Super-resolution method for tiny target recognition, electronic equipment and storage medium

Abstract

The invention relates to a super-resolution method for tiny target recognition, electronic equipment and a storage medium, and belongs to the technical field of image processing. In order to solve the problems of low resolution and smooth gray difference of a to-be-detected image, an interpretable gradient-guided image super-resolution model is designed, and high-fidelity SR reconstruction of the low-resolution image is realized; in order to solve the problem that a large amount of interference noise exists in the background, a microcrack pixel-level selection module is designed on the basis of an SR model, and the influence of the interference noise is reduced to a certain extent while high-fidelity reconstruction of a microcrack area is achieved. According to the method, high-fidelity super-resolution reconstruction of weak feature microcrack images such as the surface of the tantalum capacitor end electrode can be realized, the influence of background noise is inhibited, and the recognition accuracy of a tiny target is improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a super-resolution method, electronic equipment, and storage medium for tiny target recognition. Background Art

[0002] In the field of small target recognition, accurately capturing and identifying small targets with weak features has always been a key issue in the industry. Its recognition effect is directly related to many key aspects such as product quality control and industrial production efficiency, and has important engineering application value. However, small targets are extremely small, and the weak features they carry are even more subtle and difficult to discern.

[0003] For example, crack defects on the surface of the terminal electrodes of tantalum capacitors seriously affect the service life of the capacitors. Timely detection of product crack defects during the manufacturing process and reducing the defective rate are important links in ensuring the quality of tantalum capacitors leaving the factory, and have important engineering significance. However, tantalum capacitors are millimeter-level capacitors, and the weak characteristic microcracks they produce are even smaller. Since the selection of lens magnification needs to consider the integrity of the terminal electrode imaging and the constraints of the actual detection environment, the resolution of the image to be detected is low, making it difficult to directly identify weak characteristic microcrack defects through the detection algorithm. If the resolution of the image during the recognition process is improved by modifying the hardware, the cost is high, and the actual manufacturing process and application will also be limited. At the same time, the background noise on the surface of the terminal electrode of the tantalum capacitor itself will have a significant impact on the subsequent recognition algorithm, further increasing the possibility of false detection. Summary of the Invention

[0004] To address the problems of low image resolution and high noise in small target recognition tasks, and to achieve the acquisition of high-fidelity and high-resolution images to be detected, the present invention provides a super-resolution method, electronic device, and storage medium for small target recognition.

[0005] The present invention first provides a visual gradient interpretable super-resolution method for small target recognition, comprising the following steps: Step 1: Obtain the image to be reconstructed; Step 2: construct an interpretable gradient-guided image super-resolution network model IGGS, and use the IGGS model to perform super-resolution reconstruction on the image to be reconstructed; The IGGS model includes a feature extraction unit, a feature mapping module GGFM guided by gradient information, an upsampling unit, a pixel reconstruction unit, and a global residual unit. The expression of the forward propagation process of the IGGS model is:

[0006] Where, represents the image to be reconstructed; represents the super-resolution image output by the IGGS model; represents the upsampling unit; Represents the network model feature fitting function; represents the feature extraction unit; Represents the GGFM module; represents the pixel reconstruction unit; The forward propagation process expression of the GGFM module is:

[0007] represents the input features, Represents a feature mapping unit, which is used to learn and fit high-frequency features in images. Represents the gradient operation unit, which is used to obtain image gradients in different directions.

[0008] Optionally, the gradient differential operator expression in the gradient operation unit is:

[0009] Where, 、 、 、 Represents the gradient differential operators of the image in the horizontal direction, vertical direction, upper left to lower right direction, and upper right to lower left direction respectively.

[0010] Optionally, the method further includes: For the image output by the IGGS model, horizontal texture noise is filtered out using a horizontal mean filter operator; The filtered image is convolved with four microcrack line detection operators in different directions, and the maximum value of the four detection results is taken to obtain the line detection response of the microcrack; A dilation operation is performed on the microcrack line detection response result, and the expanded image is bilinearly interpolated to make its resolution consistent with the output feature layer of the pixel reconstruction unit. It is then activated by a selective response activation function designed based on the hyperbolic tangent function, and finally multiplied pixel-wise with the output feature of the pixel reconstruction unit to obtain the final processed image.

[0011] Optionally, the feature extraction unit includes three convolutional layers and two activation layers.

