A method and system for removing duplicates from commodity image recognition results

Abstract

The present invention provides a method and system for deduplicating commodity image recognition results. First, the method quickly filters out dissimilar images using image similarity, reducing the number of images to be processed subsequently, lowering the requirement for computing power, and improving the processing speed. Then, the method adopts a key-point matching strategy to determine the overlapping area, and uses the intersection and union comparison of commodity recognition results within the overlapping area to perform secondary verification on the overlapping area, reducing the misrecognition of the overlapping area caused by incorrect key-point matching and improving the deduplication accuracy. Finally, the method marks the overlapping commodities and deduplicates the commodity recognition summary results according to the marking results, achieving a fast and accurate commodity deduplication effect.

Description

Technical Field

[0001] The present invention relates to the field of computers, and particularly to a method and system for removing duplicates from the recognition results of commodity images. Background Art

[0002] Customer visits are an important task in the fast-moving consumer goods field. Salespersons visit customers, take pictures of the shelves, and use artificial intelligence to detect the commodities in the images to complete tasks such as display detection and stocking rate statistics. However, for many large supermarkets and hypermarkets, there are often many extremely long shelves or horizontal cabinets, and multiple images need to be taken to capture the entire shelf completely. When taking multiple images, there will inevitably be overlapping areas between the images. If the recognition results in the overlapping areas of the images are not de-duplicated, the aggregated results of multiple images will be greater than the actual display results, that is, an accurate commodity stocking details cannot be obtained. In order to achieve de-duplication of commodity images, currently, image stitching is mainly used to obtain panoramic images of shelves or freezers. However, image stitching is highly complex and has strict requirements for taking pictures, which will reduce the visit efficiency of salespersons. Summary of the Invention

[0003] Object of the Invention: The present invention aims to propose a method and system for removing duplicates from the recognition results of commodity images, which can obtain accurate commodity stocking details without reducing the visit efficiency of salespersons.

[0004] Summary of the Invention: To achieve the above object, the present invention proposes the following technical solutions:

[0005] In a first aspect, a method for removing duplicates from the recognition results of commodity images is provided, including:

[0006] Performing image acquisition on a target shelf to obtain a group of commodity images to be detected;

[0007] Inputting the group of commodity images to be detected into a pre-trained commodity recognition model to obtain the commodity recognition results of each commodity image to be detected;

[0008] For each commodity image to be detected, initially determining whether there is an overlapping area based on the similarity between the commodity image to be detected and other commodity images to be detected; taking the commodity images to be detected without overlapping areas as the first commodity images, and taking the remaining commodity images to be detected as the second commodity images;

[0009] For each second commodity image, performing key point matching between the second commodity image and the remaining second commodity images, and determining whether there is an overlapping area based on the key point matching result; for the second commodity images with overlapping areas, determining the position of the overlapping area in the second commodity image based on the key point matching result;

[0010] For two second product images with overlapping regions, respectively count the product sets within the overlapping regions of the two second product images, and calculate the intersection over union ratio of these two product sets;

[0011] If the intersection over union ratio corresponding to the two second product images is greater than the preset intersection over union ratio threshold, then mark the products in the overlapping region of any one of the two second product images as overlapping products;

[0012] Summarize the product recognition results of the product image to be detected, and deduplicate the product recognition summary results based on the marking results of the overlapping products.

[0013] As an optional implementation manner of the method described in the first aspect, for each product image to be detected, based on the similarity between the product image to be detected and other product images to be detected, preliminarily determine whether there is an overlapping region in the product image to be detected, specifically including:

[0014] For each product image to be detected, extract the image features of the product image to be detected;

[0015] Based on the image features of the product image to be detected, calculate the similarity between the product image to be detected and other product images to be detected;

[0016] If the similarity between two product images to be detected is greater than the preset similarity threshold, then preliminarily determine that there is an overlapping region between the two product images to be detected; otherwise, determine that there is no overlapping region between the two product images to be detected.

[0017] As an optional implementation manner of the method described in the first aspect, for each second product image, perform key point matching between the second product image and the remaining second product images, and based on the key point matching results, determine whether there is an overlapping region in the second product image, specifically including:

[0018] For each second product image, perform key point detection on the second product image and extract key point features;

[0019] Based on the key point features, perform key point matching between the second product image and the remaining second product images, and calculate the ratio of the number of successfully matched key points to the number of key points of the second product image as the key point matching value;

[0020] If the key point matching value between two second product images is greater than the preset key point matching threshold, then determine that there is an overlapping region between the two second product images; otherwise, determine that there is no overlapping region between the two second product images.

