A method and system for visual recognition of bottle bases

Abstract

The present application relates to the field of visual detection, and particularly relates to a method and system for visual identification of bottle bottom, the method first acquires single frame or continuous multi-frame bottle bottom image in response to the turning of the packaging bottle bottom into position; then performs edge detection and feature extraction on the bottle bottom image to obtain a feature vector; then, determines whether the feature vector matches the pre-configured template feature; if yes, calculates the coordinates of the target detection object according to the feature vector; if no, puts the bottle bottom image into a preset image feature prediction and compensation model to output the coordinates of the target detection object. Compared with the prior art, the method solves the problems of inaccurate positioning and low efficiency of the existing bottle bottom detection technology.

Description

Technical Field

[0001] This invention relates to the field of visual inspection technology. More specifically, this invention relates to a method and system for visual recognition of bottle bottoms. Background Technology

[0002] Visual inspection technology is commonly used in industrial production lines to detect product defects or process alignment issues. It primarily acquires images using a vision camera and employs feature analysis algorithms to perform defect analysis and obtain inspection results. The application of this technology is mainly based on the assumption that the camera and the target being inspected are relatively stationary. However, in real-world scenarios, the camera itself can vibrate due to equipment vibration; and when the object being inspected suddenly transitions from a moving state to a stationary state, it is inevitably subject to vibration or slippage due to inertia. These vibrations or slippage phenomena lead to blurred visual imaging, resulting in inaccurate or unobtainable inspection results.

[0003] To address the aforementioned issues, Chinese patent application CN115930813A discloses a machine vision-based method and device for anti-micro-shake deformation and displacement monitoring. It primarily uses regression fitting techniques to repair blurry images caused by camera shake. While this technology can improve image clarity to some extent, it still has the following drawbacks: (1) This method is only applicable to scenarios where the camera shakes but the target is stationary, and is not suitable for high-speed rotating packaging bottle inspection lines. In particular, when the packaging bottle is flipped to the bottom, the shaking of the gripping robot or the slippage between the bottle bottom and the platform can cause dynamic blurring or deviation in the visual positioning of the mark point on the bottle bottom. At this time, the object being detected is moving while the background is stationary. Using the above method, the background will be directly used as a "shaky image" for fitting, weakening the features of the bottle bottom image and resulting in inaccurate detection results.

[0004] (2) Although the number of effective frames has increased, the image processing latency will be extended for industrial application systems, which is not conducive to efficient detection and alignment on the production line.

[0005] Therefore, the existing technology mainly suffers from inaccurate bottle bottom detection and low efficiency. Summary of the Invention

[0006] To address the aforementioned technical problems of inaccurate and inefficient bottle bottom detection, this invention discloses a method and system for visual recognition of bottle bottoms.

[0007] In a first aspect, the present invention discloses a method for visual recognition of bottle bottoms, comprising: In response to the bottom of the packaging bottle being flipped into place, a single frame or multiple consecutive frames of bottle bottom images are captured; Edge detection and feature extraction are performed on the bottle bottom image to obtain a feature vector; Determine whether the feature vector matches the pre-configured template features; If so, calculate the coordinates of the target object based on the feature vector; If not, the bottle bottom image is placed into a preset de-shake model for image feature prediction and compensation to output the coordinates of the target object.

[0008] Beneficial effects: The moment the upstream station flips into place, the vision camera begins to quickly photograph the bottom of the bottle, obtaining one or more frames of bottle bottom images. Then, edge detection and feature extraction are performed on the bottle bottom images to obtain feature vectors. Next, the feature vectors are matched with template features using a template comparison method. If a match is found, the coordinates of the target object are directly calculated from the feature vectors, thereby improving efficiency. If a match is not found, it indicates that the bottle bottom image is a dynamic blur or biased image. In this case, a de-shake model is used to predict and compensate for its image features, thereby achieving image restoration in scenarios where the detected object is moving while the background remains stationary. This strengthens the true image features of dynamic blur or biased images, thus improving the accuracy of the detection results.

