A multi-scale blob detection method

Abstract

The application provides a multi-scale glue point detection method, and belongs to the technical field of glue point detection; the method comprises the following steps: adopting an intelligent laser lens dispensing platform to collect images before laser chip bonding; establishing an LLAS data set based on the collected images; establishing a multi-level attention guide context aggregation network model; applying the multi-level attention guide context aggregation network model to process images in the LLAS data set, and segmenting out glue point targets; and obtaining glue point detection results based on the segmented glue point targets. The application proposes a multi-level attention guide context aggregation network (MACAN) based on a self-built LLAS data set. The purpose is to realize real-time and robust detection of glue points on laser chips, and to provide necessary quality assurance for realizing high coupling efficiency of high-power laser lenses.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of glue point detection, and relates to a multi-scale glue point detection method. BACKGROUND

[0002] High-power lasers have wide application requirements in the fields of optical communication, industrial processing, medical instruments, national defense and military, etc. These fields have very high requirements for the efficiency, stability and service life of the laser. As one of the core optical elements, the packaging process of the collimating lens directly affects the beam quality, transmission efficiency and overall stability of the system. However, the coupling and packaging technology of the collimating lens still faces many challenges, especially in the dispensing process, the position, shape and size of the glue point significantly affect the coupling accuracy of the collimating lens, and the glue point defect becomes one of the main factors affecting the performance of the optical system. Therefore, it is necessary to detect the quality of the glue point of the collimating lens in the coupling and packaging process of the high-power laser.

[0003] In recent years, thanks to the rapid development of convolutional neural networks, defect detection technology using deep learning networks has shown remarkable results in many industrial defect recognition tasks. However, for laser collimating lens glue point detection, there are still three main challenges:

[0004] (1) Double target detection introduces more complexity to the glue point detection task. The collimating lens is fixed by two glue points. The coupling accuracy of the collimating lens is not only affected by the position and shape of a single glue point, but more importantly, it is affected by the accurate relative position relationship between the two glue points. In double target detection, there is a spatial position relationship between the two targets, which needs the model to effectively understand and consider the relationship between the two targets, especially when there is a large difference in the scale between the two targets, the network's receptive field and the design of the detection framework become more complex.

[0005] (2) Random shape and other multi-scale defects challenge the accurate detection of glue points. Under ideal conditions, the shape of the two glue points is regular and circular, and the target has a proper scale in the entire field of view, and there is a high contrast between the target and the background. However, in actual dispensing operations, glue points often have potential problems such as different degrees of position deviation, size deviation and random shape. The glue point shape is irregular, and the overall size is small, which may challenge the accuracy of detection.

[0006] (3) Complex background may affect the accurate detection of the glue point. The features of the glue point are not only difficult to distinguish from the background color of the target area, but the overflow of the large-area chip solder below further increases the difficulty of accurately identifying the glue point from the background. This situation not only poses a challenge to traditional image processing methods based on grayscale information, but even for deep learning methods, the network is required to have high robustness to ensure accurate identification of the target. SUMMARY

[0007] Based on the deficiencies in the prior art, the present application provides a method for detecting multi-scale glue points in the bonding of laser collimating lenses based on a Multi-level Attention-Driven Fusion Network (MADF-Net) network model, aiming to realize more real-time and robust detection of laser collimating lens glue points and provide necessary quality assurance for high coupling efficiency of high-power laser collimating lenses.

[0008] The present application provides a multi-scale glue point detection method, comprising the following steps:

[0009] Step one, using an intelligent laser lens dispensing platform, image acquisition is performed on the laser chip before bonding; and based on the acquired image, an LLAS dataset is established;

[0010] A multi-level attention-guided context aggregation network model is established;

[0011] Step two, applying the multi-level attention-guided context aggregation network model to process the images in the LLAS dataset, and segmenting out the glue point target;

[0012] Step three, based on the segmented glue point target, the glue point detection result is obtained.

[0013] The multi-level attention-guided context aggregation network model uses a U-Net model as the backbone network, and through the attention mechanism based on Kolmogorov-Arnold representation, it fuses the interactive self-attention guided Kolmogorov-Arnold representation module, the multi-perception space attention module, the KAN enhanced channel attention module and the multi-dimensional feature fusion module.

[0014] The multi-level attention-guided context aggregation network model comprises sequentially connected down-sampling layers, fusion layers and up-sampling layers, the fusion layers comprise an interactive self-attention-guided Kolmogorov-Arnold representation module, a multi-perception spatial attention module, a KAN-enhanced channel attention module and a multi-dimensional feature fusion module; the KAN-enhanced channel attention module is applied to the fourth layer in the down-sampling process, the fifth layer in the down-sampling process and the fourth layer in the up-sampling process; and the multi-dimensional feature fusion module is applied to the skip connection layers of the fourth layer and the fifth layer;

[0015] The specific process of applying the multi-level attention-guided context aggregation network model to process the images in the LLAS dataset is as follows:

[0016] The input feature image T∈(B,C,H,W) is subjected to three down-sampling layers arranged in succession to obtain a remolding feature map T1, an extreme value feature map T2 and a mean value feature map T3, respectively;

[0017] The extreme value feature map T2 and the mean value feature map T3 are taken as inputs and are transmitted to a 3×3 deep convolution block to obtain image information T′2 and image information T′3, respectively;

[0018] The number of channels in the U-Net model is halved;

[0019] The image information T′2 is transmitted to the interactive self-attention-guided Kolmogorov-Arnold representation module, and the image information T′3 is transmitted to the multi-perception spatial attention module; the image information T′2 and the image information T′3 are subjected to processing by the interactive self-attention-guided Kolmogorov-Arnold representation module and the multi-perception spatial attention module, respectively, to obtain high-resolution feature information;

[0020] The up-sampling layer receives the high-resolution feature information processed by the interactive self-attention-guided Kolmogorov-Arnold representation module and the multi-perception spatial attention module, and obtains segmented gel point targets based on the high-resolution feature information.

[0021] The interactive self-attention-guided Kolmogorov-Arnold representation module comprises an additive interactive self-attention module and a Kolmogorov-Arnold network, and the Kolmogorov-Arnold network is provided with two layers;

[0022] The operation process of the additive interactive self-attention module is as follows:

[0023] 1) The input feature image T∈(B,C,H,W) is subjected to a channel expansion convolution operation to expand in the channel dimension to obtain T′∈(B,3C,H,W);

[0024] 2) Q, K and V are obtained through channel block operation, where Q∈(B,C,H,W), K∈(B,C,H,W), V∈(B,C,H,W), Q is Query, K is Key, and V is Value;

[0025] 3) The similarity function is defined as the sum of the context scores of Q and K:

[0026] Sim(Q,K)=F(Q)+F(K), stF(Q)=S(C(Q));

[0027] Where: F(·) is the context mapping function, and it contains the necessary information interaction; F(·) is specifically defined as spatial attention S(·)∈(B,C,H,W) and channel attention C(·)∈(B,C,H,W);

[0028] 4) Based on the similarity function definition, which is the sum of the context scores of Q and K, the final output feature Attn(Q,K,V) is obtained;

[0029] After further processing the output feature Attn(Q,K,V) with depthwise separable convolution, the original input image T∈(B,C,H,W) is added as a residual to obtain the final output feature T″∈(B,C,H,W) of the additive interactive self-attention module.

