Target identification method and system based on computer vision

Abstract

The invention provides a target identification method and system based on computer vision, relates to the technical field of computer vision identification, and realizes accurate detection and identification of text labels in an industrial factory building. The method comprises the steps of receiving an original picture in an RGB format in an industrial factory building as input, performing text region detection on the original picture by adopting an improved DB-net network model, and outputting a detected text region point set and a corresponding minimum bounding rectangular frame; according to the minimum bounding rectangular frame of each text area, intercepting a corresponding text image from the input image, performing character recognition on the intercepted text image by adopting a CRNN network model, and outputting recognized character content and original position information; and carrying out classification screening on the identified text content, distinguishing the labeling content text and the position scale text, and eliminating wrong and invalid information. The target identification system based on the computer vision applies the target identification method based on the computer vision.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision recognition, and more specifically, to a computer vision-based target recognition method and system. Background Art

[0002] With the rapid development of deep learning technology, target detection and recognition technology has been widely used in various fields. However, in specific scenarios such as industrial plants, the targets are highly uniform in appearance, color, and specifications, making it difficult for traditional image feature-based methods to effectively distinguish them. In contrast, text labels, as important identification information of the targets, have become the key to distinguishing and identifying these targets. However, text labels in industrial plants often have challenges such as small fonts, irregular shapes, concentrated distribution, easy adhesion, and being greatly affected by lighting and shooting angles, resulting in problems such as missed detection and misrecognition in text detection and recognition using existing technologies. Summary of the invention

[0003] The purpose of the present invention is to provide a computer vision-based target recognition method and system, which realizes accurate detection and recognition of text labels in industrial plants by combining an improved DB-net network model and a CRNN network model, and adopts a classification and screening mechanism to eliminate errors and invalid information, thereby improving the accuracy and reliability of recognition.

[0004] In order to achieve the above object, the present invention provides a target recognition method based on computer vision, the method comprising:

[0005] Step S1: receiving the original picture in RGB format in the industrial plant as input, using the improved DB-net network model to detect the text area of ​​the original picture, and outputting the detected text area point set and the corresponding minimum circumscribed rectangular frame;

[0006] Step S2: According to the minimum circumscribed rectangular frame of each text area, the corresponding text image is intercepted from the input image, and the text image is recognized by using the CRNN network model, and the recognized text content and original position information are output;

[0007] Step S3: Classify and screen the recognized text content, distinguish the label content text and the position scale text, and eliminate errors and invalid information.

[0008] Preferably, in step S1, the improved DB-net network model is used to perform text area detection on the original image, and the detected text area point set and the corresponding minimum circumscribed rectangular frame are output, including:

[0009] Step S1.1: ResNet is used as the backbone network for feature extraction, and deep convolution calculations are performed on the input original image to obtain feature layers of multiple scales; a feature pyramid network is constructed, and the feature layers of multiple scales are input into the feature pyramid network. Based on the bottom-up path of the feature pyramid network, fractional feature maps of 1 / 2, 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the size of the original image are obtained, and upsampling, fusion, and convolution calculations are performed on the low-resolution feature maps in the fractional feature maps to obtain feature maps of 1 / 4 of the size of the original image;

[0010] Step S1.2: Input the feature map of 1 / 4 the size of the original image obtained in step S1.1 into the CRAFT network for convolution operation to obtain the text level probability map P1 and the character level probability map P2 respectively;

[0011] Step S1.3: Based on the text-level probability map P1, a differentiable binarization process is performed through a learnable adaptive threshold to obtain a preliminary text region, and based on the character-level probability map P2, the text region boundary is refined to obtain a text region mask;

[0012] Step S1.4: Apply morphological processing to the text area mask obtained in step S1.3, execute a contour detection algorithm on the text area mask after morphological processing, extract the contour point set of each text area, and calculate the minimum circumscribed rectangular box of each text area based on the extracted contour point set.

[0013] Preferably, the differentiable binarization processing formula is:

[0014]

[0015] Among them, k is the expansion factor, which is empirically set to 50, P i，j , T i，j are the values ​​of the text-level probability map P1 and the adaptive threshold map at pixel coordinates (i, j), is a robust binary image.