[0009] Preferably, the de-shake model includes a Kalman filter compensation algorithm and a complex atom model. The bottle bottom image is placed into the preset de-shake model for image feature prediction and compensation, including: The bottle bottom images of multiple consecutive frames are input into the Kalman filter compensation algorithm to predict the corrected image trajectory; Based on the corrected image trajectory, calculate the average motion vector of the bottle bottom image during the exposure time; The average motion vector is input into the complex atom model for image inpainting to obtain the coordinates of the target object.

[0010] Preferably, the complex atom model includes a point spread function and a frequency domain deconvolution algorithm; the average motion vector is input into the complex atom model to obtain the coordinates of the target object, including: The average motion vector is analyzed in two dimensions using the diffusion function to obtain the analytical value; the analytical value is used to stretch the pixels in the bottle bottom image into line segments with angles. The analytical values ​​are converted into a frequency domain representation and then fed into a frequency domain deconvolution algorithm to obtain the de-jitter image; Feature extraction is performed on the de-shaking image to obtain the de-shaking feature vector; the feature vector contains the coordinates of the target object.

[0011] Preferably, edge detection and feature extraction are performed on the bottle bottom image, including: A preset circle-finding algorithm is used to perform edge detection on the bottle bottom image to obtain the image edges; Image edges and their two-dimensional coordinates are encoded and converted into feature vectors.

[0012] Preferably, before the bottom of the packaging bottle is flipped into place, the method of the present invention further includes: A template image for configuring the bottle bottom image; Store the template image in the template database.

[0013] Preferably, determining whether a feature vector matches a pre-configured template feature includes: Extract template features from the template image; The similarity between the template features and the feature vector is calculated to obtain the similarity value; If the similarity value is greater than the similarity threshold, it is judged as yes.

[0014] Preferably, before calculating the similarity between the template features and the feature vector, the method of the present invention further includes: The template images stored in the template database are layered according to a preset pyramid layering configuration; wherein, the pyramid layering configuration is as follows: Lower resolution template images are retrieved first.

[0015] Preferably, the cosine similarity algorithm is used for similarity calculation.

[0016] Preferably, before performing edge detection and feature extraction on the bottle bottom image, the method of the present invention further includes: Denoise and / or enhance the contrast of the bottle bottom image.

[0017] In a second aspect, the present invention also discloses a system for visual recognition of bottle bottoms, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the method for visual recognition of bottle bottoms described in the first aspect is implemented.

[0018] The beneficial effects of this invention are as follows: (1) Compared with the existing technology, for bottle bottom detection, this invention proposes a strategy of "first fast matching, then compensation and repair" to improve the overall detection efficiency. On this basis, if the object being detected is moving while the background is stationary, the image feature prediction and compensation of the blurred image is realized by configuring a de-shaking model, which strengthens the real image features of the dynamically blurred or deviated image, thereby improving the accuracy of the detection results.

[0019] (2) Compared with the prior art, the method of the present invention introduces the search circle algorithm to perform edge detection on the bottle bottom image. When local jitter occurs, although there is an error in a single point due to jitter, when the search circle algorithm fits the circle, these positive and negative local errors will be canceled out at the mathematical calculation level, thereby further improving the accuracy of the detection results.

[0020] (3) Compared with the prior art, the method of the present invention uses template matching for feature matching, which has higher reliability and efficiency. Attached Figure Description

[0021] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein: Figure 1 This is a flowchart of the method for visual recognition of bottle bottoms in Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the visual configuration system in Embodiment 1 of the present invention; Figure 3 This is a diagram showing the effect of successful feature vector matching in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the system structure for visual recognition of bottle bottoms in Embodiment 2 of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] Example 1 like Figure 1 As shown, this embodiment discloses a method for visual recognition of bottle bottoms, including: S10: In response to the bottom of the packaging bottle being flipped into place, acquire a single frame or multiple consecutive frames of bottle bottom images.

[0025] It should be explained that the method in this embodiment is converted into an execution script embedded in the industrial control computer of the inspection station. After the clamping mechanism transports the packaged product to the inspection station and flips it, the industrial control computer drives the vision camera to perform single or continuous shooting actions, thereby obtaining a single frame or multiple consecutive frames of bottle bottom images.

[0026] Furthermore, after obtaining a single frame or multiple consecutive frames of bottle bottom images, image preprocessing is required to enhance features. Image preprocessing includes denoising and / or contrast enhancement of the bottle bottom images.

