A method and system for small target detection based on lattice-like superpixels

Abstract

This invention proposes a small target detection method and system based on lattice-like superpixels, comprising: constructing an improved YOLOv5 detection model for small target detection, wherein the improved YOLOv5 detection model includes a backbone feature network, a neck network, and a detection head; the backbone feature network includes a lattice-like superpixel generation module, several CBL modules, and a CSP1 module, wherein the lattice-like superpixel generation module includes several convolutional layers and max-pooling layers for generating image features of lattice superpixels; and inputting the small target image to be detected into the improved YOLOv5 detection model for detection to obtain the small target detection result. The lattice-like superpixels generated by the lattice-like superpixel module have both segmentation accuracy and topological consistency, and integrating the lattice-like superpixel module into YOLOv5 has a better effect on small target detection, improving detection accuracy.

Description

Technical Field

[0001] This invention belongs to the field of small target detection, and in particular relates to a method and system for small target detection based on lattice-like superpixels. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In recent years, with the continuous development of computer technology, computer vision has been integrated into many areas of daily life. Object detection technology is an important research direction in the field of computer vision. It is an image segmentation method based on the geometric and statistical features of objects, combining object segmentation and recognition into one process. Its accuracy and real-time performance are crucial capabilities of the entire system. Especially in complex scenes where multiple objects need to be processed in real time, automatic object extraction and recognition become particularly important.

[0004] From the traditional framework of manually designed features and shallow classifiers to the end-to-end detection framework based on deep learning, object detection has become increasingly mature and has wide application value in intelligent transportation systems, intelligent monitoring systems, military target detection, and surgical instrument positioning in medical navigation surgery.

[0005] Currently, deep learning-based object detection and recognition have become mainstream methods, which can be mainly represented as: extraction of deep features from images to object recognition and localization based on deep neural networks, with the main deep neural network model used being the Convolutional Neural Network (CNN). Many algorithms have been proposed for object detection, such as the YOLO series. The YOLO series, as a representative algorithm for one-stage object detection, has wide applications in this field.

[0006] Although many object detection algorithms have been proposed, the main challenge in the field of object detection is the detection of small objects. Small objects are generally considered to be objects whose absolute size in an image is less than 32×32 pixels, or objects whose width and height are less than 1 / 10 of the entire image. Due to their small size, the available features are limited, and their semantic information usually appears in shallower feature maps. As the number of neural network layers increases, the detailed information related to small objects may completely disappear. Therefore, small object detection based on convolutional neural networks is currently a difficult problem in the field of object detection algorithms. Summary of the Invention

[0007] To overcome the shortcomings of the prior art, the present invention provides a small target detection method and system based on lattice-like superpixels. The lattice-like superpixels generated by the lattice-like superpixel module have both segmentation accuracy and topological consistency. Furthermore, integrating the lattice-like superpixel module into YOLOv5 has a better effect on small target detection and improves detection accuracy.

[0008] To achieve the above objectives, a first aspect of the present invention provides a small target detection method based on lattice-like superpixels, comprising:

[0009] An improved YOLOv5 detection model for small object detection is constructed. The improved YOLOv5 detection model includes a backbone feature network, a neck network, and a detection head. The backbone feature network includes a lattice superpixel generation module, several CBL modules, and a CSP1 module. The lattice superpixel generation module includes several convolutional layers and max pooling layers to generate image features of lattice superpixels.

[0010] The small target image to be detected is input into the improved YOLOv5 detection model for detection, and the small target detection result is obtained.

[0011] A second aspect of the present invention provides a small target detection system based on lattice-like superpixels, comprising:

[0012] Model building module: Constructs an improved YOLOv5 detection model for small object detection. The improved YOLOv5 detection model includes a backbone feature network, a neck network, and a detection head. The backbone feature network includes a lattice superpixel generation module, several CBL modules, and a CSP1 module. The lattice superpixel generation module includes several convolutional layers and max pooling layers to generate image features for lattice superpixels.

[0013] Small target detection module: Input the image of the small target to be detected into the improved YOLOv5 detection model for detection, and obtain the small target detection result.

[0014] A third aspect of the present invention provides a computer device, comprising: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor communicates with the memory via the bus, and when the machine-readable instructions are executed by the processor, a small target detection method based on lattice superpixels is performed.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs a small target detection method based on lattice-like superpixels.

[0016] The above one or more technical solutions have the following beneficial effects:

[0017] In this invention, the lattice-shaped superpixels generated by the lattice-shaped superpixel module have both segmentation accuracy and topological consistency. Furthermore, integrating the lattice-shaped superpixel module into YOLOv5 has a better effect on small target detection and improves detection accuracy.