[0016] Preferably, the morphological processing includes a dilation operation and an erosion operation, and the formula of the dilation operation is:

[0017]

[0018] Among them, A is the binary text area mask, C is the structural element, is the expansion operator, is the reflection of C and Z is the pixel position in the output image.

[0019] Preferably, refining the text region boundary based on the character-level probability map P2 to obtain a text region mask includes:

[0020] Adaptive threshold processing is applied to the character-level probability map P2, and pixels above the threshold are regarded as potential character centers.

[0021] Use the minimum spanning tree graph theory method to connect adjacent character centers to form potential character chains;

[0022] According to the connected character centers, the boundary pixels are iterated and expanded outward to form the boundary of the text area;

[0023] A corresponding text area mask is formed based on the text area boundary, and each pixel in the mask is marked as 1 or 0 according to whether it belongs to the text area, where 1 means it belongs and 0 means it does not belong.

[0024] Preferably, the text-level probability map P1 is used to directly identify the text area, and the value of each pixel point is the probability of the point being identified as text;

[0025] The character-level probability map P2 takes each single character as the core, and the value of each pixel is the probability that the point is at the center of a certain character, decreasing from the maximum value at the center of the character to the edge of the character.

[0026] Compared with the prior art, the target recognition method based on computer vision provided by the present invention has the following beneficial effects: The present invention provides a target recognition method based on computer vision, the method comprising: receiving an original picture in RGB format in an industrial plant as input, using an improved DB-net network model to detect text areas on the original picture, and outputting the detected text area point set and the corresponding minimum circumscribed rectangular frame; according to the minimum circumscribed rectangular frame of each text area, intercepting the corresponding text image from the input image, using a CRNN network model to perform text recognition on the intercepted text image, and outputting the recognized text content and original position information; classifying and screening the recognized text content, distinguishing between labeling content text and position scale text, and eliminating errors and invalid information. The present invention realizes accurate detection and recognition of text labels in industrial plants through improved DB-net network models and CRNN network models, effectively solving the problems of missed detection and misrecognition existing in traditional methods.

[0027] The present invention also provides a computer vision-based target recognition system, the system comprising:

[0028] The text area detection module is used to receive the original image in RGB format in the industrial plant as input, use the improved DB-net network model to detect the text area of ​​the original image, and output the detected text area point set and the corresponding minimum external rectangular frame;

[0029] The text recognition module is used to intercept the corresponding text image from the input image according to the minimum circumscribed rectangular frame of each text area, use the CRNN network model to perform text recognition on the intercepted text image, and output the recognized text content and original position information;

[0030] The text classification and screening module is used to classify and screen the recognized text content, distinguish between label content text and position scale text, and eliminate errors and invalid information.

[0031] Preferably, the text recognition module uses an improved DB-net network model to detect text areas on the original image, and outputs the detected text area point set and the corresponding minimum circumscribed rectangular frame, including:

[0032] ResNet is used as the backbone network for feature extraction, and deep convolution calculations are performed on the input original image to obtain feature layers of multiple scales; a feature pyramid network is constructed, and feature layers of multiple scales are input into the feature pyramid network. Based on the bottom-up path of the feature pyramid network, fractional feature maps of 1 / 2, 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the size of the original image are obtained, and upsampling, fusion, and convolution calculations are performed on the low-resolution feature maps in the fractional feature maps to obtain feature maps of 1 / 4 of the size of the original image;

[0033] The feature map with a size of 1 / 4 of the original image is input into the CRAFT network for convolution operation to obtain the text-level probability map P1 and the character-level probability map P2 respectively;

[0034] Based on the text-level probability map P1, a differentiable binarization process is performed through a learnable adaptive threshold to obtain a preliminary text region. Based on the character-level probability map P2, the text region boundary is refined to obtain a text region mask.

[0035] Morphological processing is applied to the text region mask, and a contour detection algorithm is executed on the text region mask after morphological processing to extract the contour point set of each text region. Based on the extracted contour point set, the minimum bounding rectangle of each text region is calculated.