[0012] Optionally, the upsampling unit adopts a bilinear interpolation method.

[0013] Optionally, the pixel reconstruction unit adopts a single convolution layer.

[0014] Optionally, the calculation expression of the feature extraction unit is:

[0015] in, Represents the output features of the feature extraction unit; represents the convolutional layer; Represents the ReLU activation function.

[0016] Optionally, the bilinear interpolation method includes: Know the target point required Four adjacent coordinates and corresponding pixel values: 、 、 and , linear interpolation is performed in the horizontal direction to obtain and The specific calculation formula for the pixel value is as follows:

[0017]

[0018] Where, 、 、 、 、 、 Represents the coordinate points 、 、 、 、 、 Pixel value of After completing the horizontal linear interpolation through the above formula, then through and The coordinate value of is interpolated with the obtained pixel value in the vertical direction, and the expression is:

[0019] Will and After the pixel value of the target point is substituted into the above formula, Pixel value Can be obtained from this.

[0020] The present invention provides a small target recognition method, which uses any of the above-mentioned super-resolution methods for small target recognition to perform super-resolution reconstruction on an image to be recognized, and then performs target recognition based on the reconstructed image.

[0021] The present invention provides an electronic device, characterized by comprising a memory and a processor; The memory is used to store computer programs; The processor is configured to implement any of the above methods when executing the computer program.

[0022] The present invention provides a computer-readable storage medium, characterized in that a computer program is stored on the storage medium, and when the computer program is executed by a processor, the method described in any one of the above items is implemented.

[0023] Beneficial effects: Aiming at the recognition of small targets, the present invention constructs an interpretable gradient-guided image super-resolution network model to perform super-resolution reconstruction of the image to be identified, introduces a feature mapping module GGFM guided by gradient information, performs feature mapping and gradient operation on the input features at the same time, then integrates the obtained gradient information into the mapping feature layer, and then performs a local residual operation. This module performs further nonlinear mapping processing on the input feature layer, increases the network depth so that the network model has better fitting mapping ability, and at the same time, the structure of the local residual further reduces the fitting difficulty of the mapping module; in addition, the present invention introduces a gradient operation unit and customizes the gradient differential operator. Under the convolution operation of the gradient differential operator, the gradient information of the image containing direction and amplitude can be obtained, which effectively solves the problems of low resolution and smooth grayscale difference of the image to be detected, and realizes high-fidelity super-resolution reconstruction of low-resolution images.

[0024] In one embodiment of the present invention, a microcrack pixel-level selection module can be designed and introduced based on the IGGS model, which realizes high-fidelity reconstruction of the microcrack area while reducing the influence of interference noise to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a flow chart of the method for correcting the end electrode image in Example 3 of the present invention.

[0026] Figure 2 This is a flow chart of a method for segmenting an end electrode image in embodiment three of the present invention.

[0027] Figure 3 This is a structural diagram of the interpretable gradient-guided image super-resolution network model in Example 2 of the present invention.

[0028] Figure 4 This is a structural diagram of the feature mapping module guided by gradient information in Example 2 of the present invention.

[0029] Figure 5 This is a structural diagram of the micro-crack pixel-level selection module in the second embodiment of the present invention.

[0030] Figure 6 This is a flowchart of obtaining gradient information in multiple directions in the second embodiment of the present invention.

[0031] Figure 7 Schematic diagram of the bilinear interpolation method in the second embodiment of the present invention.

[0032] Figure 8 It is the structure diagram of the residual network module. DETAILED DESCRIPTION

[0033] To make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0034] Example 1: This embodiment provides a super-resolution method for small object recognition, including: Step 1: Obtain the image to be reconstructed; Step 2: Construct an interpretable gradient-guided image super-resolution network model IGGS, and use the IGGS model to perform super-resolution reconstruction on the image to be reconstructed; The IGGS model includes a feature extraction unit, a gradient-guided feature mapping module GGFM, an upsampling unit, a pixel reconstruction unit, and a global residual unit. The forward propagation process of the IGGS model is expressed as:

[0035] Where, represents the image to be reconstructed; Represents the super-resolution image output by the IGGS model; represents the upsampling unit; Represents the network model feature fitting function; represents the feature extraction unit; Represents the GGFM module; represents pixel reconstruction unit; The forward propagation process expression of the GGFM module is:

[0036] represents the input features, Represents a feature mapping unit, which is used to learn and fit high-frequency features in images. Represents the gradient operation unit, which is used to obtain image gradients in different directions.