[0027] In this embodiment, existing mean filtering, Gaussian filtering, or bilateral filtering algorithms can be used. Image denoising can be achieved. Furthermore, image contrast enhancement can be achieved by using automated contrast enhancement wrapper functions in the OpenCV or Pillow libraries.

[0028] It should be noted that in order to provide a specific and reliable matching template for the template comparison mechanism, the following steps also need to be performed before step S10 above: S100: Template image for configuring the bottle bottom image.

[0029] like Figure 2 As shown, the system relies on a visual configuration system to configure corresponding template images for specific models of support products. During this process, key parameters such as positioning accuracy, similarity threshold (out of 1000), camera exposure time, and shooting range also need to be entered.

[0030] It is particularly important to note that this process also requires checking the communication status between the camera and the industrial control computer, and adjusting the camera at the physical level, such as adjusting the camera aperture, preferably f / 2.8-f / 16; the smaller the aperture value, the brighter the image. Aperture adjustment varies depending on the type of bottle. For example, for milky white bottles, the aperture is best adjusted to f / 4.2-f / 5; for other colors, f / 2.8 can be used. This camera aperture adjustment can reduce images that are too dark or too bright, enhance feature visibility, reduce the possibility of duplicate judgments, thereby reducing positioning time and improving efficiency. Regarding the camera focal length, a sample needs to be provided, and the focusing knob rotated until the template image is clear.

[0031] S200: Store the template image in the template database.

[0032] After completing the specific template and parameter configurations, click the save button to save the template image into the template database.

[0033] S300: Layers the template images stored in the template database according to the preset pyramid layering configuration.

[0034] In this pyramid layering configuration, lower-resolution template images are prioritized. More specifically, when constructing the image pyramid, it needs to support hierarchical searching from low to high resolution in the subsequent matching process. This is because the lower the resolution, the fewer template features are required for matching. If a match can be successfully made directly using a lower-resolution template, the efficiency of visual detection can be significantly improved. It should be noted that the similarity threshold configuration differs for template images of different resolutions; the similarity threshold for lower-resolution template images is lower than that for higher-resolution template images.

[0035] S20: Perform edge detection and feature extraction on the bottle bottom image to obtain a feature vector.

[0036] Step S20 above includes: S21: Use a preset circle search algorithm to perform edge detection on the bottle bottom image to obtain the image edge.

[0037] It should be explained that the above circle finding algorithm includes an edge extraction module, the Hough circle transform function, and a least squares fitting algorithm.

[0038] Specifically, for the implementation of step S21, the aforementioned edge extraction module is invoked, which uses the Canny operator to extract the edge points of the bottle bottom image, thereby obtaining the image edge.

[0039] S22: Encode the image edges and their two-dimensional coordinates into feature vectors.

[0040] Specifically, the Hough circle transform function is used to accumulate votes on the center coordinates and radius in the parameter space, mapping edge points in the image space to a three-dimensional parameter space. Candidate circles are then determined through local maxima. Unlike existing localization and bounding box techniques, this function exhibits better robustness to noise and partial occlusion.

[0041] Then, based on the initial localization using the Hough circle transform, a least-squares fitting algorithm is used to select effective edge points of the candidate circle. By minimizing the sum of squares of the difference between the distance from the point to the circle center and the radius, the optimal two-dimensional coordinates of the circle center and its radius are obtained. Compared with existing technologies, this method can achieve sub-pixel-level accurate localization. Simultaneously, methods such as RANSAC are used to remove noise points, further improving the stability of the fitting. Finally, the final localization contour is determined using the two-dimensional coordinates of the circle center and its radius, and the gradient direction and position of each edge point in the final localization contour are calculated, and the feature vector is obtained through encoding.

[0042] By employing the above technical solution to address local jitter, this embodiment uses sub-pixel level edge contour extraction during edge detection and feature extraction, combined with least squares method for geometric fitting of the found circle. Since localized minute jitter is randomly distributed in the image pixel space, the statistical averaging effect of multi-point geometric fitting can naturally offset the coordinate offset error of local pixels. Therefore, if the current frame image only has localized minute jitter, its feature vector can still remain accurate and reliable.