[0036] Preferably, the differentiable binarization processing formula is:

[0037]

[0038] Among them, k is the expansion factor, which is empirically set to 50, P i，j， T i，j are the values ​​of the text-level probability map P1 and the adaptive threshold map at pixel coordinates (i, j), is a robust binary image.

[0039] Preferably, the morphological processing includes a dilation operation and an erosion operation, and the formula of the dilation operation is:

[0040]

[0041] Among them, A is the binary text area mask, C is the structural element, is the expansion operator, is the reflection of C and Z is the pixel position in the output image.

[0042] Compared with the prior art, the beneficial effects of a computer vision-based target recognition system provided by the present invention are the same as the beneficial effects of a computer vision-based target recognition method described in the above technical solution, which will not be elaborated here.

[0043] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0045] Figure 1 A flow chart of a computer vision-based target recognition method provided by an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0046] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0047] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0048] The embodiment of the present invention provides a method for object recognition based on computer vision. Figure 1 A flowchart of a computer vision-based target recognition method provided by an embodiment of the present invention is shown. Figure 1 As shown, the method includes:

[0049] Step S1: receiving the original picture of the industrial plant in RGB format as input, using the improved DB-net network model to detect the text area of ​​the original picture, and outputting the detected text area point set and the corresponding minimum circumscribed rectangular frame.

[0050] It should be noted that in step S1, the improved DB-net network model is used to detect the text area of ​​the original image, and the detected text area point set and the corresponding minimum circumscribed rectangular frame are output, including:

[0051] Step S1.1: ResNet is used as the backbone network for feature extraction, and deep convolution calculations are performed on the input original image to obtain feature layers of multiple scales; a feature pyramid network is constructed, and the feature layers of multiple scales are input into the feature pyramid network. Based on the bottom-up path of the feature pyramid network, fractional feature maps of 1 / 2, 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the size of the original image are obtained, and upsampling, fusion, and convolution calculations are performed on the low-resolution feature maps in the fractional feature maps to obtain feature maps of 1 / 4 of the size of the original image;

[0052] Step S1.2: Input the feature map of 1 / 4 the size of the original image obtained in step S1.1 into the CRAFT network for convolution operation to obtain the text level probability map P1 and the character level probability map P2 respectively;

[0053] Step S1.3: Based on the text-level probability map P1, a differentiable binarization process is performed through a learnable adaptive threshold to obtain a preliminary text region, and based on the character-level probability map P2, the text region boundary is refined to obtain a text region mask.

[0054] Specifically, the text-level probability map P1 is used to directly identify the text area, and the value of each pixel point is the probability of the point being identified as text; the character-level probability map P2 is centered on each single character, and the value of each pixel point is the probability that the point is at the center of a certain character, decreasing from the maximum value at the center of the character to the edge of the character.

[0055] Furthermore, the differentiable binarization processing formula is:

[0056]

[0057] Among them, k is the expansion factor, which is empirically set to 50, P i，j , T i，j are the values ​​of the text-level probability map P1 and the adaptive threshold map at pixel coordinates (i, j), is a robust binary image.

[0058] Furthermore, the text region boundary is refined based on the character-level probability map P2 to obtain a text region mask, including:

[0059] Adaptive threshold processing is applied to the character-level probability map P2, and pixels above the threshold are regarded as potential character centers.

[0060] Use the minimum spanning tree graph theory method to connect adjacent character centers to form potential character chains;

[0061] According to the connected character centers, the boundary pixels are iterated and expanded outward to form the boundary of the text area;

[0062] A corresponding text area mask is formed based on the text area boundary, and each pixel in the mask is marked as 1 or 0 according to whether it belongs to the text area, where 1 means it belongs and 0 means it does not belong.

[0063] Step S1.4: Apply morphological processing to the text area mask obtained in step S1.3, execute a contour detection algorithm on the text area mask after morphological processing, extract the contour point set of each text area, and calculate the minimum circumscribed rectangular box of each text area based on the extracted contour point set.

[0064] Specifically, the morphological processing includes a dilation operation and an erosion operation, and the formula of the dilation operation is:

[0065]

[0066] Among them, A is the binary text area mask, C is the structural element, is the expansion operator, is the reflection of C and Z is the pixel position in the output image.