[0037] Example 2: This embodiment provides a visual gradient interpretable super-resolution method for small object recognition. An interpretable gradient guided image super-resolution network model (IGGS) is constructed. The model consists of a feature extraction unit, a gradient-guided feature mapping module (GGFM), an upsampling unit, a pixel reconstruction unit, and a global residual unit. The overall model structure is as follows: Figure 3 As shown in Figure 1, it mainly consists of convolutional layers and activation layers. The specific parameters are shown in Table 1.

[0038] Table 1 Parameters of the interpretable gradient-guided super-resolution network model

[0039] The purpose of the above IGGS model is to enhance the resolution of the image to be detected. The image to be detected will pass through the above units in sequence to achieve super-resolution reconstruction. The expression of the forward propagation process of the overall network model is:

[0040] Where, A low-resolution image representing the input to the network model; Represents the super-resolution image output by the network model; Represents the forward propagation function of the proposed super-resolution reconstruction network model (Super-Resolution, SR); Represents the upsampling unit in the model; Represents the network model feature fitting function; Represents the feature extraction unit in the network model; Represents the gradient-guided feature mapping module in the network model; Represents the pixel reconstruction unit in the network model.

[0041] Each unit of the IGGS model is introduced below.

[0042] (1) Feature extraction unit The feature extraction unit first directly convolves the low-resolution image to extract features into a vector. This vector then contains multiple feature maps containing abstract feature information. The activation layer selects and acquires the extracted features, while also enhancing the network's nonlinear fitting capabilities. To ensure that the feature extraction unit can better capture detailed features, each convolutional layer uses a convolution kernel size of 3×3 and a stride of 1, which fully ensures the meticulousness of feature extraction. The feature extraction unit consists of three convolutional layers and two activation layers. The calculation expression is:

[0043] in, Represents the output features of the feature extraction unit; Represents a convolutional layer with a kernel size of 3×3 and a stride of 1; Represents the Rectified Linear Unit (ReLU), which is the ReLU activation function.

[0044] (2) Gradient-guided feature mapping module (GGFM) The GGFM module is the core part of the IGGS model. It adds the gradient information of the feature layer in the process of feature nonlinear mapping. At the same time, the local residual structure further reduces the fitting difficulty of the mapping module. The proposed SR network contains two feature mapping modules, which are located before and after the upsampling unit, such as Figure 3 The function of the GGFM module is to further perform nonlinear mapping on the input feature layer, increase the network depth and thus make the network model have better fitting mapping capabilities.

[0045] The GGFM module includes a feature mapping unit, a gradient operation unit, and a local residual unit, such as Figure 4 The feature mapping network module structure consists of a trainable convolutional layer, a custom convolutional layer, and an activation layer. The overall network model contains two feature mapping network modules. The input feature space sizes of the front and back modules are 150×150 and 300×300, respectively. Taking the first feature mapping network module as an example, the specific parameters are shown in Table 2.

[0046] Table 2 Parameters of feature mapping module guided by gradient information

[0047] The GGFM module performs feature mapping and gradient operations on the input features at the same time, then integrates the obtained gradient information into the mapping feature layer, and then performs local residual operations. The forward propagation process expression of the GGFM module is:

[0048] Where, represents the input features of the feature mapping network module, represents the output features of the feature mapping network module, Represents the operation of the gradient calculation unit.

[0049] The feature mapping unit plays a role in learning and fitting within the feature mapping network module. After the input features undergo nonlinear fitting by the feature mapping unit, the result is calculated using the same set of convolution kernels. The network structure of the feature mapping unit is consistent with the feature extraction unit mentioned above, both using a convolution kernel of size 3×3 and a stride of 1. This aims to improve the sophistication of feature extraction and learning and fitting. The feature mapping unit primarily learns and fits high-frequency features in the image, specifically the differences between low-resolution and high-resolution images.

[0050] The pixel gradient information in an image has multiple directions, and the gradient values ​​in different directions have different degrees of difference. That is, the gradient information of image pixels has vector characteristics, but the difficulty lies in how to determine the gradient guidance information required for each pixel in the image. Therefore, to solve the above problem, this embodiment uses a multi-directional custom gradient convolution kernel to obtain gradient information in multiple directions, such as Figure 6 shown.

