Efficient positioning method for anode copper plate clamp based on computer vision

Abstract

The invention relates to the technical field of computer vision and robots, and discloses an efficient positioning method for an anode copper plate clamp based on computer vision. According to the method, multi-view information of the copper plate is collected in a designated area through the depth camera and the RGB camera, accurate positioning and posture estimation of the copper plate are achieved in combination with deep learning, and therefore a mechanical arm is guided to complete an efficient and stable grabbing task. The method comprises the following steps: firstly, using a depth camera and an RGB (Red, Green and Blue) camera to shoot a depth color map of the whole batch of anode copper plates in a specified area from three angles of front, side and upper, and preprocessing the obtained image; secondly, a deep learning model based on multi-view information fusion for the anode copper plates is designed, and the model can directly output the surface center point coordinates and the inclination angle of each anode copper plate in the whole batch; and finally, the model is trained, so that each anode copper plate in the whole batch can be accurately positioned by the model, and finally efficient, accurate and stable operation of grabbing work of a mechanical arm is helped.

Description

Technical Field

[0001] The present invention relates to the field of computer vision and robotics, and in particular to an efficient positioning method for an anode copper plate fixture based on computer vision, which is particularly suitable for the automated grasping and handling of heavy copper plates in an industrial environment. Background Art

[0002] The handling and gripping of copper plates is a critical step in the production and processing of copper plates. Traditional handling methods rely on manual operation or simple mechanical devices, which are subject to low efficiency, poor precision, and significant safety hazards. With the development of industrial automation technology, robotic arms have gradually been introduced to the task of handling copper plates. However, due to the heavy weight of the copper plates and the complex placement (tilted placement and stacking of multiple plates), the gripping accuracy and stability of the robotic arms face significant challenges.

[0003] Existing robotic gripping systems typically rely on pre-set gripping paths and fixed gripping points, making them difficult to adapt to changes in the placement and orientation of copper plates. Alternatively, positioning must be performed by taking photos or other methods before each grasp, significantly reducing the efficiency of the robotic arm's grasping and handling. Therefore, achieving efficient, accurate, and stable positioning and gripping of copper plates has become a key research topic in industrial automation. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the present invention provides an efficient positioning method for anode copper plate fixture based on computer vision, which solves the above technical problems.

[0005] To achieve the above object, the present invention provides the following technical solution: a computer vision-based efficient positioning method for an anode copper plate fixture, comprising the following steps:

[0006] S1. Multi-view data acquisition: Use a depth camera and an RGB camera to capture a set of depth and RGB images of the copper plate from three angles: front, side, and top. Ensure that the image resolution and depth accuracy meet the requirements of subsequent processing.

[0007] S2. Preprocess the image obtained in S1, perform denoising and filtering on the depth image to remove noise points; perform contrast enhancement and histogram equalization on the RGB image to improve image quality; concatenate the image (B, C, h, w, D) obtained in S2 in the w dimension to obtain the image (B, C, h, 3w, D), where B represents the batch, C represents the RGB channel, h and w represent the length and width of the image respectively, and D represents the depth of the image;

[0008] S3. Establish a deep learning model based on multi-view information fusion, take the image (B, C, h, 3w, D) obtained in S2 as input, perform feature extraction, feature compression and other processes in a multi-layer neural network, and finally obtain the surface center point coordinates and roll angle (B, N, 5) of each copper plate in the entire batch, where N represents the number of copper plates in the image recognized by the model, and 5 represents the output position features of the five dimensions of each anode copper plate in the entire batch, namely (x, y, z, α, β), (x, y, z) are the surface center point coordinates of each copper plate, (α, β) are the tilt angles of each copper plate, α is the pitch angle (rotation around the Y axis), and β is the yaw angle (rotation around the Z axis);

[0009] S4. The model established in S3 is trained on a data set. The data set is 2000 sets of copper plate images collected, each set containing a depth image and an RGB image spliced from three images of the front, side, and top. A loss function is used to measure the gap between the predicted value and the true value. The Adam optimizer is used with a learning rate of 0.001, a batch size of 32, and 100 training rounds. The loss function value is gradually reduced through training and learning, thereby improving the accuracy of the model. Specifically, the present invention adopts a loss function that combines position loss and angle loss. The position loss adopts Euclidean distance, and the angle loss adopts cosine similarity loss. The formula is as follows:

[0010]

[0011] Among them, L position represents the Euclidean distance (position loss), (x i ,y i ,z i ) is the true value of the coordinates of the center point on the surface of the i-th anode copper plate, is the model predicted value of the coordinate of the center point of the surface of the i-th anode copper plate, L angle represents the cosine similarity loss (angle loss), (α i ,β i ) is the true value of the tilt angle of the i-th anode copper plate, is the model predicted value of the tilt angle of the i-th anode copper plate, and L represents the total loss function.

[0012] As a preferred technical solution of the present invention, step S3 includes the following steps:

[0013] S3.1. Input the image (B, C, h, 3w, D) obtained in S2, and pass it through one initial convolutional layer, two residual blocks, one flattening layer, and two fully connected layers to obtain the feature (B, 1). This shows that the model can recognize the number N of copper plates in the image.

[0014] The specific layers are as follows:

[0015] The first layer is the initial convolutional layer, with an input shape of (B, C, h, 3w, D) and an output shape of (B, 32, h, 3w, D);

[0016] The second layer is residual block 1, with an input shape of (B, 32, h / 2, 3w / 2, D / 2) and an output shape of (B, 64, h / 4, 3w / 4, D / 4).

[0017] The third layer is residual block 2, with an input shape of (B, 64, h / 4, 3w / 4, D / 4) and an output shape of (B, 128, h / 8, 3w / 8, D / 8);

[0018] The fourth layer is the flattening layer, which flattens the multidimensional data into a one-dimensional vector. The input shape is (B, 128, h / 8, 3w / 8, D / 8), and the output shape is (B, 128*(h / 8)*(3w / 8)*(D / 8)).

[0019] The fifth layer is a fully connected layer with an input shape of (B, 128*(h / 8)*(3w / 8)*(D / 8)) and an output shape of (B, 128);

[0020] The sixth layer is a fully connected layer with an input shape of (B, 128) and an output shape of (B, 1);

[0021] S3.2. Resize the image (B, C, h, 3w, D) obtained in S2 to obtain (B, C + D, h, 3w). Then, pass it through four convolutional layers, one global average pooling layer, and one flattening layer to obtain the data (B, 64), which serves as the input of the dynamic fully connected layer.

[0022] The global pooling layer can effectively compress the spatial dimensions (height and width) of the feature map to 1×1 while retaining the channel information, providing high-quality global features for the subsequent fully connected layer; global average pooling takes the average of all spatial positions of each channel, and its formula is:

[0023]

[0024] Among them, the output [b,c,1,1] is the value of the bth sample, cth channel, i-th row, and j-th column in the input feature map, h is the height of the feature map, and w is the width of the feature map. is the normalization factor used to calculate the mean;

[0025] The specific layers are as follows:

[0026] The first layer is convolutional layer 1, with an input shape of (B, C+D, h, 3w) and an output shape of (B, 32, h, 3w);

[0027] The second layer is convolutional layer 2, with an input shape of (B, 32, h, 3w) and an output shape of (B, 64, h, 3w);

[0028] The third layer is convolutional layer 3, with an input shape of (B, 64, h, 3w) and an output shape of (B, 128, h, 3w);

[0029] The fourth layer is convolutional layer 4, with an input shape of (B, 128, h, 3w) and an output shape of (B, 256, h, 3w);

[0030] The fifth layer is the global average pooling layer, with an input shape of (B, 256, h, 3w) and an output shape of (B, 256, 1, 1);

[0031] The sixth layer is the flattening layer, with an input shape of (B, 256, 1, 1) and an output shape of (B, 256);

[0032] In addition, skip connections are added between the input and convolutional layer 2, convolutional layer 1 and convolutional layer 3, and convolutional layer 2 and convolutional layer 4. 1×1 convolution is used for projection skip connections to ensure that the input and output shapes match. The residual network is introduced to implement residual learning through skip connections, which can solve the gradient vanishing and degradation problems of deep networks.