[0067] Step S2: According to the minimum circumscribed rectangular frame of each text area, the corresponding text image is intercepted from the input image, and the intercepted text image is subjected to text recognition using the CRNN network model, and the recognized text content and original position information are output.

[0068] It should be noted that, first, all potential text regions are detected in the input image using image processing techniques such as edge detection, contour finding or deep learning models such as EAST, CTPN, etc. These regions are usually represented in the form of minimum bounding boxes, which accurately surround the text lines or text blocks in the image. Then, according to the minimum bounding box of each text region obtained in the previous step, the corresponding image region is intercepted from the original input image. This step ensures that the subsequent text recognition operation can focus on the text part in the image, thereby improving the accuracy and efficiency of recognition. The intercepted text image is then sent to the CRNN (Convolutional Recurrent Neural Network) network model for recognition. CRNN combines the advantages of convolutional neural network (CNN) in image feature extraction and the ability of recurrent neural network (RNN) in sequence data modeling, and is particularly suitable for processing text sequences in images. The model can output the recognized text content and may retain or infer the position information of these texts in the original image through coordinate mapping or relative position information. In addition to the text content itself, the CRNN model or subsequent processing steps also need to record or infer the specific location information of each recognized text segment in the original image, such as the coordinates of the upper left corner and the lower right corner.

[0069] Step S3: Classify and screen the recognized text content, distinguish the label content text and the position scale text, and eliminate errors and invalid information.

[0070] It should be noted that before content classification, it is necessary to clearly distinguish the standards for labeling content text and position scale text. Usually based on the semantic features and format of the text content, such as numbers, letter combinations, and the location of occurrence, such as labels on product packaging, scales on instrument dials, etc. All text contents identified in step S2 are analyzed one by one, and their categories are judged by natural language processing (NLP) technology, such as word segmentation, part-of-speech tagging, named entity recognition, etc. or simple rule matching. For example, the position scale text is identified by matching a numeric string with a regular expression, or a specific labeling content is identified by keyword matching. During the classification process, it is also checked whether there are errors in the identified text content, such as OCR recognition errors or invalid information, such as content that is completely inconsistent with the classification standards. These errors and invalid information need to be eliminated to ensure the accuracy and reliability of subsequent processing. The text content after classification and screening is integrated according to the category to provide clear and orderly data support for subsequent applications, such as data analysis, information extraction, report generation, etc. At the same time, the classification standards, rules or model parameters used in the classification process can also be retained as needed for subsequent optimization and adjustment.

[0071] Compared with the prior art, the computer vision-based target recognition method provided by the embodiment of the present invention has the following beneficial effects: the present invention realizes accurate detection and recognition of text labels in industrial plants through the improved DB-net network model and CRNN network model, effectively solving the problems of missed detection and misrecognition existing in traditional methods. The improved text detection algorithm can cope with complex situations such as small fonts, irregular shapes, dense distribution, and easy adhesion, while considering the influence of factors such as lighting and shooting angle, thereby improving the robustness of the system. The modules of the system work together to realize a fully automated process from text detection to recognition, greatly improving processing efficiency and accuracy.

[0072] The embodiment of the present invention further provides a computer vision-based target recognition system, the system comprising:

[0073] The text area detection module is used to receive the original image in RGB format in the industrial plant as input, use the improved DB-net network model to detect the text area of ​​the original image, and output the detected text area point set and the corresponding minimum external rectangular frame;

[0074] The text recognition module is used to intercept the corresponding text image from the input image according to the minimum circumscribed rectangular frame of each text area, use the CRNN network model to perform text recognition on the intercepted text image, and output the recognized text content and original position information;

[0075] The text classification and screening module is used to classify and screen the recognized text content, distinguish between label content text and position scale text, and eliminate errors and invalid information.

