A wristwatch AI intelligent recognition system

Abstract

The present invention discloses a wristwatch AI intelligent recognition system, wherein a target detection model infers a wristwatch region in an input image, identifies a position and a boundary box of the wristwatch, extracts a feature vector of the wristwatch using a feature extraction model, obtains a high-dimensional feature vector describing the appearance of the wristwatch and establishes a vector library, and after a picture of a wristwatch to be recognized is input, similarity is calculated between the feature vector of the picture to be recognized and the feature vector stored in the vector library, a picture most similar to the picture to be recognized is retrieved from the vector library and fed back, and a search is performed using the most matching vector set in the vector library in a "picture-by-picture" manner, and finally a watch image most similar to the input image and a watch model corresponding to the image are returned; this method effectively solves the problem of insufficient accuracy in traditional wristwatch classification and recognition, and provides a new idea for AI in the field of wristwatch recognition.

Description

Technical Field

[0001] The present invention relates to the technical field of wristwatch identification, and in particular to a wristwatch AI intelligent identification system. Background Art

[0002] With the expansion of the watch market, especially in e-commerce platforms and second-hand markets, there are more and more types of watches, and the efficiency and accuracy of manual classification cannot be met. Therefore, the introduction of artificial intelligence technology is particularly important in the issue of watch classification. Deep learning has a good effect in processing complex and large-scale data sets. It can learn similar features of the same category from a large amount of data through the complex structure of neural networks, so as to better classify pictures, and can also identify complex patterns that are difficult to capture with traditional methods.

[0003] At present, there are a lot of brands, styles, design styles, materials and other aspects of watches. Deep learning can automatically identify the features of different brands, styles, functions, designs and other dimensions by learning from watch images. However, in the field of watch recognition, accurate and fast recognition still faces the following problems: When facing a newly launched watch style, the classification model may not be able to predict well when the number of pictures may not meet the requirements of model training. Each training model is limited by computing power and the training time is too long. The collected watch pictures are relatively messy, which is manifested in the uneven distribution of regions and locations of watches. This inconsistency may affect the accuracy of the AI ​​recognition model. Existing feature extraction algorithms often fail to achieve ideal results in vector search when dealing with large-scale classification tasks. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a wristwatch AI intelligent recognition system, which can solve the above-mentioned problems in the existing technology.

[0005] The present invention is achieved through the following technical solutions: a watch AI intelligent recognition system of the present invention includes a feature extraction model, a target detection model, a vector database and an online execution process. The target detection model infers the watch area in the input image, identifies the position and boundary box of the watch, uses the feature extraction model to extract the feature vector of the watch, obtains the high-dimensional features describing the appearance of the watch and establishes a vector library. After inputting the watch picture to be identified, the feature vector of the picture to be identified is calculated for similarity with the feature vector stored in the vector library, and the picture most similar to the picture to be identified is retrieved from the vector library and fed back.

[0006] For a further technical solution, the process of building a feature extraction model is as follows: the first step is to manually label the collected watch pictures with the watch models, with one model being one category; the second step is to select some pictures from the labeled pictures as training sets, validation sets, and test sets; the third step is to train the feature extraction model, training on millions of data based on the transformer or CNN model to obtain the optimal classification model.

[0007] For a further technical solution, the process of building a target detection model is as follows: the first step is to manually annotate the collected watch pictures and manually circle the area on the front of the watch in the picture; the second step is to select some pictures from the annotated pictures as training sets, verification sets and test sets; the third step is to train the target detection model based on YOLO or other neural network models that can be used for target detection, and obtain a target detection model suitable for watches.

[0008] A further technical solution is that if a watch picture does not have a front view of the watch, it will be marked as other.

[0009] For a further technical solution, the process of building a vector database is as follows: in the first step, the target detection model is used to infer the watch area in the input image to identify the position and bounding box of the watch; in the second step, based on these areas, the feature extraction model is used to extract the feature vector of the watch to obtain high-dimensional features describing the appearance of the watch; in the third step, a vector library is established, the extracted watch feature vectors are stored in the vector library, and each feature vector is stored together with the corresponding watch model.

[0010] A further technical solution is to use high-dimensional features for vector retrieval.

[0011] A further technical solution is to build a vector library and quickly find watches similar to the input image by querying the feature vector when performing vector search.

[0012] A further technical solution is that the online execution process includes the following steps: the first step is that after the image to be identified is inferred by the target detection model for the watch area, the system will identify the position of the watch in the image and generate a corresponding bounding box; the second step is to input the calibrated watch area into the feature extraction model for feature extraction to obtain a high-dimensional feature vector of the area; the third step is the similarity search of the image, the feature vector of the image to be identified is calculated for similarity with the feature vector stored in the vector library, the image most similar to the image to be identified is retrieved from the vector library, and these results are fed back.