[0030] The operation process of the Kolmogorov-Arnold network is as follows:

[0031] The final output feature T″∈(B,C,H,W) of the additive interactive self-attention module is reconstructed into a two-dimensional patch feature T″. r ∈(B*H*W,C);

[0032] The two-dimensional patch feature T″ r Projecting ∈(B*H*W,C) to a lower dimension The original dimension C is then embedded back into the original tensor dimension;

[0033] The final output features T″∈(B,C,H,W) of the additive interactive self-attention module are added as residuals to obtain the final IKA output image T. IKA ∈(B,C,H,W).

[0034] The operation process of the multi-sensory spatial attention module is as follows:

[0035] (i) Extract features from the input feature image T∈(B,C,H,W) by passing it through four convolutional kernels of sizes 1×1, 3×3, 5×5 and 7×7 respectively.

[0036] (ii) using average pooling to compress the image features extracted in step (i), and then restoring the image to the original size by nearest neighbor interpolation to obtain a remodeled feature map T1e(B,1,H*W);

[0037] (iii) obtaining extreme value feature information and mean value feature information of the input feature image T e(B,C,H,W) by global maximum pooling and global average pooling operations on the second branch and the third branch of the multi-perception spatial attention module, respectively, and reconstructing the extreme value feature information and the mean value feature information into extreme value feature map T2e(B,H*W,1) and mean value feature map T3e(B,1,H*W), respectively;

[0038] (iv) performing matrix multiplication on the extreme value feature map T2and the mean value feature map T3to obtain a feature map T4e(B,H*W,H*W);

[0039] Performing matrix multiplication on the remodeled feature map T1and the feature map T4, and then performing reconstruction and Sigmoid activation operations on the input feature image T e(B,C,H,W) to obtain a spatial attention feature map T5e(B,C,H,W).

[0040] The operation process of the KAN enhanced channel attention module is as follows:

[0041] (I) The input feature image T e(B,C,H,W) is compressed by global maximum pooling and global average pooling, respectively;

[0042] (II) The compressed image features are reconstructed to obtain reconstructed feature maps T'1e(B,1,C) and T'2e(B,1,C), respectively;

[0043] The reconstructed feature maps T'1e(B,1,C) and T'2e(B,1,C) are fused by a one-dimensional convolution kernel with a size of 3, and then activated by a Sigmoid function and reconstructed to score the original input feature image T e(B,C,H,W) to obtain the final channel attention output feature T'3e(B,C,H,W).

[0044] The multi-dimensional feature fusion module includes a low layer, a current layer and a high layer, and the low layer, the current layer and the high layer;

[0045] Let the low layer input feature be The current layer input feature is T e(B,C,H,W), and the high layer input feature is

[0046] The operation process of the multi-dimensional feature fusion module is as follows:

[0047] The low layer input feature is fused with the current layer input feature T e(B,C,H,W) and the high layer input feature T' e(B,C,H,W) by a one-dimensional convolution kernel with a size of 3, and then activated by a Sigmoid function and reconstructed to score the original input feature image T e(B,C,H,W) to obtain the final multi-dimensional feature fusion output feature T'' e(B,C,H,W). and high-level input features are aligned to ensure the same feature size; and high-level input features are aligned to ensure the same feature size;

[0048] low-level input features After being down-sampled by the detail-preserving down-sampling, low-level features T' are obtained low ∈(B,C,H,W);

[0049] After being refined by a 1x1 convolution layer, the current layer input features T∈(B,C,H,W) are extracted to obtain current layer features T'∈(B,C,H,W);

[0050] high-level input features After being up-sampled by the linear interpolation method, high-level features T' are obtained high ∈(B,C,H,W);

[0051] low-level features T' low , current layer features T' and high-level features T' high After being activated by the 1x1 convolution operation and the Sigmoid function respectively, spatial weight matrices Wlow, Wcur and Whigh from different layers are generated, and each element in the spatial weight matrices Wlow, Wcur and Whigh represents the possibility of the corresponding point belonging to the target;

[0052] After being spliced, the spatial weight matrices Wlow, Wcur and Whigh are activated by the 1x1 convolution operation and the Sigmoid function to generate the final spatial weight matrix W;

[0053] low-level features T' low , current layer features T' and high-level features T' high are fused by the intelligent channel integration module to obtain ICI fusion features T ICI ∈(B,C,H,W);

[0054] The spatial weight matrix W and the T1 fused by the channel are multiplied element by element to realize the joint weighted fusion in the spatial and channel dimensions, and the final multi-dimensional feature fusion module output result T is obtained MFF ∈(B,C,H,W).

[0055] The multi-dimensional feature fusion module comprises a detail-preserving down-sampling module and an intelligent channel integration module.

[0056] The operation process of the detail-preserving down-sampling module is as follows:

[0057] The input low-level input features The input image is cut into four parts symmetrically along its height and width.

[0058] The four features after segmentation and compression, along with the features after max pooling and average pooling, are concatenated and then finely extracted using a 1×1 convolutional kernel. The final output is the downsampled low-level feature T′. low ∈(B,C,H,W);

[0059] The operation process of the intelligent channel integration module is as follows:

[0060] For low-level features T′ that are already aligned in dimension low Current layer features T′ and high-level features T′ high Using channel splitting technology, the feature map is split into four equal parts along the channel, which are then used to obtain low-level segmentation features. Current layer segmentation features and high-level segmentation features Among them, T′ low_i For the i-th partition of the low-level features, T′ i For the i-th partition of the current layer features, T′ high_i Let i be the i-th partition of the high-level features, where i = (1, 2, 3, 4).

[0061] Low-level segmentation features Current layer segmentation features High-level segmentation features The following formula is used to calculate the weighting of a channel in a smart channel integration module based on the characteristics of the target:

[0062] p i =Sigmoid(T′) i );

[0063]

[0064] T ICI =Concat(T′) g_1 ,T′ g_2 ,T′ g_3 ,T′ g_4 );

[0065]

[0066] Where, p i For feature T′ i The feature weights, T′, obtained after applying the Sigmoid activation function. g_i This represents the intelligent aggregation result for each segment;

[0067] The four equal parts after segmentation are spliced in the channel dimension to obtain T ICI ∈R B×C×H×W ;

[0068] When p i > 0.5, the multi-level attention guided context aggregation network model enhances the weight of the fine-grained features transmitted by the low-level network;

[0069] When p i < 0.5, the multi-level attention guided context aggregation network model enhances the weight of the semantic-rich global features transmitted by the high-level network.

[0070] Compared with the prior art, the present application has the following beneficial effects:

[0071] The present application proposes a multi-level attention guided context aggregation network (MACAN) based on a self-built LLAS dataset. The purpose is to realize real-time and robust detection of glue points on laser chips and provide necessary quality assurance for realizing high coupling efficiency of high-power laser lenses. The multi-level attention driving mechanism enhances various attention mechanisms by widely applying KAN networks in different convolutional layers, effectively guiding complex nonlinear information to concentrate in the target area. The fusion network improves the encoding efficiency and robustness of the network by adaptively selecting and fusing target information from multi-level features in the convolutional layer. IKA reduces the multiplication operation in the traditional self-attention module by adding an interactive self-attention (AISA) submodule, thereby reducing the computational cost. This submodule introduces channel and spatial attention mechanisms, enhancing the interaction between spatial and channel information. In addition, the application of KAN network effectively captures global context information, enabling accurate extraction of target features while suppressing the interference of complex background noise. MPSA enhances the model's ability to capture different scale targets by introducing a multi-perception mechanism in the spatial attention module to address the dual target task of the present application. KECA enhances the model's ability to model complex nonlinear relationships by incorporating KAN networks into the channel attention module, which is applied in deep networks, thereby improving the model's performance in handling complex tasks and high-dimensional channel data challenges. The MFF module suppresses noise in the fusion process by adaptively selecting and fusing target information from multi-level features, enhancing feature representation capabilities and significantly improving the network's encoding efficiency and robustness.

