that the two endpoints of each numerical range and any value between the two endpoints can be selected. To avoid redundancy, the present invention describes preferred embodiments.

[0110] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0111] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for determining the endpoint of deep-fried French fries, characterized in that, Includes the following steps: Multiple spectral images of fried French fries samples under different frying conditions were acquired at multiple wavelengths to obtain multiple spectral images; Multiple acquired spectral images are preprocessed to eliminate interference from non-target factors, and the target region is extracted from each spectral image. Image features and cross-band response features are extracted from the target region. The cross-band response features include the average reflectance, reflectance standard deviation, and channel combination feature parameters corresponding to each wavelength of the target region. The image features include color features, texture features, edge features, and morphological features. The neural network convolutional model is trained by inputting spectral images at multiple wavelengths, the cross-band response features, and the image features into the neural network convolutional model to obtain a neural network convolutional model for outputting the fried state response. The spectral images of the French fry sample to be tested at multiple wavelengths are acquired, along with the corresponding image features and cross-band response features. The fried state response of the French fry sample to be tested is output using a trained neural network convolutional model.

2. The determination method according to claim 1, characterized in that, The wavelengths are 450 nm, 470 nm, 490 nm, 510 nm, 530 nm, 550 nm, 570 nm, 590 nm, 610 nm, 630 nm, 650 nm, 700 nm, 750 nm, 800 nm, 850 nm and 900 nm.

3. The determination method according to claim 1, characterized in that, The preprocessing includes whiteboard correction, dark current correction, background removal, and target segmentation.

4. The determination method according to claim 1, characterized in that, The color features include the average, standard deviation, maximum, and minimum values ​​of each RGB channel within the target area, the average grayscale value, the standard deviation of grayscale, the R / G color channel ratio, the R / B color channel ratio, the G / B color channel ratio, and the proportion of dark pixels. The texture features include the gray-level variance, gray-level standard deviation, number of dark spots, area ratio of dark spots, local gray-level variation amplitude, texture contrast, texture uniformity, and texture entropy of the target area. The edge features include the number of edge pixels, edge pixel ratio, contour length, average edge gray value, average edge reflectance, and color difference or reflectance difference between the edge region and the center region of the French fry target area. The morphological features include the area, length, width, aspect ratio, perimeter, area of ​​the circumscribed rectangle, rectangularity, circularity, and number of targets in the target region.

5. The determination method according to claim 1, characterized in that, The cross-band response characteristics include the browning characteristic index, the comprehensive color difference characteristic index, and the frying endpoint determination characteristic index: Browning Characteristic Index B: ； Comprehensive color difference characteristic index C: ； F, the characteristic index for determining the frying endpoint: ； , , , , , , , , , , and This represents the average reflectance under the corresponding spectrum.

6. The determination method according to claim 1, characterized in that, The average reflectance, standard deviation of reflectance, and channel combination characteristic parameters corresponding to each wavelength are obtained from the color features, texture features, edge features, and morphological features. The channel combination characteristic parameters include the difference, ratio, and normalized difference between the average reflectance of two or more wavelength channels, as well as the difference or ratio between the standard deviations of different wavelength channels. The cross-band response is characterized by image features, average reflectance, reflectance standard deviation, and channel combination feature parameters across multiple wavelength channels.

7. The determination method according to claim 1, characterized in that, The fried French fries sample was obtained by frying French fries at 150~190℃ for 3~7 minutes.

8. A system for implementing the determination method of claim 1, characterized in that, include: Image acquisition module: contains light sources and imaging units of multiple wavelengths, used to acquire spectral images of French fry samples under illumination conditions of various wavelengths; Image preprocessing module: used for denoising, correcting, and segmenting target regions in the acquired images; Feature extraction module: used to extract color, texture, edge, shape and cross-band response features; Model building and training module: Equipped with a neural network convolutional model, used for learning the extracted features and training the model; State determination module: Used to determine the state of the French fry sample under test using a trained neural network convolutional model and obtain the output result; Display module: Used to display the output results.

Images