Industrial tiny object defect identification and detection method based on improved FPN model

By improving the FPN model, combining the hybrid convolution module and the bottleneck structure module, the network structure is optimized, and the problems of low accuracy and slow detection of industrial micro objects are solved, achieving high-precision and high-speed real-time detection effect.

CN119067913APending Publication Date: 2024-12-03UNIV OF ELECTRONICS SCI & TECH OF CHINA
0 Cites 0 Cited by

Patent Information

Application Number
CN202410936683.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-12
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The prior art has low accuracy and slow speed when detecting industrial micro objects in real time, especially since tiny objects are easily missed or blocked, resulting in poor detection accuracy.

Method used

Using the improved FPN model, by constructing a hybrid convolution module and a bottleneck structure module, combining the classic FPN model, backbone network and neck network, the network structure is optimized to maximize the characteristic information and gradient information of tiny objects and improve detection accuracy and speed.

Benefits of technology

It realizes high-precision detection of industrial tiny objects in real time, improves detection accuracy and speed, enhances the practicality of the model, and has application value.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure BDA0004941534370000021
    Figure BDA0004941534370000021
  • Figure BDA0004941534370000022
    Figure BDA0004941534370000022
  • Figure BDA0004941534370000031
    Figure BDA0004941534370000031
Patent Text Reader

Abstract

The invention discloses an industrial tiny object defect identification and detection method based on an improved FPN model. The method comprises the following steps: constructing a hybrid convolution module and a bottleneck structure module; constructing a backbone network basic structure and a neck network basic structure according to the hybrid convolution module and the bottleneck structure module; based on the classic FPN model, the basic structure of the backbone network and the basic structure of the neck network, constructing an industrial tiny object detection model; training and optimizing the industrial tiny object detection model by adopting the image data set; and inputting a to-be-detected industrial tiny object image into the optimized industrial tiny object detection model, and carrying out defect identification detection. According to the method, the feature information of one tiny object is maximized by fusing the upper feature layer of the neck, the feature information of the tiny object is detected and stored in real time as far as possible in a mode of minimizing the bottleneck and maximizing gradient information, and the problems of lack of industrial tiny object detection precision and low speed under real-time detection can be effectively solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of object detection in machine vision, and in particular relates to an industrial tiny object defect recognition and detection method based on an improved FPN model. Background Art

[0002] At present, the use of deep learning technology for target detection has been seen everywhere in daily life. However, in practical applications, due to the small size of tiny objects, the actual accuracy of their detection is low. In the image processing process, as the size and resolution of the image increase, the background becomes more complex and the object area becomes smaller. In addition, due to feature extraction problems, small objects are easily missed, overlapped and occluded, and such phenomena will also affect the detection accuracy. Therefore, in order to improve the accuracy of detecting tiny objects, the acceptance range of the network must be expanded. However, as the acceptance range of the network expands, the detection speed of objects will decrease. Therefore, how to detect industrial tiny objects in real time has become an urgent problem to be solved. Summary of the invention

[0003] The purpose of the present invention is to address the above-mentioned deficiencies in the prior art and provide an industrial tiny object defect recognition and detection method based on an improved FPN model to solve the problem that industrial tiny objects are difficult to detect in real time and have poor detection accuracy.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is:

[0005] A method for industrial tiny object defect recognition and detection based on an improved FPN model comprises the following steps:

[0006] S1, build hybrid convolution module and bottleneck structure module;

[0007] S2. Construct the basic structure of the backbone network and the basic structure of the neck network according to the hybrid convolution module and the bottleneck structure module;

[0008] S3. Build an industrial tiny object detection model based on the classic FPN model, the basic structure of the backbone network and the basic structure of the neck network;

[0009] S4, using image datasets to train and optimize industrial tiny object detection models;

[0010] S5. Input the image of the industrial tiny object to be detected into the optimized industrial tiny object detection model to perform defect recognition detection.

[0011] Furthermore, in step S1, the hybrid convolution module and the bottleneck structure module both include a Conv-BN-Silu structure; the Conv-BN-Silu structure converts the input image into a feature map.

[0012] Furthermore, the Conv-BN-Silu structure includes a convolutional layer, a BN structure, and a SLIU structure;

[0013] The convolution layer performs convolution operations on the input image and outputs a set of feature maps; a set of feature maps enters the BN structure for batch normalization processing and outputs a set of normalized feature maps; a set of normalized feature maps is input into the SLIU structure for nonlinear transformation and outputs a set of feature maps after nonlinear transformation.