[0018] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0020] Figure 1 This is a schematic diagram of the improved YOLOv5 detection model structure in Embodiment 1 of the present invention;

[0021] Figure 2 This is a schematic diagram of the lattice-shaped superpixel generation module in Embodiment 1 of the present invention;

[0022] Figure 3 This is a schematic diagram of the CBL module in Embodiment 1 of the present invention;

[0023] Figure 4 This is a schematic diagram of the CSP1 module in Embodiment 1 of the present invention;

[0024] Figure 5 This is a schematic diagram of the CSP2 module in Embodiment 1 of the present invention;

[0025] Figure 6 This is a schematic diagram of the SPPF structure in Embodiment 1 of the present invention;

[0026] Figure 7 This is a schematic diagram of different movements of the candidate set in Embodiment 1 of the present invention. Detailed Implementation

[0027] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0028] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0029] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0030] Example 1

[0031] like Figure 1 As shown, this embodiment discloses a small target detection method based on lattice superpixels, including: constructing an improved YOLOv5 detection model for small target detection, the improved YOLOv5 detection model including a backbone feature network, a neck network and a detection head; the backbone feature network includes a lattice superpixel generation module, several CBL modules and CSP1 modules, the lattice superpixel generation module includes several convolutional layers and max pooling layers, used to generate image features of lattice superpixels;

[0032] The small target image to be predicted is input into the improved YOLOv5 detection model for detection, and the small target detection result is output.

[0033] In this embodiment, to address the issue of the poor performance of the existing YOLOv5 in small object detection, an improvement is made to the existing YOLOv5 framework. The improved YOLOv5 detection model includes a backbone feature network, a neck network, and a head. A lattice-like superpixel generation module is added to the backbone, and several operation layers are added to the neck to further expand the feature map. Finally, small object detection is added to the head part.

[0034] In this embodiment, a blank layer is added before the backbone network. The blank layer does not process the input small target detection image in any way, so that the lattice superpixel generation module can obtain pure original image information. The convolutions in the first and second CBL modules only change the feature extraction processing of the number of channels, and do not process the aspect ratio of the input image.

[0035] The backbone feature network includes a first CBL module, a second CBL module, a lattice-like superpixel generation module, a first CSP1 module, a third CBL module, a second CSP1 module, a fourth CBL module, a third CSP1 module, a fifth CBL module, a fourth CSP1 module, and an SPPF module connected in sequence.

[0036] like Figure 3 As shown, the CBL module includes convolutional layers (Conv), batch normalization layers (BN), and the SiLU activation function; Figure 4As shown, the CSP1 module includes a first branch structure and a second branch structure in parallel, a connection layer connecting the first branch structure and the second branch structure, and a CBL module connected to the connection layer. The first branch structure includes a CBL module and several first Bottleneck modules. The first Bottleneck module consists of a residual structure, a CBL module, and an add layer. The second branch structure is a CBL module. The convolutional kernels of the convolutional layers in the first to fifth CBL modules are 7, 5, 3, 3, and 3, respectively.

[0037] like Figure 6 As shown, the SPPF module includes a CBL module, a first max pooling layer MaxPool, a second max pooling layer MaxPool, a third max pooling layer MaxPool, a connection layer, and the CBL module connected in sequence. The outputs of the first CBL module, the first max pooling layer MaxPool, the second max pooling layer MaxPool, and the third max pooling layer MaxPool in the SPPF module are respectively connected to the input of the connection layer.

[0038] The first Bottleneck module consists of a CBL-1 module (with 1 convolutional kernel), a CBL-2 module (with 3 convolutional kernels), and an add layer connected in sequence. The residual structure is used to connect the input of the CBL-1 module and the input of the add layer.

[0039] Since the backbone feature network is a relatively deep network, the CSP1 module in the backbone network increases the gradient value of backpropagation between layers through residual structure, avoiding gradient vanishing due to deepening, thus enabling the extraction of finer-grained features without worrying about network degradation.

[0040] High-level semantic features of an image refer to what can be seen. For example, when extracting low-level features from a face, we can extract the face's outline, nose, eyes, etc., while high-level features are displayed as a face. High-level features have richer semantic information, but the target location is relatively coarse. In this embodiment, the CBL module and CSP1 module in the backbone feature network are used to extract high-level semantic features of the image.

[0041] Superpixel generation is a type of image oversegmentation algorithm. It aggregates pixels with the same or similar features (such as color, distance, shape, etc.) in an image to form a larger superpixel with a more specific feature, and uses this newly generated superpixel as the basic unit for subsequent image processing tasks. Compared to previous methods of representing image pixels with a grid, superpixels can significantly reduce the complexity of image processing and better fit the edges of objects. In some object detection fields, especially small object detection, using superpixels to aggregate information from shallow convolutional neural networks can better transmit it to deeper neural networks, thus often achieving higher accuracy than ordinary convolution.