[0072] In addition to the purposes, features and advantages described above, the present application has other purposes, features and advantages. The present application will be further described in detail below with reference to the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0073] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application, illustrate the preferred embodiments of the application and assist in

[0074] Figure 1 is a structural schematic diagram of the MACAN model in the embodiment of the application;

[0075] Figure 2 is a structural schematic diagram of the IKA module in the embodiment of the application;

[0076] Figure 3 is a structural schematic diagram of the MPSA module in the embodiment of the application;

[0077] Figure 4 is a structural schematic diagram of the KECA module in the embodiment of the application;

[0078] Figure 5 is a structural schematic diagram of the MFF module in the embodiment of the application;

[0079] Figure 6 is a structural schematic diagram of the DPD module in the embodiment of the application;

[0080] Figure 7 is a structural schematic diagram of the ICI module in the embodiment of the application;

[0081] Figure 8 is a result schematic diagram of the ablation experiment in the experimental example of the application;

[0082] Figure 9 is a result schematic diagram of the comparative experiment in the experimental example of the application;

[0083] Figure 10 is a ROC curve diagram in the experimental example of the application;

[0084] Figure 11 is a result schematic diagram of the glue point parameter measurement in the embodiment of the application. DETAILED DESCRIPTION

[0085] To make the above objects, features and advantages of the application more clear and easy to understand, the specific embodiments of the application are described in detail below with reference to the drawings. It should be noted that the drawings of the application are in a simplified form and all use non-accurate proportions, only to facilitate and clearly assist in explaining the implementation of the application; the number of several mentioned in the application is not limited to the specific number in the example of the drawings; the directions or position relationships mentioned in the application, such as 'front','middle', 'back', 'left', 'right', 'top', 'bottom', 'top', 'bottom','middle', are based on the directions or position relationships shown in the drawings of the application, and do not indicate or imply that the devices or parts referred to must have a specific direction, nor can it be understood as a limitation of the application.

[0086] Embodiment:

[0087] The multi-scale glue point detection method provided by the application comprises the following steps:

[0088] Step one, using an intelligent laser lens dispensing platform, image acquisition is performed on the laser chip before bonding; and an LLAS data set is established based on the acquired image;

[0089] A multi-level attention guided context aggregation network model is established;

[0090] Step two, the multi-level attention guided context aggregation network model is applied to process the images in the LLAS data set, and the glue point target is segmented out, so as to realize more real-time and robust detection of the glue point of the laser collimating lens, and provide necessary quality guarantee for realizing high coupling efficiency of the high-power laser collimating lens;

[0091] Step three, parameter calculation is performed on the segmented glue point target to obtain a glue point detection result.

[0092] Specifically, by performing parameter calculation on the segmented glue point target, it is judged whether the glue point is qualified or not, so as to prompt the operator to perform what operation next.

[0093] Further, the way of judging whether the glue point is qualified or not is as follows:

[0094] Glue point area;

[0095] The total area in the profile is calculated according to the segmented glue point profile, and an excessively large glue point area will lead to glue overflow and large glue offset after the collimating lens is bonded. And an excessively small glue point area will lead to the risk of loosening of the collimating lens;

[0096] Glue point roundness;

[0097] The glue point is fitted as an ellipse, and then the long axis and short axis sizes of the ellipse are judged. In an ideal state, the glue point is circular, and an excessively long long axis or an excessively short short axis will lead to risks such as glue overflow or unreliable bonding.

[0098] Glue point spacing;

[0099] According to the ellipse centers of two glue points, the glue point spacing, including horizontal spacing and vertical spacing, is measured. An excessively large horizontal distance will lead to exceeding the lens length, and an excessively small horizontal distance will lead to the glue points being concentrated in the middle section of the lens and being suspended at both ends, affecting the bonding accuracy and reliability. The vertical distance is ideally 0, and an excessively large vertical distance will lead to risks such as the glue points being separated from the lens or the lens being offset.

[0100] Further preferably, according to the design size of the collimating lens and long-term practical engineering experience, the index threshold values shown in Table 1 are determined.

[0101] Table 1: Comparison and evaluation of measurement results of glue dot quality

[0102]

[0103] In actual production, most of the glue dots meet the design requirements. In order to highlight the ability of MACAN in segmenting abnormal glue dots, in Figure 11 , five typical pictures containing normal and abnormal glue dots are selected, and the parameters of the glue dots segmented by the MACAN network and the ground truth are measured. Table 1 quantitatively analyzes the measurement results and evaluates the quality of the glue dots. As can be seen from Table 1, in the measurement results of the area, roundness, spacing and other parameters of the target, the measurement results of the method proposed in the application have smaller deviation from the true results. At the same time, the glue dot detection can cover a wide range of scales from 0.002 to 1.421 mm 2 , and has smaller deviation from the true value. Compared with the traditional method, the MACAN network reduces the need for manual inspection and improves the automation level of the system.

[0104] Further, as shown in Figure 1 , MACAN adopts the commonly used U-Net model in the field of image segmentation as the backbone network, and fuses the IKA module, MPSA module, KECA module and MFF module based on the new attention mechanism of Kolmogorov-Arnold representation, to enhance the multi-scale and multi-level information fusion. MACAN is based on the commonly used U-Net backbone network in the field of image segmentation, aiming to realize the segmentation of glue dot targets.

[0105] Further, the MACAN includes a downsampling layer, a fusion layer and an upsampling layer.

[0106] Further, the process of processing the image by the MACAN is as follows:

[0107] The input feature image T∈(B,C,H,W) is processed by three continuously arranged downsampling layers to obtain a remolding feature map T1, an extreme value feature map T2 and a mean value feature map T3, respectively;

[0108] The extreme value feature map T2 and the mean value feature map T3 are input and transmitted to a 3x3 deep convolution block to obtain image information T′2 and image information T′3, respectively. Specifically, the 3x3 deep convolution block can enhance the expression ability of local features at low computational cost.

[0109] The number of channels in the U-Net model is halved;

[0110] The image information T'2 is transmitted to the IKA module, and the image information T'3 is transmitted to the MPSA module. The image information T'2 and the image information T'3 are processed by the IKA module and the MPSA module respectively, and the target features are directed to a specific spatial position range to obtain high-resolution feature information. The IKA module can directly capture the relationship between pixels in the global range without being limited by the size of the convolution kernel. When segmenting boundaries or complex shapes, global information can help more accurately identify different regions. The MPSA module is based on a spatial attention mechanism and can highlight important regions and target positions in the image, thereby reducing interference from the background and irrelevant regions and improving the accuracy of segmentation. After feature extraction, the image information is processed by the IKA module and the MPSA module, and the target features are directed to a specific spatial position range.