[0051] Each pixel in the physical real world has gradient information in an infinite number of directions, but in an image, it has been discretized into a dot matrix representation. Therefore, the gradient information of each pixel in the image can be divided into eight directions, and the Kirsch differential operator in image processing includes edge detection in eight specific directions. At the same time, the image gradient truly reflects the grayscale changes at the high-frequency features of the image, and its essence is the derivation process of a two-dimensional discrete function. The first-order differential operator can obtain edge points and edge segments in the image, and at the same time describe the process of edge direction, while the second-order differential operator can also determine the high-frequency information of the edge. In addition, the second-order differential operator is more sensitive to noise-interfered pixels in the image, and only generates edges in high-frequency pixel areas. Therefore, in the gradient operation unit, the custom gradient differential operator used in this embodiment is designed based on the first-order differential operator, and its expression is:

[0052] Where, 、 、 、 These represent the image's gradient differential operators in the horizontal, vertical, upper left to lower right, and upper right to lower left directions, respectively. Convolution with these operators yields information about the image's gradient, including both direction and magnitude. Essentially, this involves convolving the image with a set of convolution masks to obtain the gradient value.

[0053] (3) Upsampling unit The role of the upsampling unit in the interpretable gradient-guided image super-resolution network model is to increase the resolution of the image, and then combine with other units to achieve high-fidelity SR reconstruction of the image. The upsampling unit in the SR network model proposed in this embodiment is bilinear interpolation. Figure 7 shown.

[0054] Know the target point required Four adjacent coordinates and corresponding pixel values: 、 、 and ,like Figure 7 First, linear interpolation is performed in the horizontal direction to obtain and The specific calculation formula for the pixel value is as follows:

[0055]

[0056] Where, 、 、 、 、 、 Represents the coordinate points 、 、 、 、 、 After completing the horizontal linear interpolation through the above formula, and The coordinate value of is interpolated with the obtained pixel value in the vertical direction, and the expression is:

[0057] Will and After the pixel value of the target point is substituted into the above formula, Pixel value After the feature layer passes through the upsampling unit, the resolution of all its layers is scaled to twice the original resolution and then passed to the subsequent units for processing.

[0058] (4) Pixel reconstruction unit After feature extraction, upsampling, and nonlinear mapping, the low-resolution image's features need to be restored to a high-resolution image. Finally, pixel-level loss calculations are performed with the labeled image, and the model parameters are optimized using the backpropagation algorithm. Therefore, the role of the pixel reconstruction unit is to reconstruct multiple layers of abstract features into a high-resolution image. This is achieved by a single convolutional layer, as expressed in:

[0059] Where, Represents the output features of the feature mapping module guided by gradient information before the pixel reconstruction unit, Represents the output features of the pixel reconstruction unit. The features output by the pixel reconstruction unit are the feature information that the entire network model needs to learn and fit. The channel of this feature layer is 1, which contains a large amount of high-frequency information in the label image.

[0060] (5) Global residual unit In the SR reconstruction task, the network model does not need to learn to fit the low-frequency information that is common to high-resolution and low-resolution images, but only needs to learn different high-frequency information. Therefore, the essential task of the SR network model is to learn to fit the high-frequency information, thereby realizing the SR reconstruction process of the low-resolution image. Figure 8 As shown in the figure, it can be divided into two parts: identity mapping and residual mapping. After the input features are operated by the residual block, the features of the residual part are obtained. The input features are directly used as the features of the identity mapping part and are element-wise summed with the residual mapping features.

[0061] After summing the residual mapping feature and the identity mapping feature, the observation value can be obtained. At this time, the observation value can be used as the input feature of the subsequent unit or the final output feature. The expression is:

[0062] Where, Represents the residual mapping feature in the residual block, corresponding to the output feature of the low-resolution image obtained by the SR network of this embodiment ,like Figure 8 As shown in (a) of Represents the identity mapping feature in the residual block, corresponding to the network model in this embodiment ,like Figure 8 In addition, the residual block in the SR reconstruction task can not only reduce the difficulty of learning and fitting the network model, but also avoid the problem of gradient disappearance in the network model to a certain extent. Its derivative is:

[0063] During network model training, the gradient of the residual part will not remain at -1, which to some extent avoids the problem of vanishing gradients. Therefore, the residual block in the SR network model allows the network model to only learn the high-frequency feature information of the image, while avoiding the problem of vanishing gradients in the model.