[0042] In this embodiment, the lattice-shaped superpixel generation module, compared to ordinary superpixel generation, produces lattice-shaped superpixels that combine segmentation accuracy and topological consistency. This allows for targeted application in deep learning-based image processing applications, and it performs better when integrated into the YOLOv5 network. The final result of the lattice-shaped superpixel algorithm is the superpixel label belonging to each pixel in the image. Then, based on this label, pooling is performed on the low-level features extracted earlier by the YOLOv5 network (superpixel-based pooling, which involves averaging the features of all pixels belonging to the same superpixel). See attached... Figure 1 As shown, after passing through the lattice-like superpixel generation module, i.e. the Spx layer, the feature map size changes from 640*640*64 to 160*160*64.

[0043] Specifically, such as Figure 2 As shown, the lattice-like superpixel generation module takes two inputs: the feature map of CBL and the unprocessed original image output by Blank. Blank is a layer that does nothing but record information. The original image and the feature map of CBL are input to the Spx layer. Then, superpixels are generated from the original image in the Spx layer, and superpixel pooling is performed on the feature map output by CBL based on the generated superpixels.

[0044] The network structure of the lattice-like superpixel generation module in this embodiment includes six convolutional structures with a kernel size of 3 connected in sequence, and a connection structure. Max pooling layers with a stride of 2 are added between the second and third, fourth and fifth convolutional structures, respectively. The outputs of the first, second, fourth and sixth convolutional structures after bilinear upsampling are connected to the connection structure. The convolutional structure includes a convolutional layer, a batch normalization layer and a ReLU function connected in sequence. The connection structure includes a connection layer Concat, a convolutional layer Conv and a ReLU function connected in sequence.

[0045] In the lattice-like superpixel generation module, all convolutional layers have a kernel size of 3*3, the max-pooling layer has a stride of 2, and each layer except the final connection structure has 64 output channels, while the connection structure has 15 output channels. Furthermore, the outputs of the fourth and sixth convolutional structures are upsampled and used as inputs to the connection structure; the purpose of upsampling is to obtain the resolution of the original image.

[0046] The lattice-like superpixel generation module represents the superpixel generation problem as an energy function, providing a more intuitive measurement of the homogeneity and regularity of the generated superpixels. It divides the entire image into multi-level regular grids based on the user-specified superpixel size to facilitate subsequent operations. An improved hill-climbing algorithm is used to minimize the energy function to generate superpixels with a lattice topology.

[0047] Let I be a given image, F(i) be the feature of pixel i, and S be the feature of pixel i. l For the l-th superpixel, F(S) l ) is a superpixel S l The characteristics of image oversegmentation into superpixels are that I is divided into a small region {S}. l} l The set of, i.e., I = ∪ l S l and l1≠l2, Represented as the label after pixel i is segmented, in this embodiment, superpixel S l The features are calculated by averaging the features of all the pixels in the dataset.

[0048] Pixels within a superpixel share similar characteristics, which is one method for achieving accurate oversegmentation of the image using superpixels. Therefore, the optimal superpixel {S} is generated by minimizing the following energy function. l} l :

[0049]

[0050] Among them, |S l | indicates superpixel S l The size, i.e., S l The number of pixels in image I. |I| represents the number of pixels in image I, and N is the number of superpixels generated by the user. The first term of the function makes pixels within the same superpixel have similar features, the second term makes the superpixels have regular sizes, and the parameter γ is used to balance the weight of each term.

[0051] Hill climbing is a local optimization method that starts from the current superpixel. Begin by iteratively selecting a boundary pixel or block C. i As a candidate set, and from Move to the adjacent superpixel If the energy function formula (1) decreases after the move, then the move will be accepted; otherwise, it will not be accepted, thus achieving the goal of climbing to the higher part of the mountain. This cycle continues until the minimum value of the energy function (i.e., the highest point of the mountain) is reached.

[0052] Clearly, the original hill-climbing algorithm cannot guarantee that the generated superpixels have a lattice topology. Therefore, this embodiment uses an improved constrained hill-climbing algorithm to generate superpixels with a lattice topology. Lattice topology means that the neighbors of each superpixel are fixed, the indices of its surrounding superpixels remain unchanged, and there are the same number of superpixels in each row and each column.