[0013] According to a further technical solution, the high-dimensional feature vector includes appearance information of the watch.

[0014] As a further technical solution, the similar search results of the pictures provide the user with multiple images that are most similar to the appearance of the target watch, and the watch model corresponding to the pictures is the system prediction result.

[0015] The beneficial effects of the present invention are: 1. In view of the wide variety of watches and the low accuracy of traditional classification methods, the present invention proposes an innovative solution based on the essential characteristics of the picture; through the "image search" method, the most matching vector set in the vector library is used for retrieval, and finally the watch image most similar to the input image is returned; this method effectively solves the problem of insufficient accuracy of traditional watch classification and recognition, and provides a new idea for AI in the field of watch recognition.

[0016] 2. During the research process, the size and position of the watch area in the watch picture may vary. In order to more accurately capture the features of the watch and eliminate the interference of other factors, the present invention applies the target detection technology to the feature extraction link, thereby more finely extracting the feature code of the watch to be identified, avoiding unnecessary feature identification in non-key areas.

[0017] 3. Existing feature extraction algorithms are difficult to effectively express the features of a wide variety of watch images. In order to improve the feature extraction effect, the present invention uses millions of watch images for classification training to find the most suitable feature extraction model. This method is based on the Transformer model for training, and finally obtains more accurate similar search results.

[0018] 4. New watches, antique watches, and limited edition watches cannot be trained because of their small number and scarce pictures. With the method of the present invention, these watches can also be identified, and the recognition accuracy also meets the application requirements. Specifically, the picture features of new watches, antique watches, and limited edition watches are captured, and then similar pictures are searched in the vector library. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] For ease of explanation, the present invention is described in detail with reference to the following specific embodiments and accompanying drawings.

[0020] Figure 1 Schematic diagram of the process of feature extraction and vector library establishment; Figure 2 Schematic diagram of the process of image approximate search. DETAILED DESCRIPTION

[0021] like Figure 1-Figure 2 As shown, the present invention is described in detail. For the convenience of description, the directions mentioned below are defined as follows: the up, down, left, right, front and back directions mentioned below are Figure 1The up-down, left-right, front-back directions of the projection relationship itself are consistent. A wristwatch AI intelligent recognition system of the present invention includes a feature extraction model, a target detection model, a vector database and an online execution process. The target detection model infers the watch area in the input image, identifies the position and boundary box of the watch, uses the feature extraction model to extract the feature vector of the watch, obtains the high-dimensional features describing the appearance of the watch and establishes a vector library. After the watch picture to be identified is input, the feature vector of the picture to be identified is calculated for similarity with the feature vector stored in the vector library, and the picture most similar to the picture to be identified is retrieved from the vector library and fed back. The model and specifications of the watch can be quickly identified by image search.

[0022] Advantageously, the process of building the feature extraction model is as follows: the first step is to manually label the collected watch pictures with the watch models, with one model being one category; the second step is to select some pictures from the labeled pictures as training sets, validation sets and test sets; the third step is to train the feature extraction model, based on the transformer or CNN model (other models can also be used for training), to train millions of data (tens of thousands of categories) to obtain the optimal classification model.

[0023] Advantageously, antique watches, limited edition watches, new watches, etc. are not trained because the images that can be collected are scarce and do not meet the minimum training requirement.

[0024] Advantageously, feature extraction model training is a solution proposed for the problem that existing feature extraction algorithms cannot achieve better similarity search results.

[0025] Advantageously, the process of building the target detection model is as follows: the first step is to manually annotate the collected watch images and manually circle the area on the front of the watch in the image; the second step is to select some images from the annotated images as training sets, validation sets and test sets; the third step is to train the target detection model based on YOLO or other neural network models that can be used for target detection, and obtain a target detection model suitable for watches to determine the position of the watch in the image, so that the vector extraction area is more accurate and irrelevant areas are eliminated.

[0026] Advantageously, if a watch picture does not have a front view of the watch, it is labeled as others, and in this way the pictures are labeled into two categories, namely, watches and others.

[0027] Advantageously, the training of the target detection model proposes a solution to the problem that features of some irrelevant areas are incorrectly encoded during image feature extraction.

[0028] Advantageously, the process of building a vector database is as follows: the first step is feature extraction, first using the target detection model to infer the watch area in the input image to identify the position and bounding box of the watch; the second step is based on these areas, using the feature extraction model to extract the feature vector of the watch to obtain high-dimensional features describing the appearance of the watch; the third step is to establish a vector library, store the extracted watch feature vectors in the vector library, and store each feature vector together with the corresponding watch model.