[0043] S30: Determine whether the feature vector matches the pre-configured template features.

[0044] Specifically, step S30 above includes: S31: Extract template features from the template image.

[0045] It should be noted that when step S30 is triggered, a template image is extracted from the template database in sequence according to the pyramid hierarchical configuration.

[0046] In some embodiments, in order to improve the dynamic adaptability of the method of this embodiment, a method similar to step S20 above can be used to extract template features of the template image.

[0047] In other embodiments, in order to improve the efficiency of the method of this embodiment, the edges of the template image can be predefined, and then, when the image is extracted, the gradient direction and position of each edge point of the defined edges can be calculated, and finally the template features can be obtained in the form of encoding.

[0048] In other embodiments, a gradient direction histogram or edge point set is constructed directly for the template image when it is placed.

[0049] S32: Calculate the similarity between the template features and the feature vector to obtain the similarity value.

[0050] In this embodiment, the similarity calculation algorithm used is the cosine algorithm. Specifically, the expression for the cosine algorithm is:

[0051] In the formula, Indicates similarity value; Represents the total number of feature points; cos() represents the cosine function; This represents the edge gradient direction of the i-th feature point in the template features. This represents the edge gradient direction of the i-th feature point in the template features.

[0052] S33: If the similarity value is greater than the similarity threshold, it is determined as yes.

[0053] After calculating the similarity value using the cosine algorithm described above, the similarity threshold corresponding to the template image is retrieved for comparison. If the similarity value is greater than the similarity threshold, it is determined to be true, and step S40 is executed; otherwise, step S50 is executed.

[0054] S40: If so, calculate the coordinates of the target object based on the feature vector.

[0055] In this embodiment, combined with Figure 3As shown, the target detection object can be the two-dimensional coordinates of the center of the final positioning contour, or it can be the center coordinates of industrial positioning marks such as two-dimensional labels, raised / depressed marks, openings, and Mark points within the final positioning contour. Specifically, if the target detection object is not the two-dimensional coordinates of the center, the centroid can be calculated for all pixel two-dimensional coordinates of the industrial positioning mark to obtain the center coordinates as the output result.

[0056] In some embodiments, the vertical diameter of the final positioning contour is used as the offset baseline. The perpendicular line from the center axis of the target object in the Y direction is extracted, and then the acute angle between the offset baseline and the perpendicular line from the center axis in the Y direction is calculated to obtain the offset angle of the industrial positioning mark. Finally, the center coordinates and the offset angle are used together as the output result.

[0057] S50: If not, place the bottle bottom image into the preset de-shake model for image feature prediction and compensation, and output the coordinates of the target detection object.

[0058] Specifically, the aforementioned dejitter model includes a Kalman filter compensation algorithm and a complex atom model. Further, step S50 includes: S51: Input multiple consecutive frames of bottle bottom images into the Kalman filter compensation algorithm to predict the corrected image trajectory.

[0059] It should be further explained that, to further improve efficiency, the bottle bottom image of the last frame in S30 is selected as the matching image. However, in step S51, multiple consecutive bottle bottom images need to be utilized. The specific correction image trajectory is divided into blur length and blur angle. The specific expression for blur length is as follows:

[0060] In the formula, It represents the fuzzy length, which is the relative distance between the bottom of the bottle in one frame and the bottom of the bottle in the next frame; Indicates the exposure time; Indicates the horizontal velocity component; This represents the velocity component in the vertical direction.

[0061] For the fuzzy angle, its expression is:

[0062] In the formula, Indicates an ambiguous angle. This represents the inverse cotangent function.

[0063] S52: Calculate the average motion vector of the bottle bottom image during the exposure time based on the corrected image trajectory.

[0064] Here, the blur length is the modulus of the average motion vector multiplied by the exposure time, and the blur angle is the direction of the average motion vector.

[0065] More specifically, assuming the camera exposure time is T, the average motion velocity vector of multiple frames of bottle bottom images in the image coordinate system is estimated using the Kalman filter compensation algorithm.

[0066] S53: Input the average motion vector into the complex atom model for image inpainting to obtain the coordinates of the target object.