[0021] Specifically, for the second product image with an overlapping area, determining the position of the overlapping area of the second product image based on the key point matching result specifically includes:

[0022] For the second product image with an overlapping area, calculate the convex hull containing all successfully matched key points, and use the covered area of the convex hull as the overlapping area of the second product image.

[0023] Furthermore, for two second product images with an overlapping area, respectively count the product sets within the overlapping areas of the two second product images, specifically including:

[0024] For two second product images with an overlapping area, calculate the intersection area between each product position box on each second product image and the convex hull on the second product image. If the intersection area is greater than 50% of the area of the corresponding product position box, it indicates that the current product is within the overlapping area of the second product image;

[0025] Respectively count the product sets within the overlapping areas of the two second product images.

[0026] In a second aspect, a duplicate removal system for product image recognition results is provided, including:

[0027] A data acquisition module configured to perform image acquisition on a target shelf to obtain a group of product images to be detected;

[0028] A product recognition module configured to input the group of product images to be detected into a pre-trained product recognition model to obtain product recognition results for each product image to be detected;

[0029] A filtering module configured to, for each product image to be detected, preliminarily determine whether there is an overlapping area for the product image to be detected based on the similarity between the product image to be detected and other product images to be detected; use the product images to be detected without an overlapping area as the first product images, and use the remaining product images to be detected as the second product images;

[0030] A key point matching module configured to, for each second product image, perform key point matching between the second product image and the remaining second product images, and determine whether there is an overlapping area for the second product image based on the key point matching result; for the second product image with an overlapping area, determine the position of the overlapping area of the second product image based on the key point matching result;

[0031] The overlapping product detection module is configured to, for two second product images with an overlapping area, respectively count the product sets within the overlapping area of the two second product images, and calculate the intersection-over-union ratio of the two product sets; if the intersection-over-union ratio corresponding to the two second product images is greater than a preset intersection-over-union ratio threshold, mark the products in the overlapping area of any one of the two second product images as overlapping products;

[0032] The deduplication module is configured to summarize the product recognition results of the product images to be detected, and deduplicate the product recognition summary results based on the marking results of the overlapping products.

[0033] As an optional implementation manner of the system described in the second aspect, the filtering module is specifically configured to:

[0034] For each product image to be detected, extract the image features of the product image to be detected;

[0035] Based on the image features of the product image to be detected, calculate the similarity between the product image to be detected and other product images to be detected;

[0036] If the similarity between two product images to be detected is greater than a preset similarity threshold, preliminarily determine that there is an overlapping area between the two product images to be detected; otherwise, determine that there is no overlapping area between the two product images to be detected.

[0037] As an optional implementation manner of the system described in the second aspect, the key point matching module is specifically configured to:

[0038] For each second product image, perform key point detection on the second product image and extract key point features;

[0039] Based on the key point features, perform key point matching between the second product image and the remaining second product images, and calculate the ratio of the number of successfully matched key points to the number of key points of the second product image as the key point matching value;

[0040] If the key point matching value between two second product images is greater than a preset key point matching threshold, determine that there is an overlapping area between the two second product images; otherwise, determine that there is no overlapping area between the two second product images.

[0041] Specifically, the overlapping product detection module is specifically configured to:

[0042] For a second product image with an overlapping area, calculate the convex hull containing all successfully matched key points, and use the covered area of the convex hull as the overlapping area of the second product image.

[0043] Further, the key point matching module is specifically further configured to:

[0044] For two second commodity images with overlapping regions, calculate the intersection area between each commodity position box on each second commodity image and the convex hull on the second commodity image. If the intersection area is greater than 50% of the area of the corresponding commodity position box, it indicates that the current commodity is located in the overlapping region of the second commodity image;

[0045] Respectively count the set of commodities in the overlapping regions of the two second commodity images.

[0046] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0047] The present invention first uses image similarity to quickly filter out dissimilar (i.e., non-overlapping) images, reducing the number of subsequent processed images, lowering the requirement for computing power, and improving the processing speed. Then, the present invention adopts a key point matching strategy to determine the overlapping region and uses the intersection and union of the commodity recognition results in the overlapping region to perform secondary verification on the overlapping region, reducing the misidentification of the overlapping region caused by incorrect key point matching and improving the deduplication accuracy. Finally, the present invention marks the overlapping commodities and performs deduplication on the summary result of commodity recognition according to the marking result, achieving a fast and accurate commodity deduplication effect.

