Method and determination device for determining a region of interest

Abstract

Method and determination device for determining a region of interest Method for determining a ROI in a captured image of a product moving in a production line, comprising: using a sensor, detecting leading and trailing edges of the product; using an encoder associated with the production line, obtaining a leading encoder count and a trailing encoder count; capturing the captured image of the product using an imaging unit; processing the captured image to identify an identified feature within the captured image; based on the leading encoder count and the trailing encoder count, calculating a calculated width of the product; and determining the ROI in the captured image based on the identified feature and on the calculated width of the product.

Description

[0001] METHOD AND DETERMINATION DEVICE FOR DETERMINING A REGION OF INTEREST

[0002] TECHNICAL FIELD

[0003] The present invention is directed at a method for determining a region of interest (ROI) in a captured image of a product moving in a production line. The present invention is further directed at a determination device for determining a ROI in a captured image of a product moving in a production line.

[0004] BACKGROUND ART

[0005] In a production line, it is often desired to analyze images of the produced products. Such an analysis allows ensuring that each product meets required quality standards, that a visual uniformity in the products is maintained for brand reputation purposes, and that the produced products comply with what has been declared to the tax authorities. To efficiently analyze images of the produced products, it is possible to define a region of interest (ROI) including the most critical parts of the product, and to then focus on analyzing this ROI. Manually configuring ROI settings can be a tedious and timeconsuming process. It is therefore desirable to provide an automatic ROI detection to increase overall production efficiency and reduce costs.

[0006] An objective of the present invention is the provision of an improved determination of a ROI in a captured image of a product moving in a production line.

[0007] SUMMARY OF THE INVENTION

[0008] According to a first aspect, a method for determining a region of interest (ROI) in a captured image of a product moving in a production line is provided. The method comprises: using a sensor placed alongside the production line, detecting a leading edge and a trailing edge of the product; using an encoder associated with the production line, obtaining a leading encoder count corresponding to a count of the encoder when the sensor detects the leading edge, and a trailing encoder count corresponding to a count of the encoder when the sensor detects the trailing edge; capturing the captured image of the product using an imaging unit, wherein the capture of the captured image is triggered by the detection of the leading edge and / or of the trailing edge by the sensor; processing the captured image to identify an identified feature within the captured image; based on the leading encoder count and the trailing encoder count, calculating a calculated width of the product; and determining the ROI in the captured image based on the identified feature and on the calculated width of the product. By considering both the identified feature and the calculated width of the product to determine the ROI, the ROI is determined in a more accurate manner. Identifying the identified feature allows determining an approximate position of specific characteristics of the product, such as the product’s edges or other characteristics. Combining the information about the identified feature with the width allows a more precise determination of the ROI than under consideration of the identified feature alone. In particular, this is because the width allows determining a size of the product in the captured image and considering the width therefore accounts for varying product sizes and varying camera settings. The described method allows determining the ROI for products with varying textures and shapes (by considering the identified features) and sizes (by considering the width), making the method more versatile and adapted to different types of products on the production line. The automated identification of the ROI and calculation of the product dimensions streamline the process and reduce manual intervention, thereby saving precious resources such as time and money. The automated identification of the ROI ensures consistent and reliable ROI detection, reducing the risk of errors or missed areas of interest compared to manual setup.

[0009] The production line can include a series of sequential operations to manufacture a product. The production line can include any process step for preparing, modifying, refining, producing, packaging, controlling and / or transporting the product. Preferably, the production line is provided in a factory manufacturing the product. The product can be any physical item, object or good that is in the process of being prepared, manufactured, modified, refining, transported, controlled and / or packaged in the production line. The product in particular moves in the production line in an automated manner, for example on a conveyor belt of the production line or using a robot of the production line. The production line can be used for multiple products, which are processed closely one after another.

[0010] The ROI can include a most critical part of the product for analyzing the quality, authenticity, and / or compliance with specific rules of the product. For example, the ROI is a portion of the product including a tag of the product (comprising the product name and logo, for example).