[0029] Advantageously, high-dimensional features are used for vector retrieval, among others.

[0030] Advantageously, after the vector library is built, when performing a vector search, watches similar to the input image can be quickly found by querying the feature vector, thereby achieving efficient image similarity matching.

[0031] Advantageously, the online execution process includes the following steps: in the first step, after the image to be identified is inferred by the target detection model for the watch area, the system will identify the position of the watch in the image and generate a corresponding bounding box; in the second step, the calibrated watch area is input into the feature extraction model for feature extraction to obtain a high-dimensional feature vector of the area; in the third step, the feature vector of the image to be identified is calculated for similarity with the feature vector stored in the vector library, and the images most similar to the image to be identified are retrieved from the vector library, and these results are fed back.

[0032] Advantageously, the high-dimensional feature vector includes the appearance information of the watch and provides a basis for subsequent similarity search.

[0033] Advantageously, the similar search results of the images can provide the user with several images that are most similar to the appearance of the target watch, and the watch model corresponding to the image is the system prediction result.

[0034] The above are only specific implementation methods of the present invention, but the protection scope of the present invention is not limited to this. Any changes or substitutions that are not conceived through creative work should be included in the protection scope of the present invention; therefore, the protection scope of the present invention should be based on the protection scope defined in the claims.

Claims

1. A wristwatch AI intelligent recognition system, comprising a feature extraction model, a target detection model, a vector database and an online execution process, characterized in that: The object detection model infers the watch area in the input image, identifies the position and bounding box of the watch, uses the feature extraction model to extract the feature vector of the watch, obtains high-dimensional features describing the appearance of the watch and establishes a vector library. After inputting the watch picture to be identified, the feature vector of the picture to be identified is similar to the feature vector stored in the vector library, and the picture most similar to the picture to be identified is retrieved from the vector library and fed back.

2. A wristwatch AI intelligent recognition system according to claim 1, characterized in that: The process of building the feature extraction model is as follows: the first step is to manually annotate the collected watch pictures with the watch models, with one model as one category; the second step is to select some pictures from the annotated pictures as training sets, validation sets, and test sets; the third step is to train the feature extraction model, based on the transformer or CNN neural network model, to train millions of data and obtain the optimal classification model.

3. The wristwatch AI intelligent recognition system according to claim 1, characterized in that: The process of building the target detection model is as follows: the first step is to manually annotate the collected watch pictures and circle the area on the front of the watch in the picture; the second step is to select some pictures from the annotated pictures as training sets, validation sets and test sets; the third step is to train the target detection model based on YOLO or other neural network models that can be used for target detection to obtain a target detection model suitable for watches.

4. The wristwatch AI intelligent recognition system according to claim 3, characterized in that: If a watch image does not include a front view of the watch, it will be marked as Other.

5. The wristwatch AI intelligent recognition system according to claim 1, characterized in that: The process of building a vector database is as follows: the first step is feature extraction. First, the target detection model is used to infer the watch area in the input image to identify the position and bounding box of the watch; the second step is to use the feature extraction model to extract the feature vector of the watch based on these areas, and obtain high-dimensional features that describe the appearance of the watch; the third step is to establish a vector library, store the extracted watch feature vectors in the vector library, and store each feature vector together with the corresponding watch model.

6. A wristwatch AI intelligent recognition system according to claim 5, characterized in that: High-dimensional features are used for vector retrieval.

7. The wristwatch AI intelligent recognition system according to claim 5, characterized in that: After building the vector library, when performing vector search, watches similar to the input image can be quickly found by querying the feature vector.

8. A wristwatch AI intelligent recognition system according to any one of claims 1 to 7, characterized in that: The online execution process includes the following steps: In the first step, after the image to be identified is inferred by the target detection model for the watch area, the system will identify the position of the watch in the image and generate a corresponding bounding box; In the second step, the calibrated watch area is input into the feature extraction model for feature extraction to obtain a high-dimensional feature vector of the area; In the third step, similarity search of images is performed, and the feature vector of the image to be identified is calculated for similarity with the feature vector stored in the vector library, and the image most similar to the image to be identified is retrieved from the vector library, and these results are fed back.

9. A wristwatch AI intelligent recognition system according to claim 8, characterized in that: The high-dimensional feature vector includes the appearance information of the watch.

10. The wristwatch AI intelligent recognition system according to claim 8, characterized in that: The similar search results of the image provide users with multiple images that are most similar to the appearance of the target watch, and the watch model corresponding to the image is the system prediction result.

Images