[0067] In this embodiment, the complex atom model includes a point spread function and a frequency domain deconvolution algorithm, and step S53 includes: First, the average motion vector is analyzed in two dimensions using the point spread function to obtain the analytical value. Then, the analytical value is converted into a frequency domain representation and fed into a frequency domain deconvolution algorithm to obtain the de-jitter image.

[0068] It should be noted that the above-mentioned analytical values ​​are used to stretch the pixels in the bottle bottom image into line segments with angles. The feature vector contains the coordinates of the target object being detected.

[0069] More specifically, after obtaining the average motion vector, it needs to be converted into a point spread function (PSF). The point spread function describes how an ideal feature point degenerates (elongates) into a line segment under motion blur. In the spatial domain, the point spread function caused by uniform linear motion is defined as:

[0070] In the formula, This represents the pixel coordinates of a frame of the bottle bottom image; it has a blur length of... And the angle is On the line segment of motion, if the following conditions are met... Given the condition, its pixel weights Pick Otherwise, take 0.

[0071] Subsequently, a Fourier transform needs to be performed on the point spread function in the spatial domain to convert it into a frequency domain degradation transfer function. :

[0072] In the formula, Indicates the integral symbol, The base of the logarithm of natural numbers. It represents the imaginary unit.

[0073] Next, Wiener filtering (a type of frequency domain deconvolution algorithm) is needed for image stabilization compensation. This is because in industrial settings, the bottle bottom images often suffer from motion blur and are also accompanied by system noise from the camera sensor. Directly using inverse filtering for image restoration would amplify high-frequency noise. Therefore, this embodiment preferably uses Wiener filtering as the image restoration algorithm to eliminate motion blur while suppressing noise interference. The frequency domain deconvolution formula for Wiener filtering is as follows:

[0074] In the formula, This indicates that the restored de-jitter image is represented in the frequency domain. The complex conjugate of the degradation transfer function is represented; K is a signal-to-noise ratio adjustment parameter, which is preferably set to 0.01-0.05 in this embodiment; This represents the frequency domain representation of the original bottle bottom image.

[0075] Then, regarding the above By performing an inverse Fourier transform, the de-jitter image in the spatial domain can be obtained. :

[0076] In the formula, This represents the inverse Fourier transform function.

[0077] Finally, feature extraction can be performed on the de-shake image to obtain the de-shake feature vector. The specific process is similar to step S40 above, and will not be repeated here.

[0078] Through steps S51-S53 described above, the method of this embodiment successfully restores blurred / offset images using the Kalman filter compensation algorithm and Fourier transform. It can predict and compensate for image features of blurred images when the detected object is moving while the background remains stationary, by using a de-shake model to enhance the true image features of dynamically blurred or offset images, thereby improving the accuracy of the detection results. Furthermore, step S50 is equivalent to automatically performing reverse derivation and re-judgment using the de-shake model, eliminating the need for manual visual re-judgment and further improving the accuracy of the method in this embodiment.

[0079] It should be added that after the coordinates of the target object are output, they can be directly passed to the next execution station to advance the silkscreen printing, transfer printing, hot stamping, or other processes that require precise positioning for secondary processing of packaging bottles.

[0080] Based on the above technical discussion, and differing from existing technologies, the method in this embodiment offers the following specific performance improvements: Firstly, it can support multi-product, multi-target output and uses a sub-pixel edge extraction algorithm to improve the positioning accuracy of bottle bottom feature points.

[0081] Secondly, it has a real-time tracking function, which can predict the image trajectory using consecutive frames and perform image restoration when re-evaluation is required.

[0082] Third, the latency is less than 80ms, making it particularly suitable for industrial inspection environments with jitter.

[0083] Fourth, the algorithm for finding the circle, the algorithm for edge matching, and the algorithm for Kalman filtering are introduced, which effectively eliminates the impact of slight jitter, smooth jitter, or dynamic blur on positioning accuracy, significantly reduces the deviation failure rate caused by vibration, and improves the stability and reliability of the equipment under complex working conditions.

[0084] Fifth, it can be directly transplanted to process segments that require precise positioning, promoting the high-quality production of industrial products in those segments.

[0085] Example 2 like Figure 4 As shown, this embodiment discloses a system for visual recognition of bottle bottoms, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the method for visual recognition of bottle bottoms described in Embodiment 1 is implemented.