[0014] Furthermore, the convolution layer performs convolution operation on the input image, specifically:

[0015]

[0016] Among them, a d,i,j is the i-th row and j-th column element of the feature map when the image depth is d; w d,m,n is the weight of the convolution kernel in the mth row and nth column when the image depth is d; w b is the bias term of the convolution kernel; f is the activation function; F is the output feature map of the entire convolution layer; D is the maximum depth of the image.

[0017] Furthermore, the BN structure performs batch normalization on the input feature map, which is specifically:

[0018]

[0019]

[0020] Among them, the input image data set B = {x1, x2, ..., x m}, x i is the i-th input image data in set B, μ B is the average value; is the variance; ε is a small value.

[0021] Furthermore, the SLIU structure performs nonlinear transformation on the feature map, which is specifically:

[0022] f(x)=x*σ(x)

[0023]

[0024] Among them, x is the input of the activation function, that is, the input image data value; σ(x) is the standard activation function formula; e is a natural constant.

[0025] Furthermore, the right branch of the basic structure of the backbone network in step S2 includes a hybrid convolution module, which includes two convolution layers with a 3×3 kernel, and the convolution layers generate different feature maps and gradient information; the hybrid convolution module is divided into two parts, one part goes directly downward and is connected to the bidirectional SILU convolution block through a shortcut connection part; the other part is connected to the shortcut connection part through a bottleneck structure module.

[0026] Furthermore, the industrial tiny object detection model in step S3 includes 4 backbone network basic structures and 6 neck network basic structures; among them, the four backbone network basic structures are located in the backbone, the two neck network basic structures are located in the neck, and the four neck network basic structures are located in the head.

[0027] Furthermore, the number of bottleneck layers Ci used in the basic structure of the backbone network located in the backbone of the industrial tiny object detection model is defined as:

[0028]

[0029] Among them, i is the position of the basic structure of the backbone network in the backbone network.

[0030] Furthermore, the head of the industrial tiny object detection model includes a prediction module f1, a prediction module f2 and a prediction module f3; during learning, the training ratios of the prediction module f1, the prediction module f2 and the prediction module f3 channels are set to 0.4, 0.3 and 0.3 respectively.

[0031] The industrial tiny object defect recognition and detection method based on the improved FPN model provided by the present invention has the following beneficial effects:

[0032] 1. The present invention maximizes the feature information of a tiny object by fusing the upper feature layer of the neck, minimizes bottlenecks, maximizes gradient information, and detects and saves the feature information of the tiny object as real-time as possible, which can effectively solve the problems of lack of detection accuracy and slow speed of industrial tiny objects under real-time detection.

[0033] 2. The present invention improves the classic FPN model, which can increase the speed of the trunk and neck, learn more diverse gradient information, and a model with a feature fusion layer to obtain low-level, high-resolution feature information.

[0034] 3. The present invention takes into account the lack of high-precision real-time detection models in the field of industrial tiny object defect recognition and detection, and designs a detection model for this specific field, which improves its accuracy and speed, improves its practicality, and has application value.

[0035] 4. By increasing and optimizing the network structure, the present invention enables the model to learn more diverse gradient information, especially the gradient information of tiny object features can be retained and amplified, which is conducive to the accurate detection of tiny objects. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is the network structure diagram of the classic FPN model.

[0037] Figure 2 This is the mixed convolution module of the present invention.

[0038] Figure 3 This is the bottleneck structure module of the present invention.

[0039] Figure 4 This is the Conv-BN-Silu structure of the present invention.

[0040] Figure 5 This is the SOD-C module of the present invention.

[0041] Figure 6 This is the SOP-C module of the present invention.

[0042] Figure 7 This is the network structure diagram of the industrial tiny object detection model of the present invention. DETAILED DESCRIPTION

[0043] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0044] Example 1

[0045] This embodiment provides an industrial tiny object defect recognition and detection method based on an improved FPN model, which is used to solve the problem of low accuracy and slow rate of industrial tiny object detection in the prior art, and specifically includes the following steps:

[0046] Step S1, constructing a hybrid convolution module and a bottleneck structure module;

[0047] This embodiment is based on the classic FPN model and improves it. Figure 1, the classic FPN model includes a backbone network C, a bottleneck structure P, and a head structure F; wherein the backbone network C is used to remove noise and extract feature information from the input image. The neck structure P combines the collected feature information to create a feature map with richer information. Subsequently, the head structure F detects objects based on the created feature map. The number of layers of the backbone network C and the bottleneck structure P can be different.