[0064] This embodiment is aimed at the small target recognition scenario and constructs an interpretable gradient-guided image super-resolution network model. This model is used to perform high-fidelity SR reconstruction of weak feature microcracks and their surrounding pixels. It can effectively overcome the problems of low resolution and smooth grayscale differences in the image to be detected, while reducing the impact of interference noise and improving the accuracy of small target recognition.

[0065] Example 3: This embodiment provides a method for reconstructing the surface image of an end electrode, which uses the visual gradient interpretable super-resolution method for small target recognition recorded in Example 2 to perform super-resolution reconstruction on the surface image of the end electrode, thereby obtaining a high-fidelity, high-resolution image for subsequent recognition, including the following steps.

[0066] Step 1: Acquire the end electrode image to be reconstructed.

[0067] Step 2: Correct the end electrode image, such as Figure 1 As shown, the following steps are included: Step 21: Electrode pin orientation correction: Using the pixel value variation characteristics of the end electrode pin area and its surroundings, a second-order derivative detection kernel based on central difference is designed to determine the orientation of the end electrode pins and rotate the image to a uniform orientation with the pins pointing downward. Step 22: Correcting the end electrode pin offset: Using the linear characteristics of the area above the end electrode image, the pin offset angle is obtained and corrected through image binarization, edge detection, and Hough line detection algorithm; Step 23: Acquisition of end electrode image: Use morphological opening operation to distinguish the end electrode area from the non-end electrode area, and use contour search algorithm to locate the end electrode area and cut it out.

[0068] Step 3: Use the image segmentation algorithm to divide the area to be detected, such as Figure 2 As shown, the following steps are included: Step 31: Through the horizontal position positioning step, histogram equalization, debinarization and morphological opening operations are performed on the vertical linear area of ​​the pin in the image to determine the horizontal positions of the left and right outermost areas to be detected; Step 32: Using the positive and negative electrode identification method, the fixed area image is intercepted and binarized, morphological opening operation is performed, and pixel value summation is performed to determine the positive and negative electrodes, thereby obtaining the horizontal position of the inner area to be detected; Step 33: Determine the horizontal and vertical positions of all areas to be detected, and cut out the image to be detected.

[0069] Step 4: Construct an interpretable gradient-guided image super-resolution network model (IGGS). According to the overall network model structure, build a feature extraction unit, a feature mapping module guided by gradient information, an upsampling unit, a pixel reconstruction unit, and a global residual unit.

[0070] Step 5: Prepare the experimental dataset. 1,000 images of the tantalum capacitor electrode surface were collected and divided into training, validation, and test sets with a 60% / 20% / 20% ratio. The high-resolution images were scaled to half their original size using the bicubic interpolation algorithm and used as low-resolution input images. The model was trained using the Adam optimizer in Python 3.5, PyTorch 1.1.0, and CUDA 9.0, with a learning rate of 0.0001, a batch size of 16 per training round, and 200 training rounds. After training, the low-resolution images to be tested were input into the trained model. After processing by the model's various units, a high-resolution reconstructed image was output.

[0071] During the training process, the mean square error (MSE) loss function is used to calculate the pixel difference between the reconstructed image and the true label image, and the model parameters are optimized through back propagation and gradient descent algorithms.

[0072] Step 6: Perform pixel-level selective response processing based on convolutional line detection on the image output by the super-resolution model, and use the horizontal average filter operator to perform ten horizontal filtering operations on the image to filter out horizontal texture noise.

[0073] Step 7: Convolve the filtered image with four microcrack line detection operators in different directions, take the maximum value of the four detection results, and obtain the line detection response of the microcrack.

[0074] Step 8: Dilate the microcrack line detection response using a dilation algorithm based on a neural network convolutional layer, repeating this process four times. Bilinearly interpolate the dilated image to bring its resolution into line with the output feature layer of the pixel reconstruction unit. Activate it using a selective response activation function based on the hyperbolic tangent function. Finally, perform a pixel-by-pixel multiplication with the output feature of the pixel reconstruction unit to obtain the final processed image. This achieves high-fidelity SR reconstruction of weakly characterized microcracks and their surrounding areas, while suppressing the influence of interfering noise.