[0111] The up-sampling layer receives high-resolution feature information processed by the IKA module and the MPSA module, which helps the MACAN to retain more details when restoring the resolution.

[0112] The KECA module is applied to the 4th layer in the down-sampling process, the 5th layer in the down-sampling process, and the 4th layer in the up-sampling process. During the process of processing deep network rich abstract features, the KECA module enhances the attention to key features and suppresses irrelevant or redundant features, thereby improving the expression ability of the features.

[0113] The MFF module is applied to the skip connection layers of the 4th layer and the 5th layer. By adaptively selecting and fusing target information in multi-level features, the MFF module can effectively suppress noise and retain target details, thereby enhancing the information coding efficiency and robustness of the model.

[0114] Further, referring to Figure 2 As shown in the figure, the IKA module (interactive self-attention guided Kolmogorov-Arnold representation module) includes an AISA module and a KAN network. The AISA module (additive interactive self-attention module) uses channel and spatial attention as a new form of information interaction by constructing a novel additive similarity function, thereby avoiding complex matrix multiplication and Softmax operation.

[0115] Further preferably, the operation process of the AISA module is as follows:

[0116] 1) The input feature image T∈(B,C,H,W) is expanded in the channel dimension by a channel expansion convolution operation to obtain T'∈(B,3C,H,W);

[0117] 2) Obtain Query (Q), Key (K) and Value (V) by channel chunk operation, where Q e (B, C, H, W), K e (B, C, H, W), V e (B, C, H, W);

[0118] 3) Define the sum of context scores of Q and K as similarity function:

[0119] Sim(Q, K) = F(Q) + F(K), s.t. F(Q) = S(C(Q));

[0120] Where: F(·) is the context mapping function, which contains the necessary information interaction. Specifically, F(·) is concretized as spatial attention S(·) e (B, C, H, W) and channel attention C(·) e (B, C, H, W), and residual connection is used for each spatial attention and channel attention, which adds the information before the attention module processing as residual; Through the combination of spatial and channel attention, the AISA module can promote multi-information interaction and effectively capture global context information. At the same time, the similarity function adds the results of spatial and channel attention, which is used to replace the traditional dot product attention mechanism, and depth separable convolution is used to further extract the added information. These measures significantly reduce the computational complexity, making the model more suitable for deployment on resource-constrained devices.

[0121] 4) The sum of context scores obtained by additive similarity function is further processed by depth separable convolution, which is used to weight V to obtain the final output feature Attn(Q, K, V);

[0122] The expression of output feature Attn(Q, K, V) is as follows:

[0123]

[0124] Where: Ψ(·) is a linear transformer, which is used to integrate context information; is a matrix multiplication operation.

[0125] After the output feature Attn(Q, K, V) is further processed by depth separable convolution, the original input image T e (B, C, H, W) is added as residual to obtain the final output feature T" e (B, C, H, W) of the AISA module.

[0126] Specifically, the calculation process of the entire AISA module is summarized as:

[0127] AISA(T) = DWConv(Attn(Chunk(Conv(T)))) + T.

[0128] Further preferably, in the present application, a 2-layer KAN network (Kolmogorov-Arnold network) is provided; the operation process of the KAN network is as follows:

[0129] The final output feature T'' of the AISA module is reconstructed into a series of flattened two-dimensional patches T'' r ∈(B*H*W,C);

[0130] The two-dimensional patches T'' r ∈(B*H*W,C) are projected to a lower dimension to reduce the parameter overhead, and then embedded and projected back to the original dimension C, and finally reshaped to the original tensor dimension;

[0131] The final output feature T'' of the AISA module is added as a residual to obtain the final IKA output image T IKA ∈(B,C,H,W).

[0132] Specifically, the entire process of the KAN network can be described as:

[0133] KAN(T'') = (Reshape(K2(K1(Reshape(T''))))) + T''.

[0134] Further, the entire calculation process of the IKA module can be summarized as:

[0135] IKA(T) = KAN((AISA(T))).

[0136] Further, in the dual-target detection task of the present application, the size of the gel dots has certain randomness, and even there are both large-size targets and small-size targets on the same chip. Therefore, in the channel attention module, a multi-receptive field perception mechanism is introduced to enhance the model's ability to capture different scale targets, which is called a multi-perception spatial attention module (MPSA).

[0137] Specifically, referring to Figure 3 , it is assumed that the input feature image T ∈(B,C,H,W), and the detailed steps of the MPSA module (multi-perception spatial attention module) are as follows:

[0138] (i) The input feature image T ∈(B,C,H,W) is sequentially subjected to feature extraction by 4 convolution kernels with sizes of 1×1, 3×3, 5×5 and 7×7, to obtain image features; by subjecting the input feature image T ∈(B,C,H,W) to convolution kernels of different sizes, the MACAN model is given the perception ability of multi-receptive field.

[0139] (ii) using average pooling to compress the image features extracted in step (i), and then restoring the image to the original size through nearest neighbor interpolation (NNI) to obtain a reshaped feature map T1e(B,1,H*W). Specifically, average pooling effectively compresses the redundant information in the image by averaging the pixel values in the local region, while also helping to remove some noise, which enables the network to focus on higher-level features. When restoring the image size through nearest neighbor interpolation (NNI), no new pixel values or grayscale information is introduced, but only the nearest pixels are simply copied. This method avoids introducing unnecessary assumptions, preserves the original features of the image, and reduces information distortion during image processing.

[0140] (iii) on the second branch and the third branch of the MPSA module, respectively through global maximum pooling (GMP) and global average pooling (GAP) operations, to obtain extreme value feature information and mean value feature information of the input feature image T e(B,C,H,W), and to reconstruct the extreme value feature information and the mean value feature information into extreme value feature map T2e(B,H*W,1) and mean value feature map T3e(B,1,H*W), respectively.

[0141] (iv) performing matrix multiplication on the extreme value feature map T2 and the mean value feature map T3 to obtain a feature map T4e(B,H*W,H*W).

[0142] Performing matrix multiplication on the reshaped feature map T1e(B,1,H*W) and the feature map T4e(B,H*W,H*W), and then performing reconstruction and Sigmoid activation operations on the result, to score the input feature image T e(B,C,H,W) and obtain a spatial attention feature map T5e(B,C,H,W).

[0143] Further preferably, multiple matrix multiplications are performed in the MPSA module, which can effectively calculate the spatial correlation between different regions of the input feature map, enabling the model to dynamically adjust the importance of each region in the feature map, thereby highlighting the target region and suppressing background interference. This operation not only efficiently fuses global information, but also enhances the model's perception of the target region, improving the accuracy of segmentation or detection.

[0144] Preferably, the specific process of the MPSA module can be summarized by the following equation:

[0145] T1 = Reshape (NNI (AvgP (Conv1 (T) + Conv3 (T) + Conv5 (T) + Conv7 (T))) ) ;

[0146] T2 = Reshape(GMP(T));

[0147] T3 = Reshape(GAP(T));

[0148]

[0149] wherein: Conv1(·), Conv3(·), Conv5(·) and Conv7(·) represent convolution kernels with sizes of 1x1, 3x3, 5x5 and 7x7 respectively; AvgP(·) is an average pooling operation; NNT(·) is an operation of the NNT module; Reshape is a reshape operation on a feature tensor; GMP(·) is a global maximum pooling; GAP(·) is a global average pooling; denotes a matrix multiplication operation; Sigmoid is an activation function.