[0076] Preferably, the text recognition module uses an improved DB-net network model to detect text areas on the original image, and outputs the detected text area point set and the corresponding minimum circumscribed rectangular frame, including:

[0077] ResNet is used as the backbone network for feature extraction, and deep convolution calculations are performed on the input original image to obtain feature layers of multiple scales; a feature pyramid network is constructed, and feature layers of multiple scales are input into the feature pyramid network. Based on the bottom-up path of the feature pyramid network, fractional feature maps of 1 / 2, 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the size of the original image are obtained, and upsampling, fusion, and convolution calculations are performed on the low-resolution feature maps in the fractional feature maps to obtain feature maps of 1 / 4 of the size of the original image;

[0078] The feature map with a size of 1 / 4 of the original image is input into the CRAFT network for convolution operation to obtain the text-level probability map P1 and the character-level probability map P2 respectively;

[0079] Based on the text-level probability map P1, a differentiable binarization process is performed through a learnable adaptive threshold to obtain a preliminary text region. Based on the character-level probability map P2, the text region boundary is refined to obtain a text region mask.

[0080] Morphological processing is applied to the text region mask, and a contour detection algorithm is executed on the text region mask after morphological processing to extract the contour point set of each text region. Based on the extracted contour point set, the minimum bounding rectangle of each text region is calculated.

[0081] Preferably, the differentiable binarization processing formula is:

[0082]

[0083] Among them, k is the expansion factor, which is empirically set to 50, P i，j , T i，j are the values ​​of the text-level probability map P1 and the adaptive threshold map at pixel coordinates (i, j), is a robust binary image.

[0084] Preferably, the morphological processing includes a dilation operation and an erosion operation, and the formula of the dilation operation is:

[0085]

[0086] Among them, A is the binary text area mask, C is the structural element, is the expansion operator, is the reflection of C and Z is the pixel position in the output image.

[0087] Compared with the prior art, the beneficial effects of a computer vision-based target recognition system provided by an embodiment of the present invention are the same as the beneficial effects of a computer vision-based target recognition method described in the above technical solution, which will not be elaborated here.

[0088] In addition, an embodiment of the present invention further provides an electronic device, including a bus, a transceiver, a memory, a processor, and a computer program stored in the memory and executable on the processor. The transceiver, the memory, and the processor are respectively connected via a bus. When the computer program is executed by the processor, each process of an embodiment of the above-mentioned computer vision-based target recognition method is implemented, and the same technical effect can be achieved. To avoid repetition, it will not be described here.

[0089] In addition, an embodiment of the present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the various processes of the above-mentioned computer vision-based target recognition method embodiment are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0090] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technical solution that can be easily thought of by a person skilled in the art within the technical scope disclosed by the present invention should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A computer vision-based target recognition method, characterized in that: include: Step S1: receiving the original picture in RGB format in the industrial plant as input, using the improved DB-net network model to detect the text area of ​​the original picture, and outputting the detected text area point set and the corresponding minimum circumscribed rectangular frame; Step S2: According to the minimum circumscribed rectangular frame of each text area, the corresponding text image is intercepted from the input image, and the text image is recognized by using the CRNN network model, and the recognized text content and original position information are output; Step S3: Classify and screen the recognized text content, distinguish the label content text and the position scale text, and eliminate errors and invalid information.

2. The computer vision-based target recognition method according to claim 1, characterized in that: In step S1, the improved DB-net network model is used to detect text areas on the original image, and the detected text area point set and the corresponding minimum circumscribed rectangular frame are output, including: Step S1.1: ResNet is used as the backbone network for feature extraction, and deep convolution calculations are performed on the input original image to obtain feature layers of multiple scales; a feature pyramid network is constructed, and the feature layers of multiple scales are input into the feature pyramid network. Based on the bottom-up path of the feature pyramid network, fractional feature maps of 1 / 2, 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the size of the original image are obtained, and upsampling, fusion, and convolution calculations are performed on the low-resolution feature maps in the fractional feature maps to obtain feature maps of 1 / 4 of the size of the original image; Step S1.2: Input the feature map of 1 / 4 the size of the original image obtained in step S1.1 into the CRAFT network for convolution operation to obtain the text level probability map P1 and the character level probability map P2 respectively; Step S1.3: Based on the text-level probability map P1, a differentiable binarization process is performed through a learnable adaptive threshold to obtain a preliminary text region, and based on the character-level probability map P2, the text region boundary is refined to obtain a text region mask; Step S1.4: Apply morphological processing to the text area mask obtained in step S1.3, execute a contour detection algorithm on the text area mask after morphological processing, extract the contour point set of each text area, and calculate the minimum circumscribed rectangular frame of each text area based on the extracted contour point set.