[0033] S3.3. Use the feature (B, 256) obtained in S3.2 and the number of copper plates N in the image identified by the model obtained in S3.1 as the input of the dynamic fully connected layer, output the feature (B, N×5), and then adjust the output shape to (B, N, 5). The specific process of the dynamic fully connected layer is:

[0034] Basic fully connected layer: maps the global features (B, 256) to an intermediate dimension (B, 512) to extract high-level features. The weight matrix W1 has a shape of (256, 512) and the bias b1 has a shape of (512). The formula is:

[0035] Intermediate features=Global features×W1+b1;

[0036] Dynamic weight generation: According to the number of copper plates N in the image identified by the model, N×5 features are dynamically generated. The weight matrix W2 has a shape of (512,5) and the bias b2 has a shape of (5). For each sample, the weight W2 and bias b2 are copied N times according to N to obtain the expanded weight matrix The shape is (512, N × 5), bias The shape is (N×5);

[0037] Dynamic calculation: Dynamically calculate the output features for each sample, dynamically expand the weight W2 and bias b2 according to N, and output the shape (B, N×5) according to the formula:

[0038]

[0039] Compared with the existing technology, the present invention provides an efficient positioning method for anode copper plate fixture based on computer vision, which has the following beneficial effects:

[0040] The present invention first uses a depth camera and an RGB camera to collect information from three perspectives of a batch of anode copper plates at one time. By fusing images from multiple perspectives, it can capture scene information more comprehensively and reduce the blind spots of a single perspective. At the same time, it helps to reduce the noise and error of a single perspective and improve the accuracy of the data. Secondly, the present invention designs a multi-perspective information fusion and deep learning model for anode copper plates. By inputting a depth RGB image, the model can directly output the surface center point coordinates and tilt angle of each anode copper plate in the entire batch. At the same time, the model can automatically learn features from the data, avoiding the tediousness and potential deviations of manually designed features. For the entire batch of copper plates, compared with the prior art, which requires information collection before each copper plate is grabbed, the present invention only needs to collect information for the entire batch of copper plates once, and then each copper plate in the entire batch can be located. This can shorten the time to obtain the position information of the entire batch of anode copper plates, and thus greatly improve the efficiency of the robot arm's grasping work. Finally, through model training and model optimization, the model can accurately locate each anode copper plate in the entire batch. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Schematic diagram of the process of the efficient positioning method of the anode copper plate fixture based on computer vision in the present invention;

[0042] Figure 2 This is a schematic diagram of collecting three-view information of the entire batch of anode copper plates in the present invention;

[0043] Figure 3 This is a schematic diagram of the end-to-end model structure based on multi-perspective information fusion and deep learning in the present invention. DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0045] See also Figure 1 、 Figure 2 and Figure 3 The present invention provides an efficient positioning method for an anode copper plate fixture based on computer vision, comprising the following steps:

[0046] S1, multi-view data acquisition, such as Figure 2 As shown in the figure, a depth camera and an RGB camera are used to capture depth and RGB images of a batch of anode copper plates from three angles: front, side, and top. Multi-view fusion uses multimodal data (depth + RGB) and geometric information from multiple angles to more comprehensively describe the 3D pose and surface features of the target object, and is more stable in scenes with occlusion, uneven lighting, and weak textures.

[0047] The RGB image resolution is 1920x1080, using HDR mode to reduce overexposed or underexposed areas, improve texture clarity, and adapt to dynamic lighting environments. The depth image resolution is 1280x720, with a depth range of 0.5m-3m. A time-of-flight (ToF) depth camera is used to enhance depth details in edges and inclined areas, keeping depth errors to the millimeter level. This ensures that the image resolution and depth accuracy meet the requirements of subsequent processing.

[0048] S2, preprocess the image obtained in S1, perform denoising and filtering on the depth image to remove noise points; perform contrast enhancement and histogram equalization on the RGB image to improve image quality; splice the image (B, C, h, w, D) obtained in S1 in the w dimension to obtain the image (B, C, h, 3w, D); where B represents the batch, C represents RGB channels, h , w represent the length and width of the image respectively, and D represents the depth of the image; the image stitching process is as follows:

[0049] S2.1. Image alignment and resolution adaptation: By pre-calibrating the internal and external parameters of the depth camera and RGB camera, the geometric transformation relationship between multiple viewpoints (front, side, and top) is determined to ensure the physical spatial alignment of images from different viewpoints. The depth image (1280×720) is upsampled to the resolution of the RGB image (1920×1080) through bilinear interpolation to ensure the consistency of the spatial dimensions of the two.