[0048] When implemented, the present invention has low requirements for photographing, high deduplication accuracy, can meet the needs of various business scenarios, and is particularly suitable for the commodity statistics requirements of ultra-long shelves or horizontal cabinets. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic flowchart of a method for deduplicating commodity image recognition results involved in an embodiment;

[0050] Figure 2 It is a structural diagram of a system for deduplicating commodity image recognition results involved in an embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. However, it should be understood that the present invention can be implemented in various forms. Some exemplary and non-limiting embodiments presented in the following drawings and described below are not intended to limit the present invention to the specific embodiments described.

[0052] It should be understood that, under the condition of technical feasibility, the technical features listed for different embodiments above can be combined with each other to form additional embodiments within the scope of the present invention. In addition, the specific examples and embodiments described in the present invention are non-limiting, and corresponding modifications can be made to the structures, steps, and sequences described above without departing from the protection scope of the present invention.

[0053] Please refer to Figure 1 , this embodiment proposes a method for deduplicating the recognition results of commodity images, and this method includes steps S100 to S110.

[0054] S100: Collect images of the target shelf to obtain a group of commodity images to be detected.

[0055] The above-mentioned group of commodity images to be detected may include one or more images, and the group of commodity images to be detected should include images of all commodities on the target shelf.

[0056] S102: Input the group of commodity images to be detected into a pre-trained commodity recognition model to obtain the commodity recognition results of each commodity image to be detected.

[0057] The above-mentioned commodity recognition results include commodity positions and commodity codes, and the commodity codes are used to represent the categories of commodities.

[0058] Before performing this step, it is necessary to pre-train a commodity recognition model so that the model has the ability to recognize commodity positions and commodity codes from the input commodity images. The above-mentioned commodity recognition model can be implemented by selecting a target detection model, such as target detection algorithms like Faster R-CNN model or SSD. The specific steps for training the commodity recognition model may include:

[0059] (1) Collect commodity image samples in various commodity display or placement scenarios to construct a training sample set.

[0060] (2) Add commodity position box marks and commodity codes to each commodity image sample.

[0061] (3) Input the commodity image samples in the training sample set into the commodity recognition model one by one to obtain the predicted commodity detection boxes and commodity codes. Construct a first loss function based on the predicted commodity detection boxes and the actual commodity position boxes, and construct a second loss function based on the predicted commodity codes and the actual commodity codes. Then, perform a weighted sum of the first loss function and the second loss function to obtain a total loss function, and use this total loss function to update the parameters of the commodity recognition model until a commodity recognition model that meets the requirements is obtained.

[0062] The trained commodity recognition model can be used to recognize the commodity images to be detected to obtain the commodity positions and commodity codes in the commodity images to be detected.

[0063] S104: For each commodity image to be detected, preliminarily determine whether there is an overlapping area in the commodity image to be detected based on the similarity between the commodity image to be detected and other commodity images to be detected; use the commodity images to be detected without overlapping areas as the first commodity images, and use the remaining commodity images to be detected as the second commodity images.

[0064] Specifically, deep learning models such as VGG16, ResNet, and VIT can be used to extract the image features of the commodity images to be detected. Then calculate the similarity of the feature vectors between every two commodity images to be detected. The methods for calculating the similarity of feature vectors include cosine distance, Euclidean distance, etc. Determine whether the similarity between every two commodity images to be detected is greater than a preset similarity threshold. If it is greater than the preset similarity threshold, it means that these two commodity images to be detected are highly similar and there may be an overlapping area, and further judgment is required; otherwise, it means that these two commodity images to be detected are not similar and there is no overlapping area.

[0065] In this step, the similarity between the images to be detected is used to preliminarily screen out the images to be detected that may have overlapping areas. Use the commodity images to be detected without overlapping areas as the first commodity images, and use the remaining commodity images to be detected as the second commodity images. For the commodity images to be detected without overlapping areas, there is no need to perform subsequent steps such as determining the position of the overlapping area and removing duplicates, reducing the number of images to be processed subsequently, reducing the requirement for computing power, and improving the processing speed of the entire solution.

[0066] S106: For each second commodity image, perform key point matching between the second commodity image and the remaining second commodity images, and determine whether there is an overlapping area in the second commodity image based on the key point matching result; for the second commodity image with an overlapping area, determine the overlapping area in the second commodity image based on the key point matching result.