[0150] Further, referring to Fig. 2, the detailed steps of the KECA module (KAN enhanced channel attention module) are as follows: Figure 4

[0151] (I) The input feature image T ∈ (B, C, H, W) is compressed by global maximum pooling (GMP) and global average pooling (GAP) respectively, which can avoid the over-attention to a certain type of features caused by using a single pooling method.

[0152] (II) The compressed image features are reconstructed to obtain reconstructed feature maps T'1 ∈ (B, 1, C) and T'2 ∈ (B, 1, C) respectively; the compressed image features are reconstructed to use two KAN networks (named KAN1 and KAN2 respectively) to more effectively capture the complex nonlinear dependence between channel features, and the specific working process of the KAN network is the same as that in the IKA module.

[0153] The reconstructed feature maps T'1 ∈ (B, 1, C) and T'2 ∈ (B, 1, C) are fused by a one-dimensional convolution kernel with a size of 3, and then activated by a Sigmoid function and reconstructed to score the original input feature image T ∈ (B, C, H, W) to obtain the final channel attention output feature T'3 ∈ (B, C, H, W).

[0154] Preferably, the specific process of the KECA module can be summarized as follows:

[0155] T'1 = Reshape(KAN1(Reshape(GMP(T))));

[0156] ​T'2 = Reshape(KAN2(Reshape(GAP(T))));

[0157]

[0158] wherein the specific calculation process of KAN1(·) and KAN2(·) is the same as that of KAN(·) in the IKA module; Conv1D(·) is a one-dimensional convolution kernel with a size of 3.

[0159] Further, in the laser collimator glue point detection of the present application, the glue point has a random shape. This requires the deep learning network not only to accurately identify the position of the glue point, but also to clearly identify the irregular contour of the glue point, and to effectively suppress noise interference. Therefore, a multi-dimensional feature fusion module (MFF) is set on the skip connection layers of the fourth and fifth layers of the MACAN network, aiming to fuse high-level image features and low-level image features, so that the extracted features contain clear texture and edge information, and are not affected by noise.

[0160] Referring to Figure 5 , the MFF module (multi-dimensional feature fusion module) is used to fuse image features from three different layers; that is, the MFF module includes a low layer, a current layer and a high layer, and the low layer, the current layer and the high layer;

[0161] Let the low layer input feature be the current layer input feature be T∈(B,C,H,W), and the high layer input feature be

[0162] The specific operation process is as follows:

[0163] 1) Align the low layer feature with the image feature T∈(B,C,H,W) of the current layer and the high layer feature with the image feature T∈(B,C,H,W) of the current layer, to ensure the same feature size;

[0164] 2) Generate a spatial weight matrix and perform fusion;

[0165] The low layer feature T′ low , the current layer feature T′ and the high layer feature T′ high are respectively processed by a 1x1 convolution operation and a Sigmoid function activation, to generate the spatial weight matrices Wlow, Wcur and Whigh from different layers, and each element in the spatial weight matrices Wlow, Wcur and Whigh represents the possibility that the corresponding point belongs to the target;

[0166] The spatial weight matrix Wlow, Wcur and Whigh are spliced, and after 1*1 convolution operation and Sigmoid function activation, the final spatial weight matrix W is generated. The spatial weight matrix W combines the reinforced target details and accurate target positioning, while effectively suppressing the background noise.

[0167] 3), intelligently fuse the channels of image features of each layer;

[0168] Low-level feature T' low , current layer feature T' and high-level feature T' high are fused by the intelligent channel integration module (ICI) to obtain ICI fusion feature T ICI ∈(B,C,H,W); The ICI module can intelligently select different layers of channels for fusion according to the characteristics of the target.

[0169] Element-wise multiplication of the spatial weight matrix W and the channel fused T1 is performed, realizing joint weighted fusion in spatial and channel dimensions, and obtaining the final multi-dimensional feature fusion module (MFF) output result T MFF ∈(B,C,H,W). Element-wise multiplication of the spatial weight matrix and the channel fusion feature can effectively integrate spatial information and channel information, improve the precision and robustness of the model, and especially perform better in the task of accurately identifying target details from complex backgrounds.

[0170] Further preferably, in order to preserve target details, the low-level image feature is first down-sampled by a detail-preserving down-sampling (DPD) to obtain a low-level feature T' low ∈(B,C,H,W);

[0171] The current layer image feature T∈(B,C,H,W) is refined and extracted by a 1*1 convolution layer to obtain a current layer feature T'∈(B,C,H,W);

[0172] The high-level image feature is up-sampled by a linear interpolation method to obtain a high-level feature T' high ∈(B,C,H,W).

[0173] Preferably, referring to Figure 6 , the operation process of the DPD module (detail-preserving down-sampling module) is as follows:

[0174] The input low-level image feature is cut into four parts in a symmetrical manner along the height and width directions;

[0175] The input features are also subjected to maximum pooling operation and average pooling operation for feature extraction, taking into account the characteristics of the maximum pooling operation that can extract the most significant features in the local area, and the characteristics of the average pooling operation that can smooth details, enhancing the expression ability of the model and improving the robustness.

[0176] The four features after cutting and compression, and the features after maximum pooling and average pooling are spliced, and after a 1x1 convolution kernel is used for fine feature extraction, the final output is the down-sampled feature T' low ∈(B,C,H,W).

[0177] Further preferably, the calculation process of the DPD module is summarized by the formula:

[0178] DPD(T low )=Conv(Concat(Cut(T low ),Maxpool(T low ),Avgpool(T low )))。

[0179] Preferably, referring to FIG. 1, the operation process of the ICI module (intelligent channel integration module) is as follows: Figure 7

[0180] For the low-level feature T′ low , the current layer feature T′ and the high-level feature T′ high that have been aligned in the dimension, the channel splitting technology is used to split the feature map into 4 equal parts in the channel, respectively obtaining the low-level split feature the current layer split feature and the high-level split feature , wherein T′ low_i is the i-th divided part of the low-level feature, T′ i is the i-th divided part of the current layer feature, and T′ high_i is the i-th divided part of the high-level feature, i=(1, 2, 3, 4);

[0181] The low-level split feature the current layer split feature and the high-level split feature are calculated according to the following formula, so that the ICI module can intelligently select which layer of the channel to increase the weight according to the characteristics of the target:

[0182] p i =Sigmoid(T′ i );

[0183]

[0184] T​ICI = Concat(T' g_1 , T' g_2 , T' g_3 , T' g_4 ) ;

[0185]

[0186] where p i is the feature weight of T' i after the activation function Sigmoid, and T' g_i represents the intelligent aggregation result of each partition.

[0187] The four equal parts after partitioning are spliced in the channel dimension to obtain T ICI ∈ R B×C×H×W In such a setting, when p i > 0.5, MACAN enhances the weight of the fine-grained features delivered by the low-level network; if p i < 0.5, MACAN enhances the weight of the semantic-rich global features delivered by the high-level network, thereby achieving intelligent aggregation of features at different levels.