[0050] S2.2. Multimodal data fusion: The depth image is treated as a single-channel data (depth value) and merged with the three channels (R, G, B) of the RGB image to form a four-channel (RGB-D) input. The RGB-D images of different perspectives are horizontally spliced along the width dimension (W) to generate a wide image.

[0051] The fusion of multi-view and multi-modal data reduces single-view blind spots and enhances the model's ability to perceive the copper plate's posture. Through splicing in the pre-processing stage, the model avoids repeated processing of multi-view data during inference, improving end-to-end computing speed.

[0052] S3. Establish a deep learning model based on multi-view information fusion, such as Figure 3 As shown, the image (B, C, h, 3w, D) obtained by S2 is used as input, and feature extraction and feature compression are performed in the multi-layer neural network to finally obtain the surface center point coordinates and tilt angle (B, N, 5) of each copper plate in the entire batch; N Represents the number of copper plates in the image recognized by the model, and 5 represents the five-dimensional position features of each copper plate in the entire batch of outputs, namely (x, y, z, α, β), (x, y, z) is the coordinate of the center point of the copper plate surface, (α, β) is the tilt angle of the copper plate, α is the pitch angle (rotation around the Y axis), and β is the yaw angle (rotation around the Z axis);

[0053] S3.1, input S2 obtained image (B, C, h, 3w, D), through 1 layer of initial convolution layer, 2 layers of residual block, 1 layer of flattening layer, two layers of fully connected layer to obtain the feature (B, 1). The number of copper plates in the image recognized by the model N A tensor equal to (B,1). The specific structure of each layer is as follows:

[0054] The first layer is the initial convolution layer, which includes one convolution layer and one maximum pooling layer. The convolution kernel size of the convolution layer is (3,3,3), the stride is 1, the padding is 1, the input shape is (B,C,h,3w,D), and the output shape is (B,32,h,3w,D); the output of the convolution layer is normalized to make the data distribution more stable, thereby alleviating the gradient vanishing and gradient exploding problems; the ReLU activation function is added after the convolution layer to improve the nonlinearity introduced so that the neural network can learn complex patterns, while also alleviating the gradient vanishing problem and improving the training effect of deep networks; the maximum pooling layer has a pooling kernel size of (2,2,2), a stride of 2, an input shape of (B,32,h,3w,D), and an output shape of (B,32,h / 2,3w / 2,D / 2);

[0055] The second layer is residual block 1, which includes convolutional layer 1, convolutional layer 2 and maximum pooling layer; the input shape is (B, 32, h / 2, 3w / 2, D / 2), and the output shape is (B, 64, h / 4, 3w / 4, D / 4);

[0056] First, convolutional layer 1 can extract local features of the data. The convolution kernel size is (3,3,3), the stride is 1, the padding is 1, the input shape is (B,32,h / 2,3w / 2,D / 2), and the output shape is (B,64,h / 2,3w / 2,D / 2). The output of the convolutional layer is normalized to make the data distribution more stable, thereby alleviating the gradient vanishing and gradient exploding problems. The ReLU activation function is added after the convolutional layer to improve the nonlinearity introduced so that the neural network can learn complex patterns. At the same time, it can alleviate the gradient vanishing problem and improve the training effect of deep networks.