[0011] The leading edge of the product can designate the foremost edge of the product. The sensor can detect the leading edge of the product as said leading edge passes the sensor. The trailing edge of the product can designate the rearmost edge of the product. The sensor can detect the trailing edge of the product as said trailing edge passes the sensor. In case of a product having an irregular shape, the leading edge is considered as being the foremost edge of the product at the portion of the product passing in front of the sensor, while the trailing edge is considered as being the rearmost edge of the product at the portion of the product passing in front of the sensor.

[0012] The sensor can be a reflective-type photoelectric sensor for determining the absence, or presence of a product by using a light transmitter (often infrared), and a photoelectric receiver. The sensor can detect the leading edge and the trailing edge without any physical contact with the object. The product can have marks or specific characteristics (shape, color, etc.) that can be detected by the sensor to detect the leading edge and the trailing edge. The sensor may emit an electromagnetic beam and detect the leading edge or the trailing edge when the electromagnetic beam is interrupted or no longer interrupted.

[0013] The encoder can be configured to constantly provide a count as the product moves in the production line. The encoder may be an absolute encoder or an incremental encoder. The encoder can be seen as a clock associated with the production line. The encoder may allow tracking the position of the product at a specific time. The leading encoder count is the count of the encoder at the moment where the sensor detects the leading edge, and the trailing encoder count is the count of the encoder at the moment where the sensor detects the trailing edge.

[0014] The imaging unit can include one or several cameras. The image capture is triggered, i.e. initiated, by the detection of the leading and / or trailing edge. In particular, the imaging unit is programmed to capture an image at a predetermined time after the sensor detects the leading and / or trailing edge, and / or at the exact moment where the sensor detects the leading and / or trailing edge. The imaging unit is in particular provided alongside the production line and pre-oriented to take an image of the product as it passes by the imaging unit on the production line. The captured image of the product moving in the production line is in particular captured without stopping the movement of the product. Preferably, one captured image is captured for each product (with said image being in particular centered around said product). Each captured image might include only one product of multiple moving products.

[0015] To process the captured image to identify the identified feature, the captured image can be run through an image analysis tool, which is for example provided on an industrial PC or on a PLC (programmable logic controller). The identified feature can be any type of characteristic expected in the image and / or in product, such as a specific logo or brand name written on the product. The identified feature can also correspond to the entire product (without the background).

[0016] The calculated width of the product is in particular obtained by calculating a difference between the leading encoder count and the trailing encoder count. In particular, knowing the distance corresponding to one pulse, the difference between the leading encoder count and the trailing encoder count can be converted into a distance (corresponding to the width). A further conversion to convert the distance (which is an absolute value expressed in meters) into a corresponding pixel count in the captured image. The width of the product may be a maximal width of the product and / or a width of the product at a location aligned with the sensor.

[0017] The ROI in the captured image is determined under consideration of both the identified feature and the calculated width of the product. By considering both the identified feature and the calculated width of the product to determine the ROI, the ROI is determined in a more accurate manner. An effort for configuring new production lines, including new cameras, is further reduced as the ROI is detected automatically.

[0018] According to an embodiment, the ROI is a polygon shaped portion, in particular a rectangular portion, of the captured image, the polygon shaped portion having at least four edges being an upper edge, a lower edge, a left edge and a right edge.

[0019] In particular, the right edge is the edge of the ROI that comes closest to the leading edge of the product, and the left edge is the edge of the ROI that comes closest to the trailing edge of the product. The words “right”, “left”, “upper” and “lower” refer to the corresponding right, left, upper and lower parts of the ROI in the orientation of the captured image and in the corresponding orientation of the product in the production line.

[0020] According to a further embodiment, the step of determining the ROI in the captured image based on the identified feature and on the calculated width of the product includes: based on the identified feature, defining positions of the upper and the lower edge of the polygon shaped portion forming the ROI in the captured image; and based on the width of the product, defining the left and the right edge of the polygon shaped portion forming the ROI in the captured image.