[0188] Preferably, the entire MFF calculation process can be summarized in the formula as follows:

[0189] W low (T low ) = Sigmoid(Conv(DPD(T low ))) ;

[0190] W cur (T) = Sigmoid(Conv(Conv(T))) ;

[0191] W high (T high ) = Sigmoid(Conv(Up(T high ))) ;

[0192] W(W low , W cur , W high ) = Sigmoid(Conv(Concat(W low , W cur , W high ))) ;

[0193]

[0194] Experimental Example:

[0195] (I) Experimental Results and Analysis

[0196] Based on the images taken by the laser chip before bonding, the LLAS dataset established contains a total of 3070 glue point pictures, and the size of the picture is 256*192 (width*height).

[0197] Specifically, 2520 (82%) of them are used for training, 280 (9%) for testing, and the remaining 270 (9%) for verification.

[0198] All experiments of the present application are calculated by 12G NVIDIA RTX 3060 GPU, the calculation process is completed based on PyTorch architecture, the Adam optimizer is used for parameter optimization, the binary cross-entropy loss function (BCE Loss) is used to calculate the loss, which is used to measure the gap between the predicted value of the model and the true label. And use the intersection over union IoU), the average intersection over union (mIoU), F1 score, Dice, receiver operating characteristic curve (ROC) and other indicators to evaluate the performance of MACAN.

[0199] Specifically, the IoU index is used to measure the overlap between the predicted result and the true label, and evaluate the spatial consistency; mIoU averages the IoU of all samples, which is used for overall performance evaluation; F1 considers precision and recall comprehensively, and reflects the balanced performance of the classification model; Dice evaluates the overlap between the predicted value and the true value by twice the intersection divided by the total area, and is more sensitive to small targets. The larger the value of the above indicators, the more accurate the model prediction. ROC describes the classification ability of the model at different thresholds, and shows the relationship between true positive rate (TPR) and false positive rate (FPR). The larger the area surrounded by the curve and the X-axis, the better the performance of the model. The specific calculation formula is as follows:

[0200]

[0201]

[0202] Wherein, A ∩ and A ∪ respectively represent the area of the intersection and the union, N and j represent the total number of samples and the jth sample, respectively, True Positive represents the number of samples correctly predicted as positive, False Positive represents the number of samples actually negative but incorrectly predicted as positive, False Negative represents the number of samples actually positive but incorrectly predicted as negative, and True Negative represents the number of samples correctly predicted as negative. Recall is also called TPR in some scenarios.

[0203] (II) Ablation experiment

[0204] To investigate the effectiveness of each module designed in MACAN, ablation experiments were carried out, and the evaluation results are shown in Table 1 and Figure 8 .

[0205] Specifically, the classic image segmentation network U-Net was taken as the baseline. The results of the ablation experiment show that the numerical evaluation performance of the baseline network is poor. By adding the KECA module in the deep layer of the U-Net backbone network, the performance of the model in processing complex high-dimensional channel data is significantly improved, and the mIoU, F1 and Dice indicators are increased by 1.6%, 0.7% and 0.9% respectively. When the IKA module is further introduced in the skip connection, the mIoU, F1 and Dice indicators are further increased by 1.3%, 0.5% and 0.7% respectively. When the MPSA module is introduced in the skip connection, the mIoU, F1 and Dice indicators are increased by 1.0%, 0.2% and 0.5% respectively. But if the processing results of the IKA and MPSA modules are spliced in the skip connection at the same time, the mIoU, F1 and Dice indicators are increased by 2.0%, 0.7% and 1.1% compared with only adding the KECA module. This shows that since the IKA module can improve the global perception ability, and the MPSA module has the ability of fine positioning of multi-scale targets, when the two are combined, a strong synergistic effect can be formed, which can more effectively cope with complex background and target size difference. Finally, the setting of the MFF module adaptively fuses the target information in the multi-level features, and further improves the mIoU, F1 and Dice indicators by 0.8%, 0.3% and 0.4% respectively. Finally, compared with the baseline network, the MACAN network improves the mIoU, F1 and Dice indicators by 4.4%, 1.7% and 2.4% respectively.

[0206] In Figure 8Some qualitative results are also given to further demonstrate the effect of each module. In the pictures of sequence (1) and sequence (2), there is no blob in itself, but the baseline network incorrectly identifies the existence of the blob. In particular, since the dirt condition on the chip in sequence (1) is more serious, the baseline network also identifies more false targets. When the KECA module is added, the false alarm problem is significantly improved as the module improves the performance in processing complex high-dimensional channel data. After adding the IKA and MPSA modules respectively, the false alarm problem is further improved, especially for sequence (2), the false alarm problem has been completely eliminated. When the model simultaneously adds the IKA and MPSA modules, sequence (1) with more complex background information also completely suppresses the occurrence of false alarm. In the pictures of sequence (3), there are blobs in the dispensing target area, but the pixel brightness of the left blob is similar to the background area, and the contrast is insufficient, so it is difficult for the baseline network to segment the accurate boundary, and there is also a certain degree of false alarm. With the addition of the KECA module, the false alarm problem is eliminated, and at the same time, the difference between the identified blob area and the true label is reduced, but the contour boundary is still not accurate enough. When the IKA and MPSA modules are added respectively, the accuracy of identifying the contour boundary is improved, and the identification result of the IKA module on the contour boundary is relatively more accurate. When the two modules are added at the same time, especially when the MFF module is further added, the identification accuracy of the contour boundary is significantly improved, and the final identification result of the MACAN network on the contour boundary of the target in sequence (3) is highly consistent with the true label. In the pictures of sequence (4), there are two very small blobs, and the left blob target is particularly small. The baseline network can identify the existence of the right blob, but the contour is quite different from the true label, and the small target on the left cannot be identified. With the addition of the KECA and IKA modules, the small target on the left cannot be identified. When the MPSA module with multi-receptive field perception mechanism is added, the small target on the left is detected. In the subsequent experiments of simultaneously adding the IKA and MPSA modules, and further adding the MFF module, the accuracy of identifying the blob is gradually improved. In the pictures of sequence (5), the blob target is ideal, and even the baseline network can accurately segment the blob target, but careful comparison of the contour shape can still show the advantage of the modules proposed in this application in improving the segmentation accuracy.

[0207] Table 2: Ablation experiment results

[0208]

[0209] (III) Comparison with other advanced methods

[0210] In the existing research reports, there is no segmentation algorithm for the microlens glue points of the laser. In order to evaluate the performance of MACAN on the LLAS dataset, the network trained on the LLAS dataset is used to form a contrast, which has achieved excellent results in the field of infrared small target recognition, medical image segmentation, etc. The selected methods include TransUNet, MTU-Net, DNA-Net, HCF-Net, MRF3Net and U-KAN. TransUNet introduces the ViT architecture into the U-Net network for medical image segmentation tasks, and MTU-Net further constructs a multi-polar feature extraction network based on ViT. DNA-Net fuses a dense nested attention network structure dedicated to infrared small target detection tasks, HCF-Net improves the detection performance of infrared small targets by integrating a hierarchical context fusion network, MRF3Net fuses multi-receptive field perception and effective feature fusion strategy for infrared small target detection, and U-KAN realizes medical image segmentation by fusing KAN network in U-Net. The quantitative results of the comparative test are shown in Table II, and MACAN benefits from the outstanding advantages of the multi-level attention driven fusion network in dealing with complex nonlinear problems. In the image segmentation results of all networks, the mIoU, F1 and Dice three indicators are the highest. The multi-polar feature extraction network MTU-Net based on ViT achieves the second optimal result. Due to the fusion of KAN network with strong nonlinear expression ability, U-KAN also achieves a result close to MTU-Net. TransUNet and HCF-Net perform at a medium level among all networks, while DNA-Net and MRF3Net achieve relatively poor results. The possible reason is that DNA-Net has a larger width, and MRF3Net network has a shallower depth. Although they have achieved good detection results in their respective tasks, the depth and width of the network need to be properly balanced to adapt to different task characteristics. After a series of optimized calculation strategies, MACAN achieves the best performance in terms of calculation time. However, the number of parameters exceeds that of MRF3Net, DNA-Net and U-KAN network. However, compared with TransUNet and MTU-Net networks which also integrate ViT modules, MACAN has fewer parameters due to the efficiency improvement of the traditional ViT module in the IKA module.