[0057] Next is convolution layer 2, with a kernel size of (3,3,3), a stride of 1, a padding of 1, an input shape of (B,64,h / 2,3w / 2,D / 2), and an output shape of (B,64,h / 2,3w / 2,D / 2). The output of the convolution layer is normalized to make the data distribution more stable, thereby alleviating the problems of gradient disappearance and gradient explosion. A ReLU activation function is added after the convolution layer to improve the nonlinearity introduced so that the neural network can learn complex patterns, while alleviating the gradient disappearance problem and improving the training effect of the deep network. A jump connection is added to the output of the preliminary convolution layer and the output of convolution 2 to form a standard residual block. The residual network is introduced to achieve residual learning through jump connections, which can solve the gradient disappearance and degradation problems of the deep network. The number of channels of the input and output are inconsistent, and 1x1 convolution is used to adjust the input shape. The formula of the standard residual block is: v = F(x,W i )+u, where u is the input, F(u,W i ) is the output of the two convolutional layers (residual map), and v is the final output;

[0058] Finally, there is the maximum pooling layer, which can reduce the spatial dimension of the feature map, reduce the amount of computation and prevent overfitting; the pooling kernel size is (2,2,2), the stride is 2, the input shape is (B,64,h / 2,3w / 2,D / 2), and the output shape is (B,64,h / 4,3w / 4,D / 4);

[0059] The third layer is residual block 2, which includes convolution layer 1, convolution layer 2 and maximum pooling layer. The input shape is (B, 64, h / 4, 3w / 4, D / 4) and the output shape is (B, 128, h / 8, 3w / 8, D / 8).

[0060] First, convolutional layer 1 can extract local features of the data. The convolution kernel size is (3, 3, 3), the stride is 1, the padding is 1, the input shape is (B, 64, h / 4, 3w / 4, D / 4), and the output shape is (B, 128, h / 4, 3w / 4, D / 4). The output of the convolutional layer is normalized to make the data distribution more stable, thereby alleviating the gradient vanishing and gradient exploding problems. The ReLU activation function is added after the convolutional layer to improve the nonlinearity introduced so that the neural network can learn complex patterns. At the same time, it can alleviate the gradient vanishing problem and improve the training effect of deep networks.

[0061] Next is convolution layer 2, with a kernel size of (3,3,3), a stride of 1, a padding of 1, an input shape of (B,128,h / 4,3w / 4,D / 4), and an output shape of (B,128,h / 4,3w / 4,D / 4). The output of the convolution layer is normalized to make the data distribution more stable, thereby alleviating the problems of gradient disappearance and gradient explosion. A ReLU activation function is added after the convolution layer to improve the nonlinearity introduced so that the neural network can learn complex patterns, while alleviating the gradient disappearance problem and improving the training effect of the deep network. A jump connection is added to the output of the preliminary convolution layer and the output of convolution 2 to form a standard residual block. The residual network is introduced to achieve residual learning through jump connections, which can solve the gradient disappearance and degradation problems of the deep network. The number of channels of the input and output are inconsistent, and 1x1 convolution is used to adjust the input shape. The formula of the standard residual block is: v=F(x,W i )+u, where u is the input, F(u,W i ) is the output of the two convolutional layers (residual map), and v is the final output;

[0062] Finally, there is the maximum pooling layer. The pooling layer can reduce the spatial dimension of the feature map, reduce the amount of computation and prevent overfitting. The pooling kernel size is (2, 2, 2), the stride is 2, the input shape is (B, 128, h / 4, 3w / 4, D / 4), and the output shape is (B, 128, h / 8, 3w / 8, D / 8).

[0063] The fourth layer is the flattening layer, which flattens the multidimensional data into a one-dimensional vector. The input shape is (B, 128, h / 8, 3w / 8, D / 8), and the output shape is (B, 128*(h / 8)*(3w / 8)*(D / 8)).

[0064] The fifth layer is a fully connected layer with a weight shape of (128, 128*(h / 8)*(3w / 8)*(D / 8)), a bias shape of (128), an input shape of (B, 128*(h / 8)*(3w / 8)*(D / 8)), and an output shape of (B, 128);

[0065] The sixth layer is a fully connected layer with a weight shape of (1,128), a bias shape of (1), an input shape of (B,128), and an output shape of (B,1);

[0066] S3.2. Adjust the dimensions of the image (B, C, h, 3w, D) obtained in S2, and concatenate the depth channel (D) and the RGB channel (C) in the channel dimension to obtain (B, C + D, h, 3w), retaining the height (h) and width (3w) dimensions after concatenation to ensure the integrity of multi-view spatial information; fuse RGB information (color, texture) and depth information (geometric structure) in the channel dimension, so that the model can simultaneously utilize color features and 3D shape features to enhance the perception of the copper plate surface and posture; integrate multimodal data into a single feature map to avoid processing RGB and depth data separately, reduce computational redundancy, and improve model inference speed;