[0021] In particular, the upper and lower edges of the ROI are defined as corresponding to upper and lower edges of the identified feature or as being provided at a predetermined distance to the upper and lower edges of the identified feature. The left and right edges of the ROI are at least determined based on the width. For example, the left and right edges of the ROI are determined such that a distance between the left and right edges of the ROI matches the width, is proportional thereto or at a predefined distance therefrom.

[0022] According to a further embodiment, a position of the left and the right edges of the polygon shaped portion are determined based on a position of the identified feature, in particular such that a centroid of the polygon shaped portion is the same as a centroid of the identified feature.

[0023] In particular, the left and right edges of the ROI are determined using the position of the identified feature as a reference, thereby combining the knowledge gained from both the identification of the identified feature and from the calculated width.

[0024] According to a further embodiment, the product is a bottle or a can. In particular, the bottle or the can is for holding beverages. According to a further embodiment, the ROI includes at least a part of a tag characterizing the product.

[0025] The tag characterizing the product is in particular attached to the product. The tag can include a name of the product, a brand name of the product, an associated logo, a description of the product, a price of the product and / or a list of ingredients included in the product. The tag is commonly designated as a “label”, but the term “label” is used in the context of data labelling in the following and the word “tag” is therefore preferred.

[0026] According to a further embodiment, the identified feature is identified based on a histogram of the captured image, based on feature point detection in the captured image and / or based on a Gaussian mixture model.

[0027] Different methods can be used to identify the identified feature. Histograms of the captured image can be histograms of the pixel intensity value or of pixel colors, which can be analyzed to recognize expected patterns and accordingly identify the identified feature. In the feature point detection method, salient points in the image can be analyzed to recognize expected patterns and accordingly identify the identified feature. The Gaussian mixture model can be used to determine different regions of the captured image based on a probability analysis, wherein said regions can form the identified feature or be used to identify the same.

[0028] According to a further embodiment, the product is one of multiple products moving on the production line, wherein neighboring products are touching or nearly touching one another along vertical sides of the product, except for a position of a larger gap between neighboring products, said larger gap being recognized by the sensor in detecting the leading and / or the trailing edge of the product.

[0029] In particular, the products are arranged in a line in the production line, meaning that only one product is aligned with the sensor and / or imaging unit at a time. “Neighboring products” here designates adjacent products, in particular which partly touch or nearly touch one another. “Nearly touching” can mean that there are only few millimeters (less than 0.5cm) between the products, making it difficult for the sensor to distinguish between neighboring products. The larger gap between parts of the neighboring products can correspond to a narrower portion of the products, such as a bottleneck or a recessed part of a can or other product. This larger gap is in particular larger than 0.5cm. The larger gap is large enough to allow for the sensor to distinguish between neighboring products.

[0030] According to a second aspect, a method of training a classification model is provided. Said method comprises: performing the method of the first aspect or any embodiment thereof on multiple captured images of multiple products moving in the production line, thereby determining the ROI of each of the captured images; associating a label to each of the ROIs of the captured images; and training the classification model, which is an artificial intelligence model, using the ROIs and the associated labels.

[0031] All features described in view of the first aspect or any embodiment thereof also hold for the method of the second aspect.

[0032] Multiple captured images with automatically determined ROIs (as described in view of the first aspect and embodiments thereof) can be used to train a classification model. Preferably, the products for which the ROIs are determined are visually similar products but include some differences. For example, the ROIs are determined for multiple products of the same type (for example bottles of a same size) but with varying content. The labels can indicate a characteristic of the product, such as the content of the bottles. The classification model is trained, with the ROIs and the corresponding labels as an input. The purpose of the classification model, once trained, is to receive a ROI as an input and to determine a characteristic of the product (of the same type as the labels used for the training, for example a content of the bottles).