[0067] For the second commodity images with a similarity greater than the similarity threshold, methods such as sift or superpoint can be used to extract image key points and calculate the key point features. Use image key point matching methods such as superglue to perform key point matching between every two second commodity images, and calculate the ratio of the number of successfully matched key points to the number of key points of the corresponding second commodity image. This ratio can be called the key point matching value. If this ratio is greater than the preset key point matching threshold, it means that there is an overlapping area between the two second commodity images; otherwise, it means that there is no overlapping area between the two second commodity images.

[0068] For the second product image with an overlapping region, the specific position of the overlapping region in the second product image can be determined based on the key point matching result. Specifically, for two second product images with an overlapping region, calculate the convex hulls containing all successfully matched key points in these two second product images respectively, which are the overlapping regions on these two second product images.

[0069] S108: For two second product images with an overlapping region, count the product sets in the overlapping regions of these two second product images respectively, and calculate the intersection over union (IoU) of these two product sets; if the IoU corresponding to these two second product images is greater than the preset IoU threshold, then mark the products in the overlapping region of any one of these two second product images as overlapping products.

[0070] Specifically, the product positions and product codes of the products in each image to be detected have been obtained in step S102. In this step, the intersection area between each product position box on the second product image and the convex hull in the second product image can be calculated. If the intersection area is greater than 50% of the area of the product position box, it means that the current product is located in the overlapping region, otherwise, it means that the current product is not in the overlapping region. Count all the product sets located in the overlapping region, and calculate the IoU of the product sets in the overlapping regions of the two second product images with an overlapping region. The calculation method is as follows: Assume that the product set in the overlapping region on image A is {'a': 3, 'b': 2, 'c': 1}, and the product set in the overlapping region on image B is {'a': 2, 'b': 2, 'd': 1}, where 'a': 3 means the quantity of product a is 3, 'b': 2 means the quantity of product b is 2, and so on. Then A ∪ B = {'a': 3, 'b': 2, 'c': 1, 'd': 1}, A ∩ B = {'a': 2, 'b': 2}. Then the IoU of the product sets in the overlapping regions of image A and image B Compare this IoU with the preset IoU threshold, and this IoU threshold can be adaptively set according to requirements, such as it can be set to 50%.

[0071] If the IoU is greater than or equal to the preset IoU threshold, it means that there is indeed product overlap between these two second product images. At this time, mark the products in the overlapping regions of these two second product images to indicate that these products are overlapping products, and exclude these overlapping products in subsequent statistics. More specifically, for two second product images that indeed have an overlapping region, only mark the overlapping products in one of the second product images.

[0072] If the intersection over union (IoU) is less than a preset IoU threshold, it indicates that the previous judgment on the overlapping regions of these two second product images is incorrect. This may be due to the similar scenes and similar displayed products of these two second product images. In this case, no overlapping label processing is performed on these two second product images. By using the IoU of the product sets within the overlapping regions, the misidentification of overlapping regions caused by similar scenes and similar displayed products can be reduced, and the recognition accuracy of overlapping regions can be improved.

[0073] S110: Aggregate the product recognition results of the product images to be detected, and deduplicate the aggregated product recognition results based on the labeling results of overlapping products.

[0074] In step S102, the product recognition results of each image to be detected have been obtained. At this time, these product recognition results can be aggregated to obtain a preliminary aggregated product recognition result.

[0075] Next, according to the overlapping labels in step S108, remove these overlapping products from the preliminary aggregated product recognition result to obtain an accurate aggregated product recognition result.

[0076] Corresponding to the above method for deduplicating product image recognition results, this embodiment also provides a system for deduplicating product image recognition results. This system can be used to implement the above method for deduplicating product image recognition results. As Figure 2 shown, this system includes:

[0077] A data acquisition module, configured to collect images of the target shelf to obtain a set of product images to be detected.

[0078] A product recognition module, configured to input the set of product images to be detected into a pre-trained product recognition model to obtain the product recognition results of each product image to be detected.

[0079] A filtering module, configured to, for each product image to be detected, preliminarily determine whether there is an overlapping region for this product image to be detected based on the similarity between this product image to be detected and other product images to be detected; use the product images to be detected with overlapping regions as the first product images, and use the remaining product images to be detected as the second product images.

[0080] A key point matching module, configured to, for each second product image, perform key point matching between this second product image and the remaining second product images, and determine whether there is an overlapping region for this second product image based on the key point matching results; for the second product images with overlapping regions, determine the position of the overlapping region in this second product image based on the key point matching results.