[0211] Table 3: Comparison results of MACAN with other 6 advanced methods on LLAS dataset

[0212]

[0213] In Figure 9In the middle, 6 pictures with typical characteristics are selected, and the verification results of different networks are qualitatively compared. In the picture of sequence (1), there are typical irregular shapes, especially the right side of the glue point, which is more random; In the picture of sequence (2), there are a large number of solder flux under the chip, which makes it difficult to distinguish the outline of the glue point from the background; The picture in sequence (3) is also shown in sequence (4) of ablation experiment, which has the characteristics of small size target in complex background, and can be used to verify the effect of different networks in identifying small targets. The chips shown in sequences (4) and (5) are another common type of chips in the LLAS data set. This type of chip is pasted on a black substrate, which is extremely close in color to the glue point, greatly increasing the difficulty of image recognition. Sequence (6) is a relatively ideal image recognition picture, and the target is relatively clear in the picture. Thanks to the advantages of KAN network in MACAN in dealing with complex nonlinear problems, and the strong ability of IKA to inherit the ViT module in capturing context information and long distance dependence, the model can clearly identify the boundary profile of the target. Through the integration of multi-receptive field perception mechanism, MACAN can accurately extract the target at different scales. In addition, the multi-layer fusion module further enhances the ability of MACAN to accurately extract the target from complex background. With these advantages, MACAN has achieved excellent recognition results in all images. DNA-Net has relatively poor accuracy in target contour recognition, and the possible reason is that the LLAS data set has limited sample diversity, making it difficult to fully train the DNA-Net network with dense nested design, and its advantages cannot be fully utilized. MRF3Net network also performs poorly in these typical image recognition tasks. Its relatively shallow network depth makes it particularly difficult to accurately segment the target in sequences (4) and (5) with complex background, but its multi-receptive field perception mechanism enables it to successfully identify the target in sequence (3) with small target perception task. The image segmentation results of TransUNet, HCF-Net and U-KAN network are relatively close, but U-KAN has higher segmentation accuracy when dealing with images like sequences (4) and (5) where the target and background have extremely close gray values. The MTU-Net network based on ViT multi-level feature extraction module achieves segmentation results second only to MACAN.

[0214] In Figure 10 The ROC curve is used to provide another intuitive expression of the classification performance of each network. The larger the area enclosed by the curve and the X-axis, the larger the value of the area under the curve (AUC), and the better the performance of the network. As can be seen from the figure, the curve of the proposed network is closest to the upper left corner, and the value of AUC is the largest, indicating that MACAN can meet the highest true positive rate at the lowest false positive rate.

[0215] In order to solve the challenge of detecting the defects of the laser collimator lens glue point in the complex background, especially in the presence of random shape and complex background, the MACAN based on deep learning is proposed. MACAN introduces a multi-pole attention mechanism and an adaptive selection of multi-layer fusion module to extract defects with random shape and multi-scale. And through the KAN, multi-receptive field perception mechanism and other modules, the traditional attention mechanism is optimized to further suppress the interference brought by the complex background. Based on the ablation experiment and extensive comparison experiment of the self-built data set LLAS, the superiority and robustness of MACAN are proved. However, the current research still has certain limitations, for example, since the data set is collected from the glue point image in the formal production after the point glue equipment is fully debugged, most of the glue points in the data set are ideal and easy to identify, so the pictures containing obvious defect features are relatively few, and the sample diversity is insufficient. When the point glue equipment is not ideal in debugging effect, more complex features appear, and the probability of artificial recheck will increase. At the same time, although MACAN network has achieved excellent segmentation accuracy, compared with other lighter networks, it is not superior in operation time and parameter amount. Therefore, in future research, the types and quantities of data sets will be further expanded, and the training effect of the model will be further improved. At the same time, the lightweight research will improve the advantages of the model in the calculation time and the parameter amount without reducing the accuracy, which is more conducive to the deployment of the model.

[0216] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Those skilled in the art can make various modifications and changes to the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A multi-scale blob detection method, characterized in that, The method comprises the following steps: Step 1: using an intelligent laser lens dispensing platform to collect images of the laser chip before bonding; and establishing an LLAS dataset based on the collected images; establishing a multi-level attention-guided context aggregation network model; Step 2: applying the multi-level attention-guided context aggregation network model to process the images in the LLAS dataset to segment out the glue point target; Step 3: obtaining a glue point detection result based on the segmented glue point target; The multi-level attention-guided context aggregation network model adopts a U-Net model as a backbone network, and fuses an interactive self-attention-guided Kolmogorov-Arnold representation module, a multi-perception spatial attention module, a KAN-enhanced channel attention module and a multi-dimensional feature fusion module based on an attention mechanism based on Kolmogorov-Arnold representation; The multi-level attention-guided context aggregation network model comprises sequentially connected down-sampling layers, a fusion layer and up-sampling layers, and the fusion layer comprises an interactive self-attention-guided Kolmogorov-Arnold representation module, a multi-perception spatial attention module, a KAN-enhanced channel attention module and a multi-dimensional feature fusion module; The KAN-enhanced channel attention module is applied to the 4th layer in the down-sampling process, the 5th layer in the down-sampling process and the 4th layer in the up-sampling process; The multi-dimensional feature fusion module is applied to the skip connection layers of the 4th layer and the 5th layer; The specific process of applying the multi-level attention-guided context aggregation network model to process the images in the LLAS dataset is as follows: The input feature image T is in (B, C, H, W), and the remodeled feature image T1, the extreme value feature image T2 and the mean value feature image T3 are obtained through the three down-sampling layers arranged in succession; The extreme value feature image T2 and the mean value feature image T3 are input and transmitted to a 3*3 deep convolution block to obtain image information T2' and image information T3', respectively; The number of channels in the U-Net model is halved; The image information T2' is transmitted to the interactive self-attention-guided Kolmogorov-Arnold representation module, and the image information T3' is transmitted to the multi-perception spatial attention module; the image information T2' and the image information T3' are processed by the interactive self-attention-guided Kolmogorov-Arnold representation module and the multi-perception spatial attention module, respectively, to obtain high-resolution feature information; The up-sampling layer receives the high-resolution feature information processed by the interactive self-attention-guided Kolmogorov-Arnold representation module and the multi-perception spatial attention module, and obtains the segmented glue point target based on the high-resolution feature information.