3. The computer vision-based target recognition method according to claim 2, characterized in that: The differentiable binarization processing formula is: Among them, k is the expansion factor, which is empirically set to 50, P i，j , T i，j are the values ​​of the text-level probability map P1 and the adaptive threshold map at pixel coordinates (i, j), is a robust binary image.

4. The computer vision-based target recognition method according to claim 2, characterized in that: The morphological processing includes a dilation operation and a corrosion operation. The formula of the dilation operation is: Among them, A is the binary text area mask, C is the structural element, is the expansion operator, is the reflection of C and Z is the pixel position in the output image.

5. The computer vision-based target recognition method according to claim 2, characterized in that: Based on the character-level probability map P2, the text region boundary is refined to obtain the text region mask, including: Adaptive threshold processing is applied to the character-level probability map P2, and pixels above the threshold are regarded as potential character centers. Use the minimum spanning tree graph theory method to connect adjacent character centers to form potential character chains; According to the connected character centers, the boundary pixels are iterated and expanded outward to form the boundary of the text area; A corresponding text area mask is formed based on the text area boundary, and each pixel in the mask is marked as 1 or 0 according to whether it belongs to the text area, where 1 means it belongs and 0 means it does not belong.

6. The computer vision-based target recognition method according to claim 2, characterized in that: The text-level probability map P1 is used to directly identify the text area, and the value of each pixel is the probability of the point being identified as text; The character-level probability map P2 takes each single character as the core, and the value of each pixel is the probability that the point is at the center of a certain character, decreasing from the maximum value at the center of the character to the edge of the character.

7. A computer vision-based target recognition system, characterized in that: include: The text area detection module is used to receive the original image in RGB format in the industrial plant as input, use the improved DB-net network model to detect the text area of ​​the original image, and output the detected text area point set and the corresponding minimum external rectangular frame; The text recognition module is used to intercept the corresponding text image from the input image according to the minimum circumscribed rectangular frame of each text area, use the CRNN network model to perform text recognition on the intercepted text image, and output the recognized text content and original position information; The text classification and screening module is used to classify and screen the recognized text content, distinguish between label content text and position scale text, and eliminate errors and invalid information.

8. The computer vision-based target recognition system according to claim 7, characterized in that: The text recognition module uses an improved DB-net network model to detect text areas on the original image, and outputs the detected text area point set and the corresponding minimum circumscribed rectangular frame, including: ResNet is used as the backbone network for feature extraction, and deep convolution calculations are performed on the input original image to obtain feature layers of multiple scales; a feature pyramid network is constructed, and feature layers of multiple scales are input into the feature pyramid network. Based on the bottom-up path of the feature pyramid network, fractional feature maps of 1 / 2, 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the size of the original image are obtained, and upsampling, fusion, and convolution calculations are performed on the low-resolution feature maps in the fractional feature maps to obtain feature maps of 1 / 4 of the size of the original image; The feature map with a size of 1 / 4 of the original image is input into the CRAFT network for convolution operation to obtain the text-level probability map P1 and the character-level probability map P2 respectively; Based on the text-level probability map P1, a differentiable binarization process is performed through a learnable adaptive threshold to obtain a preliminary text region. Based on the character-level probability map P2, the text region boundary is refined to obtain a text region mask. Morphological processing is applied to the text region mask, and a contour detection algorithm is executed on the text region mask after morphological processing to extract the contour point set of each text region. Based on the extracted contour point set, the minimum enclosing rectangle of each text region is calculated.

9. The computer vision-based target recognition system according to claim 8, characterized in that: The differentiable binarization processing formula is: Among them, k is the expansion factor, which is empirically set to 50, P i，j , T i，j are the values ​​of the text-level probability map P1 and the adaptive threshold map at pixel coordinates (i, j), is a robust binary image.

10. The computer vision-based target recognition system according to claim 8, characterized in that: The morphological processing includes a dilation operation and a corrosion operation. The formula of the dilation operation is: Among them, A is the binary text area mask, C is the structural element, is the expansion operator, is the reflection of C and Z is the pixel position in the output image.

Images