[0033] Since the ROIs are automatically determined in the first aspect and embodiments thereof, enough input data can be provided to the classification model for training. This further allows performing better training of the classification model with an increased number of input data, thereby rendering the classification model more reliable for its classification tasks.

[0034] According to a third aspect, a method for classifying a captured image of a product moving in a production line is provided. The method comprises: performing the method of the first aspect or any embodiment thereof to determine a region of interest (ROI) of the captured image; and predicting a predicted label associated with the product by inputting the ROI of the captured image into a classification model, in particular into the classification model trained according to the method of the second aspect, and by obtaining the predicted label as an output of the classification model.

[0035] All features described in view of the first aspect or any embodiment thereof or in view of the second aspect or any embodiment thereof also hold for the method of the third aspect.

[0036] The ROI determined in the method of the first aspect or any embodiment thereof is input into a trained classification model (preferably the one of the second aspect) to obtain, as an output of the classification model, the predicted label. The predicted label can indicate a characteristic of the product, such as a content of a bottle. Using the ROI to perform classification by the classification model is advantageous in that the image to be classified (i.e. the ROI) does not include any background or unnecessary information, thus decreasing the probability of misclassification.

[0037] According to an embodiment, the method of the third aspect further comprises: comparing the predicted label with a predefined authentic label; and if the predicted label matches the predefined authentic label with a probability that is higher than a predetermined threshold, determining that a type of the product complies with an expected type of product.

[0038] The predefined authentic label can be a label indicating a product characteristic that is expected from the product. Preferably, the expected product characteristic can be the expected type of product. Determining whether the product complies with the expected type of product is particularly advantageous when checking that the product complies with the product type for which a tax has been paid, thereby avoiding fraud.

[0039] According to a fourth aspect, a computer-implemented program product stored on a machine- readable medium and comprising machine-readable instructions for executing the method according to the first, second or third aspect or any embodiment thereof.

[0040] More specifically, the computer-implemented program product comprises instructions for making the sensor, the encoder and the imaging unit function as described above. All features described with respect to the first, second or third aspect equally hold for the computer program product of the fourth aspect.

[0041] According to a fifth aspect, a determination device for determining a region of interest (ROI) in a captured image of a product moving in a production line is provided. The determination device comprises: a sensor configured to be placed alongside the production line and to detect a leading edge and a trailing edge of the product; an encoder associated with the production line, the encoder being configured to obtain a leading encoder count corresponding to a count of the encoder when the sensor detects the leading edge, and a trailing encoder count corresponding to a count of the encoder when the sensor detects the trailing edge; an imaging unit configured to capture the captured image of the product, wherein the capture of the captured image by the imaging unit is triggered by the detection of the leading edge and / or trailing edge by the sensor; a processing unit configured to process the captured image to identify an identified feature within the captured image; a calculation unit configured to calculate a calculated width of the product based on the leading encoder count and the trailing encoder count; and a ROI determination unit configured to determine the ROI in the captured image based on the identified feature and on the calculated width of the product.

[0042] All features described in view of the first aspect or any embodiment thereof also hold for the determination device of the fifth aspect.

[0043] The processing unit, the calculation unit and the ROI determination unit can be part of a processor or a PLC, for example.

[0044] According to an embodiment, the determination device is configured to perform the method according to the first, second or third aspect or of any embodiment thereof.

[0045] The present invention will be described more fully hereinafter with reference to the accompanying figures in which like numerals represent like element throughout the different figures, and in which prominent aspects and features of the invention are illustrated.

[0046] BRIEF DESCRIPTION OF THE FIGURES

[0047] Fig. 1 shows a production system including a production line and a determination device;

[0048] Fig. 2 shows a method for determining a ROI in a captured image of a product;

[0049] Fig. 3 shows an example of a captured image;

[0050] Fig. 4 shows an example of an identified feature in the captured image of Fig. 3;

[0051] Fig. 5 shows an example of a calculated width of a product;

[0052] Fig. 6 shows an example of a determined ROI in the captured image of Fig. 3;

[0053] Fig. 7 shows an example of a training of a classification model;

[0054] Fig. 8 shows an example of a classification of a captured image; and

[0055] Fig. 9 shows an example of neighboring products.