2. The multi-scale blob detection method of claim 1, wherein, The interactive self-attention-guided Kolmogorov-Arnold representation module comprises an additive interactive self-attention module and a Kolmogorov-Arnold network, and the Kolmogorov-Arnold network is provided with two layers; The operation process of the additive interactive self-attention module is as follows: 1) The input feature image T e (B, C, H, W) is expanded in the channel dimension by a channel expansion convolution operation to obtain T' e (B, 3C, H, W); 2) Q, K and V are obtained by a channel blocking operation, wherein Q e (B, C, H, W), K e (B, C, H, W), V e (B, C, H, W), Q is Query, K is Key, and V is Value; 3) The similarity function is defined as the sum of the context scores of Q and K: Sim(Q, K) = F(Q) + F(K), s.t. F(Q) = S(C(Q)); Wherein: F(·) is a context mapping function, and it contains necessary information interaction; F(·) is specifically spatial attention S(·) e (B, C, H, W) and channel attention C(·) e (B, C, H, W); 4) Based on the similarity function defined as the sum of the context scores of Q and K, the final output feature Attn(Q, K, V) is obtained; After the output feature Attn(Q, K, V) is processed by a depth separable convolution, the original input image T e (B, C, H, W) is added as a residual to obtain the final output feature T" e (B, C, H, W) of the additive interaction self-attention module; The operation process of the Kolmogorov-Arnold network is as follows: The final output feature T" of the additive cross-attention module is reconstructed into a two-dimensional patch feature T" r ∈(B*H*W,C); Projecting the two-dimensional patch feature T" r ∈(B*H*W,C) to a lower dimension Post-embedding back to the original dimension C, reshaped to the original tensor dimension; The final output feature T" of the additive interaction self-attention module is added as a residual to obtain the final IKA output image T IKA ∈(B,C,H,W).

3. The multi-scale blob detection method of claim 2, wherein, The operation process of the multi-perception spatial attention module is as follows: (i) The input feature image T e (B, C, H, W) is sequentially subjected to feature extraction by 4 convolution kernels with sizes of 1x1, 3x3, 5x5 and 7x7 to obtain image features; (ii) The image features extracted in step (i) are subjected to information compression by average pooling, and then the image is restored to the original size by nearest neighbor interpolation to obtain a reshaped feature map T1 e (B, 1, H*W); (iii) The extreme value feature information and the mean value feature information of the input feature image T e (B, C, H, W) are obtained by global maximum pooling and global average pooling operations on the second branch and the third branch of the multi-perception spatial attention module respectively, and the extreme value feature information and the mean value feature information are respectively reconstructed into an extreme value feature map T2 e (B, H*W, 1) and a mean value feature map T3 e (B, 1, H*W); (iv) The extreme value feature map T2 and the mean value feature map T3 are subjected to matrix multiplication to obtain a feature map T4 e (B, H*W, H*W); The reshaped feature map T1 and the feature map T4 are subjected to matrix multiplication, and the result is subjected to reconstruction and Sigmoid activation operation, and then the input feature image T e (B, C, H, W) is scored to obtain a spatial attention feature map T5 e (B, C, H, W).

4. The multi-scale blob detection method of claim 3, wherein, The operation process of the KAN enhanced channel attention module is as follows: (I) The input feature image T e (B, C, H, W) is subjected to feature compression by global maximum pooling and global average pooling respectively; (II) The compressed image features are reconstructed to obtain a reconstructed feature map T1' e (B, 1, C) and a reconstructed feature map T2' e (B, 1, C) respectively; The reconstructed feature map T1' belongs to (B, 1, C), the reconstructed feature map T2' belongs to (B, 1, C), and the original input feature image T belongs to (B, C, H, W) are scored to obtain a final channel attention output feature T3' belonging to (B, C, H, W) after being fused by a one-dimensional convolution kernel with a size of 3, activated by a Sigmoid function and reconstructed.

5. The multi-scale blob detection method of claim 4, wherein, The multi-dimensional feature fusion module comprises a low layer, a current layer and a high layer, and the low layer, the current layer and the high layer are connected in series. Let the low-level input feature be The current layer input feature is T e (B, C, H, W), and the high-level input feature is The operation process of the multi-dimensional feature fusion module is as follows: aligning the low-level input features with the current layer input features T∈(B,C,H,W) and the high-level input features with the current layer input features T∈(B,C,H,W) to ensure the same feature size; Low-level input features Down-sampling is first performed with detail preservation, obtaining low-level features T' low ∈(B,C,H,W); After the current layer input feature T belongs to (B, C, H, W) is refined and extracted by a 1x1 convolution layer, the current layer feature T' belongs to (B, C, H, W) is obtained; High-level input features Up-sampling is performed using linear interpolation to obtain high-level features T' high ∈(B,C,H,W); low-level features T' low , current-level features T', and high-level features T' high After being respectively processed by a 1x1 convolution operation and a Sigmoid function activation, spatial weight matrices Wlow, Wcur, and Whigh from different levels are generated, and each element in the spatial weight matrices Wlow, Wcur, and Whigh represents the possibility that its corresponding point belongs to the target. The spatial weight matrix Wlow, Wcur and Whigh are spliced, and the final spatial weight matrix W is generated after being operated by a 1x1 convolution operation and activated by a Sigmoid function; Low-level features T' low , current layer features T', and high-level features T' high are fused by the intelligent channel integration module to obtain ICI fusion features T ICI ∈(B, C, H, W); The T1 fused with the spatial weight matrix W is multiplied element by element to realize joint weighted fusion in the spatial and channel dimensions, and obtain the final multi-dimensional feature fusion module output result T MFF ∈(B,C,H,W).

6. The multi-scale blob detection method of claim 4, wherein, The multi-dimensional feature fusion module comprises a detail reservation down-sampling module and an intelligent channel integration module. The operation process of the detail reservation down-sampling module is as follows: Input low-level input features The input image is cut into four parts symmetrically along its height and width. The four features after cutting compression are spliced, and after maximum pooling and average pooling, a 1x1 convolution kernel is used for fine feature extraction, and finally the low-level features T after downsampling are output l low∈(B,C,H,W); The operation process of the intelligent channel integration module is as follows: T′ low , the current layer feature T′ and the high layer feature T′ high , using the channel split technology, the feature map is split into 4 equal parts in the channel, respectively obtaining low layer segmentation features current layer segmentation features and high layer segmentation features wherein, T′ low_i is the i-th divided part of the low layer feature, T′ i is the i-th divided part of the current layer feature, T′ high_i is the i-th divided part of the high layer feature, i=(1, 2, 3, 4). Low-level segmentation features Current layer segmentation features And high-level segmentation features The calculation is performed according to the following formula, so that the intelligent channel integration module intelligently selects which layer's channel to give more weight according to the characteristics of the target: p i = Sigmoid(T' i ); T ICI = Concat(T' g_1 , T' g_2 , T' g_3 , T' g_4 ); Wherein, p i is the feature T' i After the activation function Sigmoid, the feature weight T' g_i represents the intelligent aggregation result of each partition; The four equal parts after segmentation are spliced in the channel dimension to obtain T ICI ∈R B×C×H×W ; When p i > 0.5, the multi-level attention-guided context aggregation network model enhances the weight of the fine-grained features delivered by the low-level network; When p i The multi-level attention guided context aggregation network model enhances the weight of the rich semantic global features delivered by the high-level network when p

Images