[0067] Take (B, C+D, h, 3w) as input, pass through 4 convolutional layers, 1 global average pooling layer and 1 flattening layer to obtain data (B, 64);

[0068] The global pooling layer can effectively compress the spatial dimensions (height and width) of the feature map to 1×1 while retaining the channel information, providing high-quality global features for the subsequent fully connected layer; global average pooling takes the average of all spatial positions of each channel, and its formula is:

[0069] Output Input [b,c,i,j],

[0070] Among them, the output [b,c,1,1] is the value of the bth sample, cth channel, i-th row, and j-th column in the input feature map, h is the height of the feature map, and w is the width of the feature map. is the normalization factor used to calculate the mean;

[0071] The specific structure of each layer is as follows:

[0072] The first layer is convolutional layer 1, with a kernel size of (3,3,3), a stride of 1, a padding of 1, an input shape of (B,C+D,h,3w), and an output shape of (B,32,h,3w). A ReLU activation function is added after the layer to improve the nonlinearity of the neural network so that it can learn complex patterns, alleviate the gradient vanishing problem, and improve the training effect of deep networks.

[0073] The second layer is convolutional layer 2, with a kernel size of (3,3,3), a stride of 1, a padding of 1, an input shape of (B,32,h,3w), and an output shape of (B,64,h,3w). A ReLU activation function is added after the layer to improve the nonlinearity of the neural network so that it can learn complex patterns, alleviate the gradient vanishing problem, and improve the training effect of deep networks.

[0074] The third layer is convolution layer 3, with a kernel size of (3,3,3), a stride of 1, a padding of 1, an input shape of (B,64,h,3w), and an output shape of (B,128,h,3w). A ReLU activation function is added after the layer to improve the nonlinearity of the neural network so that it can learn complex patterns, alleviate the gradient vanishing problem, and improve the training effect of deep networks.

[0075] The fourth layer is convolutional layer 4, with a kernel size of (3,3,3), a stride of 1, a padding of 1, an input shape of (B,128,h,3w), and an output shape of (B,256,h,3w). A ReLU activation function is added after the layer to improve the nonlinearity of the neural network so that it can learn complex patterns, alleviate the gradient vanishing problem, and improve the training effect of deep networks.

[0076] The fifth layer is the global average pooling layer, with an input shape of (B, 256, h, 3w) and an output shape of (B, 256, 1, 1);

[0077] The sixth layer is the flattening layer, which flattens the multidimensional data into a one-dimensional vector. The input shape is (B, 256, 1, 1) and the output shape is (B, 256).

[0078] In addition, jump connections are added between the input to convolution layer 2, convolution layer 1 to convolution layer 3, and convolution layer 2 to convolution layer 4; 1×1 convolution is used for projection jump connection to ensure that the shapes of input and output match: input to the output of convolution layer 2, use 1×1 convolution to project the input from C+D channels to 64 channels; the output of convolution layer 1 to the output of convolution layer 3, use 1×1 convolution to project the input from 32 channels to 128 channels; the output of convolution layer 2 to the output of convolution layer 4, use 1×1 convolution to project the input from 64 channels to 256 channels; introduce residual network and implement residual learning through jump connection, which can solve the gradient disappearance and degradation problems of deep network. The formula of standard residual block is:

[0079] y=F(x,W i )+x,

[0080] Where x is the input, F(x,W i ) is the output of the two convolutional layers (residual map), and y is the final output;

[0081] S3.3. Use the feature (B, 256) obtained in S3.2 and the number of copper plates N in the image identified by the model obtained in S3.1 as the input of the dynamic fully connected layer, output the feature (B, N×5), and then adjust the output shape to (B, N, 5). The specific process of the dynamic fully connected layer is:

[0082] Basic fully connected layer: maps the global features (B, 256) to an intermediate dimension (B, 512) to extract high-level features. The weight matrix W1 has a shape of (256, 512) and the bias b1 has a shape of (512). The formula is:

[0083] Intermediate features=Global features×W1+b1;

[0084] Dynamic weight generation: According to the number of copper plates N in the image identified by the model, N×5 features are dynamically generated. The weight matrix W2 has a shape of (512,5) and the bias b2 has a shape of (5). For each sample, the weight W2 and bias b2 are copied N times according to N to obtain the expanded weight matrix The shape is (512, N × 5), bias The shape is (N×5);

[0085] Dynamic calculation: Dynamically calculate the output features for each sample, dynamically expand the weight W2 and bias b2 according to N, and output the shape (B, N×5) according to the formula:

[0086]

[0087] S4. Train the model built in S3 on a dataset of 2,000 copper plate images. Each dataset contains a depth image and an RGB image stitched together from the front, side, and top views. A loss function is used to measure the difference between the predicted value and the true value. The Adam optimizer is used with a learning rate of 0.001, a batch size of 32, and 50 training rounds. The loss function is gradually reduced through training, thereby improving the accuracy of the model.

[0088] This invention uses a loss function that combines position loss and angle loss. Position loss is calculated using Euclidean distance, and angle loss is calculated using cosine similarity loss. The formula is as follows:

[0089]

[0090] Among them, L position represents the Euclidean distance (position loss), (x i ,y i ,z i ) is the true value of the coordinate of the center point of the surface of the i-th anode copper plate in the entire batch, is the model predicted value of the coordinate of the center point of the surface of the i-th anode copper plate in the entire batch, L angle represents the cosine similarity loss (angle loss), (α i ,β i ) is the true value of the tilt angle of the i-th anode copper plate in the entire batch, is the model predicted value of the tilt angle of the i-th anode copper plate in the entire batch, and L represents the total loss function.

[0091] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An efficient positioning method for anode copper plate fixture based on computer vision, characterized in that: include: Acquire multi-view image data, where the multi-view image data includes a depth image and an RGB image; Preprocessing the multi-view image data to obtain a spliced image; Inputting the stitched image into a pre-established deep learning model, wherein the deep learning model comprises at least a feature extraction module and a dynamic full connection module; Performing feature extraction on the spliced image by the feature extraction module to obtain feature data; The characteristic data is processed by the dynamic full-connection module to obtain positioning information of the anode copper plate, wherein the positioning information includes the coordinates of the surface center point and the tilt angle.

2. The method according to claim 1, wherein The acquiring of multi-view image data includes: Use a depth camera and an RGB camera to capture depth images and RGB images of the anode copper plate from three angles: front, side, and top. The resolution and depth accuracy of the depth images and RGB images meet preset requirements.

3. The method according to claim 1, wherein The preprocessing of the multi-view image data includes: Performing denoising and filtering on the depth image to remove noise points; Performing contrast enhancement and histogram equalization on the RGB image to improve image quality; The processed depth image and RGB image are stitched together in the width dimension to obtain a stitched image.

4. The method according to claim 1, wherein The feature extraction module of the deep learning model includes: An initial convolutional layer, multiple residual blocks, a flattening layer and a fully connected layer, wherein the initial convolutional layer is used to perform preliminary feature extraction on the spliced image, the residual block is used to further extract the preliminary features, the flattening layer is used to flatten the multidimensional feature data into a one-dimensional vector, and the fully connected layer is used to perform feature mapping on the one-dimensional vector.

5. The method according to claim 1, wherein The dynamic fully connected module of the deep learning model includes: A basic fully connected layer and a dynamic weight generation module, wherein the basic fully connected layer is used to map global features to intermediate dimensions and extract high-level features, and the dynamic weight generation module is used to dynamically generate feature weights according to the number of anode copper plates identified by the model.

6. The method according to claim 1, wherein The processing of the feature data by the dynamic fully connected module includes: Based on the dynamically generated weights and biases, each sample is dynamically calculated to output the surface center coordinates and tilt angle of the anode copper plate.

7. The method according to claim 1, wherein The training process of the deep learning model includes: A preset loss function is used to measure the gap between the predicted value and the true value. The loss function includes position loss and angle loss. The position loss is calculated using Euclidean distance, and the angle loss is calculated using cosine similarity.

Images