[0048] The target detection framework of this embodiment enhances the functional information of the image by adding a top-down path based on FPN, so that the entire network can obtain better detection results. The neck structure feature layers P1, P2 and P3 are generated using the classic FPN model, and high-resolution feature maps are obtained through lateral connections in the middle of the feature layers.

[0049] In order to detect industrial tiny objects, it is first necessary to use a larger image in the head network to detect the presence of tiny objects and create a feature map containing various feature information. However, since the target needs to be detected in real time, the model needs to achieve rapid detection of large-area images in this part, specifically:

[0050] Since the features of tiny objects are easily diluted when passing through this layer of the network during image processing, the present invention is designed to use a hybrid convolution module and a bottleneck structure module to preserve the features of tiny objects.

[0051] Specifically, refer to Figure 2 and Figure 3 , hybrid convolution module and bottleneck structure module. Both modules use the Conv-BN-Silu structure to convert the input image into a feature map through the convolution kernel in the convolution layer.

[0052] refer to Figure 4 , Conv-BN-Silu includes convolutional layer, BN structure and SLIU structure;

[0053] The convolution layer performs convolution operations on the input image and outputs a set of feature maps; a set of feature maps enters the BN structure for batch normalization processing and outputs a set of normalized feature maps; a set of normalized feature maps is input into the SLIU structure for nonlinear transformation and outputs a set of feature maps after nonlinear transformation.

[0054] In the convolution layer, the input is the original image, and the image is convolved through the convolution layer. The convolution layer extracts local features of the image through the convolution kernel. These features are the manifestation of edges, textures, and other significant features in the image. The output of the convolution operation is a set of feature maps, which are the internal representation of the image and contain information for further processing.

[0055] The feature map enters the Batch Normalization (BN) structure for batch normalization. The purpose of batch normalization is to normalize the feature map to solve the gradient vanishing and gradient exploding problems in the deep network training process. It adjusts the distribution of the feature map so that the input of each layer has a stable distribution, which helps the network converge and train faster. The output of batch normalization is a set of normalized feature maps.

[0056] The normalized feature map enters the SiLU structure for nonlinear transformation. SiLU is a new type of activation function proposed in recent years. It combines the advantages of linear units and gating mechanisms, which can accelerate the training of neural networks and improve the performance of the model. The output of the SiLU activation function is a set of feature maps that have undergone nonlinear transformation. These feature maps will be used in subsequent network layers to further extract and integrate information.

[0057] First, a convolution operation is performed between the input image and the kernel;

[0058] In the convolution layer, a convolution operation is performed between the input image and the convolution kernel, which is specifically:

[0059]

[0060] Among them, x ij is the i-th row and j-th column of the image, w mn is the weight of the convolution kernel in the mth row and nth column, w b is the bias term of the convolution kernel, a ij is the i-th row and j-th column element of the feature map, and f is the activation function.

[0061] Assuming that the image depth before convolution is D, the depth of the corresponding convolution kernel is also D, and the formula is:

[0062]

[0063] Secondly, the input image data is normalized through batch normalization to increase the learning speed, which specifically includes the following:

[0064] Suppose its input is u, the weight parameter and bias parameter are W and b respectively, the activation function is σ, and the batch normalization operator is BN, then the output of batch normalization is σ(BN(Wu+b));

[0065]

[0066] In this embodiment, a small batch of m input data sets B = {x1, x2, ..., x m Find the average value μ B and variance The input data is then regularized with a mean of 0 and a variance of 1; where ε is a small value to prevent division by 0.

[0067] Finally, the Silu structure is used as the activation function to scale the input image data value between 0 and 1. The formula is:

[0068] f(x)=x*σ(x)

[0069] in,

[0070] The Conv-BN-Silu structure of this embodiment improves the stability and accuracy of the model.

[0071] The hybrid convolution module structure uses two Conv-BN-Silus structures with a convolution kernel size of 3 and a stride of 1 to generate many spatial features based on the feature map and learn various features.

[0072] The bottleneck structure module adds a shortcut connection to solve the problem of gradient vanishing of tiny features in the model. This is mainly to prevent the features of tiny industrial objects from gradually disappearing as the model becomes larger, prevent input loss, and improve the learning ability of the model in the middle layer.