[0056] DETAILED DESCRIPTION

[0057] Fig. 1 shows a production system 50 including a production line 1 and a determination device 30. The determination device 30 is configured to determine a region of interest (ROI) 15 in a captured image 9 of a product 2 moving in the production line 1 by performing the method steps of Fig. 2. The determination of the ROI 15 will be described in the following based on Fig. 1 to 6. The determination device 30 includes a sensor 4, an encoder 5, an imaging unit 6, a processing unit 10, a calculation unit 11, and a ROI determination unit 12. The production line 1 includes a conveyor belt 3 on which cans, which form products 2, are being transported.

[0058] The sensor 4 is placed besides the conveyor belt 3 such that the products 2 pass in front of the sensor 4 one after another as they are transported. The sensor 4 is an infrared photoelectric sensor detecting the presence or absence of a product 2 in front of it. In step S1 of Fig. 2, the sensor 4 detects a leading edge 7 and a trailing edge 8 of the product 2 as said edges 7, 8 pass in front of the sensor 4. The sensor 4 being aligned with inclined portions (recessed portions) located at the top of the cans 2, the difference between the leading edge 7 and the trailing edge 8 does not form the maximum width of the cans 2, but rather indicates a width of the cans 2 at the inclined portions, i.e. aligned with the sensor 4. This is also shown in Fig. 5, which will be described further below. In step S2, the encoder 5 obtains a leading encoder count LEG which is the count of the encoder as the sensor 4 detects the leading edge 7, and a trailing encoder count TEC which is the count of the encoder as the sensor 4 detects the trailing edge 8.

[0059] In step S3 of Fig. 2, the imaging unit 6, which is a camera, captures the image 9 of the product 2 passing it. The timing of the image capture by the imaging unit 6 is achieved by considering the moment at which the sensor 4 detects the leading edge 7 and / or the trailing edge 8. The transmission of the information triggering the image capture is represented by an arrow from the sensor 4 to the imaging unit 6. Fig. 3 shows an example of a captured image 9, which includes an image of a single product 2 and of a background 13.

[0060] In step S4 of Fig. 2, the imaging unit 6 transmits the image 9 to the processing unit 10, which includes an image processing tool. The processing unit 10 can be provided on a computerized machine, such as a computer, optionally together with the calculation unit 11 and the ROI determination unit 12. Once the processing unit 10 receives the image 9, it analyses it to identify an identified feature 14 therein. Fig. 4 shows an example of such an identified feature 14, which is represented in a hashed manner therein. Here, the identified feature 14 is a lightning-shaped logo.

[0061] In step S5 of Fig. 2, the calculation unit 11, which is here a PLC, calculates a calculated width W of the product 2 when aligned with the sensor 4 by calculating a difference between the leading encoder count LEG and the trailing encoder count TEC. The width W is schematically represented in Fig. 5.

[0062] In step S6 of Fig. 2, the ROI 15 (shown in Fig. 6) is determined by the ROI determination unit 15. For this, the identified feature 14 and the calculated width W are used. The identified feature 14 is used to determine an upper edge 16 and a lower edge 18 of the ROI 15. Namely, the upper edge 16 of the ROI 15 is determined as corresponding to the upper edge of the identified feature 14, and the lower (bottom) edge 18 of the ROI 15 is determined as corresponding to the lower edge of the identified feature 14. Further, based on the center of the identified feature 14 and on the width W, a left edge 19 and a right edge 17 of the ROI 15 are determined such that the width W is centered around the center of the identified feature 14.

[0063] The method described above can also be used to determine the ROI 15 of a captured image 9 including several products 2. For example, the feature detection of step s4 allows determining the location of one specific product by identifying the identified feature.