[0086] The system in this embodiment also includes other components well known to those skilled in the art, such as communication interfaces. Their settings and functions are known in the art, and therefore will not be described in detail here.

[0087] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device. Any application or module described in this invention can be implemented using computer-readable / executable instructions that can be stored or otherwise maintained by such a computer-readable medium.

[0088] In the description of this specification, "multiple" means at least two, such as two, three or more, etc., unless otherwise expressly and specifically defined.

[0089] While this specification has shown and described numerous embodiments of the invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Many modifications, alterations, and alternatives will occur to those skilled in the art without departing from the spirit and essence of the invention. It should be understood that various alternatives to the embodiments of the invention described herein may be employed in the practice of this invention.

Claims

1. A method for visual recognition of bottle bottoms, characterized in that, include: In response to the bottom of the packaging bottle being flipped into place, a single frame or multiple consecutive frames of bottle bottom images are captured; Edge detection and feature extraction are performed on the bottle bottom image to obtain a feature vector; Determine whether the feature vector matches a pre-configured template feature; If so, calculate the coordinates of the target object based on the feature vector; If not, the bottle bottom image is placed into a preset de-shake model for image feature prediction and compensation to output the coordinates of the target detection object.

2. The method for visual recognition of bottle bottoms according to claim 1, characterized in that, The de-jitter model includes a Kalman filter compensation algorithm and a complex atom model. The bottle bottom image is placed into the preset de-jitter model for image feature prediction and compensation, including: The Kalman filter compensation algorithm is used to input multiple consecutive frames of bottle bottom images to predict the corrected image trajectory. Based on the corrected image trajectory, calculate the average motion vector of the bottle bottom image during the exposure time; The average motion vector is input into the complex atom model for image inpainting to obtain the coordinates of the target object.

3. The method for visual recognition of bottle bottoms according to claim 2, characterized in that, The complex atom model includes a point spread function and a frequency domain deconvolution algorithm; The average motion vector is input into the complex atom model to obtain the coordinates of the target object, including: The average motion vector input point diffusion function is subjected to two-dimensional analysis to obtain the analytical value; wherein, the analytical value is used to stretch the pixels in the bottle bottom image into line segments with angles. The parsed values ​​are converted into a frequency domain representation and then fed into the frequency domain deconvolution algorithm to obtain the de-jitter image; Feature extraction is performed on the de-jitter image to obtain a de-jitter feature vector; wherein the feature vector contains the coordinates of the target detection object.

4. The method for visual recognition of bottle bottoms according to claim 1, characterized in that, Edge detection and feature extraction are performed on the bottle bottom image, including: The image bottom is edge-detected using a preset circle-finding algorithm to obtain the image edges; The image edges and their two-dimensional coordinates are encoded and converted into the feature vector.

5. The method for visual recognition of bottle bottoms according to claim 1, characterized in that, The method further includes, prior to the bottom of the packaging bottle being flipped into place: Configure the template image for the bottle bottom image; The template image is stored in the template database.

6. The method for visual recognition of bottle bottoms according to claim 5, characterized in that, Determining whether the feature vector matches a pre-configured template feature includes: Extract the template features from the template image; The similarity between the template features and the feature vector is calculated to obtain a similarity value; If the similarity value is greater than the similarity threshold, it is determined as yes.

7. The method for visual recognition of bottle bottoms according to claim 6, characterized in that, Before calculating the similarity between the template features and the feature vector, the method further includes: The template images stored in the template database are layered according to a preset pyramid layering configuration; wherein, the pyramid layering configuration is as follows: Lower resolution template images are retrieved first.

8. The method for visual recognition of bottle bottoms according to claim 6, characterized in that, The similarity calculation uses the cosine similarity algorithm.

9. The method for visual recognition of bottle bottoms according to claim 1, characterized in that, Before performing edge detection and feature extraction on the bottle bottom image, the method further includes: The image of the bottom of the bottle is denoised and / or contrast enhanced.

10. A system for visual recognition of bottle bottoms, characterized in that, It includes a processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the method for visual recognition of bottle bottoms as described in any one of claims 1-9.

Images