[0075] This embodiment designs a visual gradient interpretable super-resolution method for small target recognition, which is used to realize the preprocessing of microcrack images. First, using the fixed shape features of the end electrode surface, an algorithm for acquiring the image of the small target surface to be detected is designed, and the image of the area to be detected is intercepted from the acquired image through the tantalum capacitor end electrode image correction algorithm and the segmentation algorithm; then, in response to the problems of low resolution and smooth grayscale difference of the image to be detected, an interpretable gradient-guided image super-resolution model is designed to achieve high-fidelity SR reconstruction of low-resolution images; in response to the problem of a large amount of interference noise in the background, a microcrack pixel-level selection module is designed based on the SR model, which realizes high-fidelity reconstruction of the microcrack area while reducing the influence of interference noise to a certain extent. The method of this embodiment can achieve high-fidelity super-resolution reconstruction of microcrack images with weak features on the surface of tantalum capacitor end electrodes, while suppressing the influence of background noise, thereby improving the accuracy of small target recognition.

[0076] Some steps in the embodiments of the present invention may be implemented using software, and the corresponding software program may be stored in a readable storage medium, such as a CD or a hard disk.

[0077] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A super-resolution method for small target recognition, characterized in that: The method comprises: Step 1: Obtain the image to be reconstructed; Step 2: construct an interpretable gradient-guided image super-resolution network model IGGS, and use the IGGS model to perform super-resolution reconstruction on the image to be reconstructed; The IGGS model includes a feature extraction unit, a feature mapping module GGFM guided by gradient information, an upsampling unit, a pixel reconstruction unit, and a global residual unit. The expression of the forward propagation process of the IGGS model is: Where, represents the image to be reconstructed; represents the super-resolution image output by the IGGS model; represents the upsampling unit; Represents the network model feature fitting function; represents the feature extraction unit; Represents the GGFM module; represents the pixel reconstruction unit; The forward propagation process expression of the GGFM module is: represents the input features, Represents a feature mapping unit, which is used to learn and fit high-frequency features in images. Represents the gradient operation unit, which is used to obtain image gradients in different directions.

2. The super-resolution method for small target recognition according to claim 1, characterized in that: The gradient differential operator expression in the gradient operation unit is: Where, 、 、 、 Represents the gradient differential operators of the image in the horizontal direction, vertical direction, upper left to lower right direction, and upper right to lower left direction respectively.

3. The super-resolution method for small target recognition according to claim 1, characterized in that: The method further comprises: For the image output by the IGGS model, horizontal texture noise is filtered out using a horizontal mean filter operator; The filtered image is convolved with four microcrack line detection operators in different directions, and the maximum value of the four detection results is taken to obtain the line detection response of the microcrack; A dilation operation is performed on the microcrack line detection response result, and the expanded image is bilinearly interpolated to make its resolution consistent with the output feature layer of the pixel reconstruction unit. It is then activated by a selective response activation function designed based on the hyperbolic tangent function, and finally multiplied pixel-wise with the output feature of the pixel reconstruction unit to obtain the final processed image.

4. The super-resolution method for small target recognition according to claim 1, characterized in that: The up-sampling unit adopts a bilinear interpolation method.

5. The super-resolution method for small target recognition according to claim 1, characterized in that: The pixel reconstruction unit adopts a single convolution layer.

6. The super-resolution method for small target recognition according to claim 1, characterized in that: The calculation expression of the feature extraction unit is: in, Represents the output features of the feature extraction unit; represents the convolutional layer; Represents the ReLU activation function.

7. The super-resolution method for small target recognition according to claim 3, characterized in that: Bilinear interpolation methods include: Know the target point required Four adjacent coordinates and corresponding pixel values: 、 、 and , linear interpolation is performed in the horizontal direction to obtain and The specific calculation formula for the pixel value is as follows: Where, 、 、 、 、 、 Represents the coordinate points 、 、 、 、 、 Pixel value of After completing the horizontal linear interpolation through the above formula, then through and The coordinate value of is interpolated with the obtained pixel value in the vertical direction, and the expression is: Will and After the pixel value of the target point is substituted into the above formula, Pixel value Can be obtained from this.

8. A small target recognition method, characterized in that: The method uses the super-resolution method for small target recognition as described in any one of claims 1 to 7 to perform super-resolution reconstruction on the image to be recognized, and then performs target recognition based on the reconstructed image.

9. An electronic device, characterized in that: including memory and processor; The memory is used to store computer programs; The processor is configured to implement the method according to any one of claims 1 to 8 when executing the computer program.

10. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images