[0064] This ROI 15 can be used for multiple purposes relating to controlling the quality and authenticity of the product 2. Two examples of such applications of the ROI 15 are illustrated in Fig. 7 and 8.

[0065] In the example of Fig. 7, the ROIs 15 of multiple products 2 are obtained following the setup described in view of Fig. 1 to 6. Here, all products 2 are bottles, but there are two different type of bottles: the first bottles containing a first soda and having a lightning shaped tag, and the second bottles containing a second soda and having a rectangular shaped-tag with a circular protrusion in the center, as shown in Fig. 7. The ROIs 15 obtained for the first bottles form a first learning dataset 23 and the ROIs 15 obtained for the second bottles form a second learning dataset 24. Each ROI 15 of the first learning dataset 23 gets a label “first soda” associated therewith while each ROI 15 of the second learning dataset 24 gets a label “second soda” associated therewith. The first learning dataset 23 and the second learning dataset 24, including their associated labels, are used to train a classification algorithm 22.

[0066] In the example of Fig. 8, the classification algorithm 22 is used to classify a ROI 15 obtained in the setup of Fig. 1 to 6 and received as an input to the classification algorithm 22. The classification algorithm 22 indicates whether the ROI 15 is a first soda (“A” in Fig. 8) of a second soda (“B” in Fig. 8).

[0067] Fig. 9 illustrates how the sensor 4 can distinguish between neighboring products 2 on the conveyor belt 3. Said neighboring products 2 are touching each other vertically along vertical sides 20, but due to the shape of the cans 2, there is a recessed portion at the top and bottom of each can 2 forming a larger gap 21, which can be sensed by the sensor 4.

[0068] The above disclosed subject-matter is to be considered illustrative, and not restrictive, and serves to provide a better understanding of the invention defined by the independent claims.

[0069] REFERENCE NUMERALS

[0070] 1 production line

[0071] 2 product

[0072] 3 conveyor belt 4 sensor

[0073] 5 encoder

[0074] 6 imaging unit

[0075] 7 leading edge

[0076] 8 trailing edge

[0077] 9 captured image

[0078] 10 processing unit

[0079] 11 calculation unit

[0080] 12 ROI determination unit

[0081] 13 background

[0082] 14 identified feature

[0083] 15 ROI

[0084] 16 upper edge

[0085] 17 right edge

[0086] 18 lower edge

[0087] 19 left edge

[0088] 20 vertical side

[0089] 21 larger gap

[0090] 22 classification model

[0091] 23 first learning dataset

[0092] 24 second learning dataset

[0093] 30 determination device

[0094] 50 production system

Claims

CLAIMS1. A method for determining a region of interest, ROI, (15) in a captured image (9) of a product (2) moving in a production line (1 ), the method comprising: using (S1 ) a sensor (4) placed alongside the production line (1 ), detecting a leading edge (7) and a trailing edge (8) of the product (2); using (S2) an encoder (5) associated with the production line (1 ), obtaining a leading encoder count corresponding to a count of the encoder (5) when the sensor (4) detects the leading edge (7), and a trailing encoder count corresponding to a count of the encoder (5) when the sensor (4) detects the trailing edge (8); capturing (S3) the captured image (9) of the product (2) using an imaging unit (6), wherein the capture of the captured image (9) is triggered by the detection of the leading edge (7) and / or of the trailing edge (8) by the sensor (4); processing (S4) the captured image (9) to identify an identified feature (14) within the captured image (9); based on the leading encoder count and the trailing encoder count, calculating (S5) a calculated width (W) of the product (2); and determining (S5) the ROI (15) in the captured image (9) based on the identified feature (14) and on the calculated width (W) of the product (2).

2. The method according to claim 1 , wherein the ROI (15) is a polygon shaped portion, in particular a rectangular portion, of the captured image (9), the polygon shaped portion having at least four edges being an upper edge (16), a lower edge (18), a left edge (19) and a right edge (17).