[0081] The overlapping product detection module is configured to, for two second product images with overlapping regions, respectively count the product sets within the overlapping regions of the two second product images, and calculate the intersection-over-union ratio of these two product sets; if the intersection-over-union ratio corresponding to these two second product images is greater than a preset intersection-over-union ratio threshold, then mark the products within the overlapping region of any one of these two second product images as overlapping products.

[0082] The duplicate removal module is configured to summarize the product recognition results of the product images to be detected, and remove duplicates from the summarized product recognition results based on the marking results of the overlapping products.

[0083] Optionally, the above-mentioned filtering module is specifically used for:

[0084] For each product image to be detected, extract the image features of this product image to be detected;

[0085] Based on the image features of this product image to be detected, calculate the similarity between this product image to be detected and other product images to be detected;

[0086] If the similarity between two product images to be detected is greater than a preset similarity threshold, then preliminarily determine that there is an overlapping region between these two product images to be detected; otherwise, determine that there is no overlapping region between these two product images to be detected.

[0087] Optionally, the above-mentioned key point matching module is specifically used for:

[0088] For each second product image, perform key point detection on this second product image, and extract key point features;

[0089] Based on the key point features, perform key point matching between this second product image and the remaining second product images, and calculate the ratio of the number of successfully matched key points to the number of key points of this second product image as the key point matching value;

[0090] If the key point matching value between two second product images is greater than a preset key point matching threshold, then determine that there is an overlapping region between these two second product images; otherwise, determine that there is no overlapping region between these two second product images.

[0091] Optionally, the above-mentioned overlapping product detection module is specifically used for: for a second product image with an overlapping region, calculate the convex hull containing all successfully matched key points, and use the covered region of the convex hull as the overlapping region of this second product image.

[0092] Optionally, the above-mentioned key point matching module is specifically further used for:

[0093] For two second commodity images with overlapping regions, calculate the intersection area between each commodity position box on each second commodity image and the convex hull of the second commodity image. If the intersection area is greater than 50% of the area of the corresponding commodity position box, it indicates that the current commodity is located within the overlapping region of the second commodity image; respectively count the sets of commodities within the overlapping regions in these two second commodity images.

[0094] The above-mentioned duplicate removal system for commodity image recognition results may include a client and a server in terms of hardware. Among them, the client serves as a data acquisition module for collecting images of the target shelf and uploading the obtained group of commodity images to be detected to the server. The server deploys a commodity recognition module, a filtering module, a key point matching module, an overlapping commodity detection module, and a duplicate removal module. The server uses the above-mentioned method for removing duplicates from commodity image recognition results to count and remove duplicates from the commodity details in the group of commodity images to be detected, and then sends the obtained accurate summary results of commodity recognition to the client for the salesman to view.

[0095] The technical features of the above-described embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as within the scope described in this specification.

[0096] The above-described embodiments merely represent several implementation manners of the present invention, and the description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention patent shall be subject to the appended claims.

Claims

1. A method for removing duplicate product image recognition results, characterized in that: include: Capture images of the target shelf to obtain a group of product images to be detected; Inputting the group of commodity images to be detected into a pre-trained commodity recognition model to obtain commodity recognition results for each commodity image to be detected; For each product image to be detected, based on the similarity between the product image to be detected and other product images to be detected, preliminarily determine whether the product image to be detected has an overlapping area; The product images to be detected without overlapping areas are used as the first product images, and the remaining product images to be detected are used as the second product images; For each second product image, perform key point matching on the second product image and other second product images, and determine whether the second product image has an overlapping area based on the key point matching result; for the second product image with an overlapping area, determine the position of the overlapping area in the second product image based on the key point matching result; For two second product images with overlapping areas, count the product sets in the overlapping areas of the two second product images respectively, and calculate the intersection-and-union ratio of the two product sets; If the IoU corresponding to the two second product images is greater than a preset IoU threshold, marking the products in the overlapping area of ​​any one of the two second product images as overlapping products; The commodity recognition results of the commodity images to be detected are summarized, and the commodity recognition summary results are deduplicated based on the marking results of the overlapping commodities.

2. The method according to claim 1, characterized in that For each product image to be detected, based on the similarity between the product image to be detected and other product images to be detected, it is preliminarily determined whether the product image to be detected has an overlapping area, specifically including: For each product image to be detected, extract the image features of the product image to be detected; Based on the image features of the commodity image to be detected, calculating the similarity between the commodity image to be detected and other commodity images to be detected; If the similarity between the two commodity images to be detected is greater than a preset similarity threshold, it is preliminarily determined that there is an overlapping area between the two commodity images to be detected; otherwise, it is determined that there is no overlapping area between the two commodity images to be detected.