[0073] Step S2: constructing the basic structure of the backbone network and the basic structure of the neck network according to the hybrid convolution module and the bottleneck structure module;

[0074] The basic structure of the backbone network, hereinafter abbreviated as SOD-C; the basic structure of the neck network, hereinafter abbreviated as SOD-P;

[0075] refer to Figure 5 , SOD-C is mainly located in the backbone, which matches the number of channels from the output red convolution layer to the final convolution layer, and retains small features through the left branch structure. In SOD-C, a hybrid convolution module is used at the right branch, which uses two convolution layers containing 3×3 kernels. This convolution layer is mainly used to reduce redundancy, generate different feature maps and gradient information, and the hybrid convolution layer is divided into two parts, one part goes directly down, and the other part diversifies the spatial information and captures the missing small features through the bottleneck structure module.

[0076] refer to Figure 6, SOD-P is mainly located in the neck, it does not use the left branch structure similar to the backbone network because it already has many abstract features, and uses hybrid convolution and bottleneck extraction feature maps to preserve spatial information. It then fuses the resulting feature maps with the output generated in the lower stage. The architecture effectively extracts different features while retaining object information outside of tiny objects and receiving gradient information in various ways in the overall model.

[0077] Step S3: Based on the classic FPN model, the basic structure of the backbone network and the basic structure of the neck network, an industrial tiny object detection model is constructed, which specifically includes the following contents:

[0078] refer to Figure 7 , the industrial micro-object detection model of this embodiment includes 4 SOD-Cs and 6 SOD-Ps, and 4 SOD-Cs are used in the trunk, including SOD-C1 module, SOD-C2 module, SOD-C3 module, and SOD-C4 module; the trunk is used to extract deep features of the image. As the module goes deeper, the depth and abstraction of the features will increase, and the spatial resolution will decrease. The neck includes SOD-P1 module and SOD-P2 module. The neck is used for further processing and adjustment of features, which includes upsampling operations to restore spatial resolution, and feature fusion and adjustment; the head includes SOD-P3 module, SOD-P4 module, SOD-P5 module, SOD-P6 module, prediction module f1, prediction module f2 and prediction module f3. The head is the place in the network used for the final task, that is, defect detection, and is responsible for generating the final detection results.

[0079] The basic structure of the network consists of a top-down path and a lateral connection. In the top-down path, feature maps of higher layers (smaller size, greater depth) are upsampled to increase spatial resolution. The lateral connection fuses the feature maps of the same layer (from the original backbone network) with the upsampled feature maps. This fusion can be achieved by element-wise addition, thereby retaining rich detail information.

[0080] Take an image of an industrial tiny object defect as the input of the industrial tiny object detection model to illustrate:

[0081] The input image first enters the backbone network. In the SOD-C1 module, the image passes through a series of convolutional layers to extract preliminary features. These features are then passed to the SOD-C2 module, which includes more convolutional layers and a pooling layer to reduce the spatial resolution of the feature map. The same process is repeated in the SOD-C3 and SOD-C4 modules, each of which extracts deeper and more abstract features.

[0082] The features then flow into the neck network. In the SOD-P1 module, upsampling layers are included to increase the spatial resolution of the feature map, and there are some shortcut connections or skip connections to combine the features of the early layers with the features of the late layers, so that more detailed information can be retained. The SOD-P2 module further processes these fused features.

[0083] Next, the features enter the head network. The features are further processed in the SOD-P3 to SOD-P6 modules, which include more convolutional layers and upsampling layers to ensure that the feature maps have enough resolution to detect tiny defects. The outputs of these modules are finally fed into the prediction modules f1, f2, and f3.

[0084] Prediction modules f1, f2, and f3 generate the final output according to the needs of the task.

[0085] Throughout the process, the features of the image flow through various parts of the network, and each module adds some specific processing steps based on the previous module. Finally, the model outputs detection results, which can be used for further analysis or decision-making.

[0086] Because as the number of layers increases, the complexity of the model also increases, and deeper layers can extract more complex features. However, as the number of layers increases, the training time and memory usage of the model increase, and the information of small object features is more likely to be lost. In addition, due to the large overhead of more layers, the increased complexity of the model will affect the real-time performance of detection.