[0053] like Figure 7 As shown, the superpixel satisfies the lattice topology in its initial state. Figure 7 (b) Candidate set C i and C j Movement at the boundary of the intersection of two superpixels does not disrupt the lattice topology, but Figure 7 (d) C i The movement of the candidate set alters the adjacency relationships between superpixels, thus disrupting the lattice topology. To preserve the lattice topology of the superpixels during optimization, four constraints must be imposed on the movement of the candidate set:

[0054] ① Each candidate set can only be moved horizontally or vertically.

[0055] ②When candidate set C i When the movement occurs at the intersection of multiple superpixels, let C... j For along C i Superpixels on the vertical path of the movement direction The adjacent candidate set in the C, either move C i Move C together during the process j Either abandon this move, such as Figure 7 (f).

[0056] ③ If only one candidate set C remains in the superpixel i Then C i It cannot be moved.

[0057] ④ If the movement disrupts the crystal lattice topology, then abandon the movement.

[0058] Constraint ① ensures that the movement of the candidate set between disjoint superpixels does not disrupt the lattice topology. Constraint ② ensures that the movement of the candidate set between intersecting superpixels does not disrupt the lattice topology. Constraint ③ prevents non-adjacent superpixels from becoming adjacent superpixels. Constraint ④ guarantees the connectivity of superpixels.

[0059] According to formula (1), when the candidate set C i Move to Superpixel S l At that time, superpixel S l The energy increment can be calculated using the following formula:

[0060]

[0061] Where F(β) represents the characteristic of β, and β represents C. i S l If β represents a block, then F(β) is the average of the features of all pixels in block β; if β represents a single pixel, then F(β) is the feature of pixel β. σ(||S l |-|I| / N|) is an indicator function, defined as:

[0062]

[0063] Using the four constraints and formula (2), we can determine whether each move of the candidate set reduces formula (1). Calculate separately... and Candidate set C i From superpixels The adjacent superpixel that has moved here, Indicates from Remove C from the middle i If satisfied Then this time C i The movement reduces formula (1), and will be accepted; otherwise, it will not be accepted.

[0064] For C i The movement occurs at the intersection points of multiple superpixels, which is the special case corresponding to constraint ②. It needs to be judged based on the following inequality:

[0065]

[0066] Where C j For along C i Superpixels on the vertical path of the movement direction In order to preserve the lattice topology of the superpixel, the neighboring candidate set in the C is moved. i C needs to be moved together during the process. j If the inequality holds after the movement, then C i and C j The movement is accepted if it is not accepted otherwise.

[0067] Among them, the image feature F(i) used in the superpixel generation module with lattice topology is attached... Figure 2 Extraction of the six convolutional structures and connection structures sequentially connected in the lattice-like superpixel generation module.

[0068] In this embodiment, the neck network adopts an FPN+PAN structure. The features extracted by the backbone feature network are fused at multiple scales before being fed into the head detection part, thus simultaneously meeting the needs of target detection and image classification.

[0069] Specifically, the neck is composed of three FPN networks and three PAN structures stacked sequentially. The FPN network includes a CBL module (with one convolutional kernel), an upsampling layer, a cross-layer connection layer, and a CSP2 module connected in sequence. The output of the first CSP1 module is connected to the input of the cross-layer connection layer in the third FPN network, the output of the second CSP1 module is connected to the input of the cross-layer connection layer in the second FPN network, and the third CSP1 module is connected to the input of the cross-layer connection layer in the first FPN network. Figure 5 As shown, the CSP2 module includes a third branch structure and a fourth branch structure in parallel, a connection layer connecting the third branch structure and the fourth branch structure, and a CBL module connected to the connection layer; the third branch structure includes a CBL module and several second Bottleneck modules, and the second Bottleneck module includes two CBL modules and an add layer connected in sequence.

[0070] In this embodiment, the PAN structure includes a CBL module (with 3 convolutional kernels), a cross-layer connection layer, and a CSP2 module connected in sequence. The output of the CBL module in the first FPN network is connected to the input of the cross-layer connection layer in the third PAN structure. The output of the CBL module in the second FPN network is connected to the input of the cross-layer connection layer in the second PAN structure. The output of the CBL module in the third FPN network is connected to the input of the cross-layer connection layer in the first PAN structure.

[0071] In this embodiment, the Head network includes four convolutional layers (each with a kernel of 1). The input of the first convolutional layer is connected to the output of the CSP2 module in the third FPN network, the input of the second convolutional layer is connected to the output of the CSP2 module in the first PAN structure, the input of the third convolutional layer is connected to the output of the CSP2 module in the second PAN structure, and the input of the fourth convolutional layer is connected to the output of the CSP2 module in the third PAN structure.