3. The method according to claim 1, characterized in that For each second product image, performing key point matching on the second product image and other second product images, and determining whether there is an overlapping area of ​​the second product image based on the key point matching result, specifically includes: For each second product image, perform key point detection on the second product image and extract key point features; Based on the key point features, the second product image is matched with other second product images in key points, and the ratio of the number of key point matches successfully to the number of key points of the second product images is calculated as the key point matching value; If the key point matching value between the two second product images is greater than a preset key point matching threshold, it is determined that there is an overlapping area between the two second product images; otherwise, it is determined that there is no overlapping area between the two second product images.

4. The method according to claim 3, characterized in that For the second product image having the overlapping area, determining the position of the overlapping area in the second product image based on the key point matching result specifically includes: For the second product image with the overlapping area, the convex hull including all the key points that are successfully matched is calculated, and the coverage area of ​​the convex hull is used as the overlapping area in the second product image.

5. The method according to claim 4, characterized in that For two second product images with overlapping areas, counting the product sets in the overlapping areas of the two second product images respectively includes: For two second product images with overlapping areas, calculate the intersection area of ​​each product location frame on each second product image and the convex hull on the second product image. If the intersection area is greater than 50% of the area of ​​the corresponding product location frame, it indicates that the current product is located in the overlapping area of ​​the second product image. The commodity sets in the overlapping area of ​​the two second commodity images are counted respectively.

6. A system for removing duplicate product image recognition results, characterized in that: include: A data acquisition module is configured to collect images of a target shelf to obtain an image group of commodities to be detected; A commodity recognition module is configured to input the commodity image group to be detected into a pre-trained commodity recognition model to obtain a commodity recognition result of each commodity image to be detected; a filtering module configured to preliminarily determine, for each commodity image to be detected, whether the commodity image to be detected has an overlapping area based on the similarity between the commodity image to be detected and other commodity images to be detected; use the commodity image to be detected without an overlapping area as the first commodity image, and use the remaining commodity images to be detected as the second commodity image; a key point matching module configured to, for each second product image, perform key point matching on the second product image with other second product images, and determine whether the second product images have overlapping areas based on the key point matching results; and for the second product images with overlapping areas, determine the positions of the overlapping areas in the second product images based on the key point matching results; The overlapping product detection module is configured to, for two second product images having overlapping areas, respectively count the product sets in the overlapping areas of the two second product images, and calculate the intersection-and-union ratio of the two product sets; if the intersection-and-union ratio corresponding to the two second product images is greater than a preset intersection-and-union ratio threshold, mark the products in the overlapping area of ​​any one of the two second product images as overlapping products; The deduplication module is configured to summarize the commodity recognition results of the commodity images to be detected, and to deduplicate the commodity recognition summary results based on the marking results of the overlapping commodities.

7. The system according to claim 6, characterized in that The filtering module is specifically used for: For each product image to be detected, extract the image features of the product image to be detected; Based on the image features of the commodity image to be detected, calculating the similarity between the commodity image to be detected and other commodity images to be detected; If the similarity between the two commodity images to be detected is greater than a preset similarity threshold, it is preliminarily determined that there is an overlapping area between the two commodity images to be detected; otherwise, it is determined that there is no overlapping area between the two commodity images to be detected.

8. The system according to claim 6, characterized in that The key point matching module is specifically used for: For each second product image, perform key point detection on the second product image and extract key point features; Based on the key point features, the second product image is matched with other second product images in key points, and the ratio of the number of key point matches successfully to the number of key points of the second product images is calculated as the key point matching value; If the key point matching value between the two second product images is greater than a preset key point matching threshold, it is determined that there is an overlapping area between the two second product images; otherwise, it is determined that there is no overlapping area between the two second product images.

9. The system according to claim 8, characterized in that The overlapping commodity detection module is specifically used for: For the second product image with the overlapping area, the convex hull including all the key points that are successfully matched is calculated, and the coverage area of ​​the convex hull is used as the overlapping area of ​​the second product image.

10. The system according to claim 9, characterized in that The key point matching module is also specifically used for: For two second product images with overlapping areas, calculate the intersection area of ​​each product location frame on each second product image and the convex hull on the second product image. If the intersection area is greater than 50% of the area of ​​the corresponding product location frame, it indicates that the current product is located in the overlapping area of ​​the second product image. The commodity sets in the overlapping areas of the two second commodity images are counted respectively.

Images