[0087] In order to send a larger image from the neck to the head, the present invention performs three upsampling to learn more features; in this process, as the continuous downsampling layers proceed, in order to reduce noise and keep small objects, the feature map is extracted from the downsampled convolutional backbone network, which is downsampled twice and then passed through the proposed overall model to reduce noise and increase gradient information, and then connected to the additional feature layer.

[0088] In this embodiment, the input and output method of the convolutional layer is modified so that the input and output channels of the gradient information maximization part are the same.

[0089] In addition, due to the large amount of data tensors calculated, the deep convolution layer and the 1×1 convolution bottleneck kernel can reduce the number of flow points; however, the additional 1×1 convolution layer is harmful to the GPU operation efficiency. Especially in the backbone, if the left branch connection method is reused in the module, too many input and output channels must be changed separately. This embodiment significantly reduces the number of changes in the channel by dividing the channel into modules and connections, so that the channel is minimized.

[0090] exist Figure 4The number of bottleneck layers Ci used in the SOD-Ci module in the backbone is defined as

[0091]

[0092] Here, i represents the position of SOD-C in the backbone network. The reason for adjusting the number of basic models of the bottleneck structure is that the more times the feature map passes through the convolution, the more gradient information is generated. However, small features may disappear. Therefore, a small number of bottleneck structures are used initially to preserve as many features as possible. Subsequently, the number is increased to highlight the features and extract various types of gradient information.

[0093] The first to third backbone modules in the overall model use SOD-C to increase the cardinality, and in the last layer, they are designed to be sensitive to tiny features using the 1×1 convolution layer of the existing SOD-C. In the bottleneck structure, there is no need to keep tiny features, so the present invention uses SOD-P to reduce computational cost and increase speed. In addition, in the red marked part, in order to reduce noise and increase gradient information, the present invention uses two convolution layers instead of one, and designs it to increase gradient information through the model. The generated feature map contains low-level, high-resolution functional information, which is brought to the neck and combined to design an additional prediction head that is more sensitive to industrial tiny objects. The prediction module f1 reduces the input SOD-P N by half, leaving the original image without severely diluted information. In addition, because it retains the upsampled data, it contains rich gradient information, which is a huge advantage for detecting small industrial objects.

[0094] In addition, the number of N in SOD-P is 3 or 4 before the output of prediction module f2 and prediction module f3. This enables learning to use richer spatial information. Only prediction module f1 and prediction module f2 with many small object features, and prediction module f3 with many spatial features are used to reduce computational costs and optimize real-time detection. Using a four-layer structure in FPN can increase the number of parameters and provide more learning opportunities for smaller objects. The number of parameters used affects real-time learning; therefore, when building the model, the number of heads is limited to 3, the learning ratio of f1 is increased, and the learning weights of f2 and f3 are reduced. When learning occurs, the training ratio of each channel is specified as 0.4, 0.3, and 0.3.

[0095] Step S4: using the image data set to train and optimize the industrial tiny object detection model, which specifically includes:

[0096] This example uses specific hyperparameter settings for training. On the PCB image dataset, the size of each training batch is 10, while on the cell image dataset it is 16. All models are trained for more than 500 cycles, and an early stopping strategy is used to prevent overfitting. The performance changes on the validation set are monitored to determine when to stop training. The performance evaluation of the model mainly uses the mean average precision (mAP) as a metric.

[0097] Specifically, the PCB image resolution of this embodiment is 4096×3000×24, which is used to detect defects smaller than 32×32 in real time. There are 5 types of defects in the PCB image data. Among them, the partial structural defects detected in this embodiment, as well as the size data of the smallest object range from 4×4 to 32×32 pixels, and are located in an area that is usually hidden, some of which even exist in fine wiring parts. This is a kind of data that is difficult to detect because it seems to greatly increase the difficulty of detecting defects. In addition, due to the complex background and high background interference, other features are highlighted, making detection more difficult. Regarding the amount of PCB data, 2070 objects were marked from 1513 images. When the PCB image is resized, tiny features may disappear; therefore, it is divided into 8 parts, the size is adjusted to 1024×1024 pixels, and then the objects are detected.

[0098] This embodiment uses 1400 images from 1513 PCB images for training, 34 images for testing, and 79 images for verification, and processes the PCB images of size 4096×3000×24 through preprocessing to use them as 1024×1024×24 data. The PCB image model uses a batch size of 10 due to its larger image size, while the unit data model uses a batch size of 16. In addition, each learning rate initially starts from 0.01 and is then multiplied by 0.01 to ensure function convergence.