3. The method according to claim 2, wherein the step of determining the ROI (15) in the captured image (9) based on the identified feature (14) and on the calculated width (W) of the product (2) includes: based on the identified feature (14), defining positions of the upper and the lower edge (16, 18) of the polygon shaped portion forming the ROI (15) in the captured image (9); and based on the width (W) of the product (2), defining the left and the right edge (17, 19) of the polygon shaped portion forming the ROI (15) in the captured image (9).

4. The method according to claim 3, wherein a position of the left and the right edges (17, 19) of the polygon shaped portion are determined based on a position of the identified feature (14), in particular such that a centroid of the polygon shaped portion is the same as a centroid of the identified feature (14).

5. The method according to any one of claims 1 to 4, wherein the product (2) is a bottle or a can.

6. The method according to any one of claims 1 to 5, wherein the ROI (15) includes at least a part of a tag characterizing the product (2).

7. The method according to any one of claims 1 to 6, wherein the identified feature (14) is identified based on a histogram of the captured image (9), based on feature point detection in the captured image (9) and / or based on a Gaussian mixture model.

8. The method according to any one of claims 1 to 7, wherein the product (2) is one of multiple products (2) moving on the production line (1 ), wherein neighboring products (2) are touching or nearly touching one another along vertical sides of the product (2), except for a position of a larger gap (21 ) between neighboring products (2), said larger gap (21) being recognized by the sensor (4) in detecting the leading and / or the trailing edge (8) of the product (2).

9. A method of training a classification model, comprising: performing the method of any one of claims 1 to 8 on multiple captured images (9) of multiple products (2) moving in the production line (1 ), thereby determining the ROI (15) of each of the captured images (9); associating a label to each of the ROIs (15) of the captured images (9); and training the classification model (22), which is an artificial intelligence model, using the ROIs (15) and the associated labels.

10. A method for classifying a captured image (9) of a product (2) moving in a production line (1 ), comprising: performing the method of any one of claims 1 to 8 to determine a region of interest, ROI, (15) of the captured image (9); and predicting a predicted label associated with the product (2) by inputting the ROI (15) of the captured image (9) into a classification model (22), in particular into the classification model (22) trained according to the method of claim 8, and by obtaining the predicted label as an output of the classification model (22).

11. The method according to claim 10, further comprising: comparing the predicted label with a predefined authentic label; and if the predicted label matches the predefined authentic label with a probability that is higher than a predetermined threshold, determining that a type of the product (2) complies with an expected type of product (2).

12. A computer-implemented program product stored on a machine-readable medium and comprising machine-readable instructions for executing the method according to any one of claims 1 to 11.

13. A determination device (30) for determining a region of interest, ROI, (15) in a captured image (9) of a product (2) moving in a production line (1), the determination device (30) comprising: a sensor (4) configured to be placed alongside the production line (1 ) and to detect a leading edge (7) and a trailing edge (8) of the product (2); an encoder (5) associated with the production line (1 ), the encoder (5) being configured to obtain a leading encoder count corresponding to a count of the encoder (5) when the sensor (4) detects the leading edge (7), and a trailing encoder count corresponding to a count of the encoder (5) when the sensor (4) detects the trailing edge (8); an imaging unit (6) configured to capture the captured image (9) of the product (2), wherein the capture of the captured image (9) by the imaging unit (6) is triggered by the detection of the leading edge (7) and / or trailing edge (8) by the sensor (4); a processing unit (10) configured to process the captured image (9) to identify an identified feature (14) within the captured image (9); a calculation unit (11) configured to calculate a calculated width (W) of the product (2) based on the leading encoder count and the trailing encoder count; and a ROI determination unit (12) configured to determine the ROI (15) in the captured image (9) based on the identified feature (14) and on the calculated width (W) of the product (2).

14. The determination device according to claim 13, which is configured to perform the method according to any one of claims 1 to 11.

Images