[0099] Step S5: input the image of the industrial tiny object to be detected into the optimized industrial tiny object detection model to perform defect recognition detection.

[0100] Although the specific implementation of the invention is described in detail in conjunction with the drawings, it should not be understood as limiting the scope of protection of this patent. Within the scope described in the claims, various modifications and variations that can be made by those skilled in the art without creative work still fall within the scope of protection of this patent.

Claims

1. A method for industrial tiny object defect recognition and detection based on an improved FPN model, characterized in that: The following steps are involved: S1, build hybrid convolution module and bottleneck structure module; S2. Construct the basic structure of the backbone network and the basic structure of the neck network according to the hybrid convolution module and the bottleneck structure module; S3. Build an industrial tiny object detection model based on the classic FPN model, the basic structure of the backbone network and the basic structure of the neck network; S4, using image datasets to train and optimize industrial tiny object detection models; S5. Input the image of the industrial tiny object to be detected into the optimized industrial tiny object detection model to perform defect recognition detection.

2. The industrial tiny object defect recognition and detection method based on the improved FPN model according to claim 1 is characterized in that: In the step S1, the hybrid convolution module and the bottleneck structure module both include a Conv-BN-Silu structure; the Conv-BN-Silu structure converts the input image into a feature map.

3. The industrial tiny object defect recognition and detection method based on the improved FPN model according to claim 2 is characterized in that: The Conv-BN-Silu structure includes a convolutional layer, a BN structure and a SLIU structure; The convolution layer performs convolution operation on the input image and outputs a set of feature maps; a set of feature maps enters the BN structure for batch normalization processing, and outputs a set of normalized feature maps; a set of normalized feature maps is input into the SLIU structure for nonlinear transformation, and outputs a set of feature maps after nonlinear transformation.

4. The industrial tiny object defect recognition and detection method based on the improved FPN model according to claim 3 is characterized in that: The convolution layer performs a convolution operation on the input image, specifically: Among them, a d,i,j is the i-th row and j-th column element of the feature map when the image depth is d; w d,m,n is the weight of the convolution kernel in the mth row and nth column when the image depth is d; w b is the bias term of the convolution kernel; f is the activation function; F is the output feature map of the entire convolution layer; D is the maximum depth of the image.

5. The industrial tiny object defect recognition and detection method based on the improved FPN model according to claim 4 is characterized in that: The BN structure performs batch normalization processing on the input feature map, which is specifically as follows: Among them, the input image data set B = *x1,x2,...,x m +, x i is the i-th input image data in set B, μ B is the average value; is the variance; ε is a small value.

6. The industrial tiny object defect recognition and detection method based on the improved FPN model according to claim 5 is characterized in that: The SLIU structure performs nonlinear transformation on the feature map, which is specifically: f(x)=x*σ(x) Among them, x is the input of the activation function, that is, the input image data value; σ(x) is the standard activation function formula; e is a natural constant.

7. The industrial tiny object defect recognition and detection method based on the improved FPN model according to claim 1 is characterized in that: The right branch of the basic structure of the backbone network in step S2 includes a hybrid convolution module, which includes two convolution layers with a 3×3 kernel, and the convolution layers generate different feature maps and gradient information; the hybrid convolution module is divided into two parts, one part goes directly downward and is connected to the bidirectional SILU convolution block through a shortcut connection part; the other part is connected to the shortcut connection part through a bottleneck structure module.

8. The industrial tiny object defect recognition and detection method based on the improved FPN model according to claim 1 is characterized in that: The industrial tiny object detection model in step S3 includes 4 backbone network basic structures and 6 neck network basic structures; wherein, the four backbone network basic structures are located in the backbone, the two neck network basic structures are located in the neck, and the four neck network basic structures are located in the head.

9. The method for industrial tiny object defect recognition and detection based on the improved FPN model according to claim 8, characterized in that: The number of bottleneck layers Ci used in the basic structure of the backbone network located in the backbone of the industrial small object detection model is defined as: Among them, i is the position of the basic structure of the backbone network in the backbone network.

10. The industrial tiny object defect recognition and detection method based on the improved FPN model according to claim 8, characterized in that: The head of the industrial tiny object detection model includes a prediction module f1, a prediction module f2 and a prediction module f3; during learning, the training ratios of the prediction module f1, the prediction module f2 and the prediction module f3 channels are set to 0.4, 0.3 and 0.3 respectively.