[0072] If the input image resolution is 640*640, after training, the final feature map sizes are 20*20, 40*40, and 80*80. Among these three feature maps, the largest, 80*80, is responsible for detecting small targets, corresponding to a receptive field of 640 / 80 = 8*8 for each grid cell of the 640*640 feature map. If the height and width of the target in the original image are less than 8 pixels, the network will have difficulty learning the target's feature information. Furthermore, many images have high resolutions. If downsampling is performed simply, a downsampling factor that is too large will easily lose data information, while a factor that is too small will require the network to store a large number of feature maps in memory during forward propagation, severely depleting GPU resources and easily leading to memory overflow. Therefore, this embodiment adds a small target detection layer in the head, generating a 160*160 feature map to detect small targets, thereby improving the detection accuracy of small targets.

[0073] Example 2

[0074] The purpose of this embodiment is to provide a small target detection system based on lattice-like superpixels, including:

[0075] Model building module: Used to build an improved YOLOv5 detection model for small object detection. The improved YOLOv5 detection model includes a backbone feature network, a neck network, and a detection head. The backbone feature network includes a lattice superpixel generation module, several CBL modules, and a CSP1 module. The lattice superpixel generation module includes several convolutional layers and max pooling layers to generate image features for lattice superpixels.

[0076] Small target detection module: This module takes the image of the small target to be predicted and inputs it into the improved YOLOv5 detection model for detection, and outputs the small target detection results.

[0077] Example 3

[0078] The purpose of this embodiment is to provide a computing device, including a processor, a memory, and a bus. The memory stores machine-readable instructions that can be executed by the processor. When the computer device is running, the processor and the memory communicate through the bus. When the machine-readable instructions are executed by the processor, a small target detection method based on lattice superpixels is performed.

[0079] Example 4

[0080] The purpose of this embodiment is to provide a computer-readable storage medium.

[0081] A computer-readable storage medium is characterized in that a computer program is stored on the computer-readable storage medium, and the computer program is executed by a processor to perform a small target detection method based on lattice-like superpixels.

[0082] The steps and methods involved in the apparatuses of Embodiments 2, 3, and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0083] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0084] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A small target detection method based on lattice-like superpixels, characterized in that, include: An improved YOLOv5 detection model for small object detection is constructed. The improved YOLOv5 detection model includes a backbone feature network, a neck network, and a detection head. The backbone feature network includes a lattice superpixel generation module, several CBL modules, and a CSP1 module. The lattice superpixel generation module includes several convolutional layers and max pooling layers to generate image features of lattice superpixels. The small target image to be detected is input into the improved YOLOv5 detection model for detection, and the small target detection result is obtained; The backbone feature network includes a first CBL module, a second CBL module, a lattice-like superpixel generation module, a first CSP1 module, a third CBL module, a second CSP1 module, a fourth CBL module, a third CSP1 module, a fifth CBL module, a fourth CSP1 module, and an SPPF module connected in sequence. The CBL module includes a convolutional layer, a batch normalization layer, and an activation function; the CSP1 module includes a first branch structure and a second branch structure in parallel, a connection layer connecting the first branch structure and the second branch structure, and a CBL module connected to the connection layer; the first branch structure includes a CBL module and several first Bottleneck modules, the first Bottleneck module is composed of a residual structure, a CBL module, and an add layer, and the second branch structure is a CBL module. The lattice-like superpixel generation module takes two inputs: the feature map of CBL and the unprocessed original image output by Blank.

2. The small target detection method based on lattice-like superpixels as described in claim 1, characterized in that, The lattice-shaped superpixel generation module includes six convolutional structures and connection structures connected in sequence. Max pooling layers are added between the second and third, fourth and fifth convolutional structures, respectively. The input end of the lattice-shaped superpixel generation module, the output end of the second convolutional structure, the output end of the fourth convolutional structure and the output end of the sixth convolutional structure are respectively connected to the connection structure. The convolutional structure includes a convolutional layer, a batch normalization layer, and a ReLU function connected in sequence; the connection structure includes a connection layer, a convolutional layer, and a ReLU function connected in sequence.

3. The small target detection method based on lattice-like superpixels as described in claim 1, characterized in that, In the lattice-like superpixel generation module, a constrained hill-climbing algorithm is used to minimize the energy function of the lattice-like superpixel generation module, as shown in the formula:

4. The small target detection method based on lattice-like superpixels as described in claim 1, characterized in that, The neck network is composed of three FPN networks and three PAN structures stacked together. The FPN network includes a CBL module, an upsampling layer, a cross-layer connection layer, and a CSP2 module connected in sequence. The output of the first CSP1 module is connected to the input of the cross-layer connection layer in the third FPN network, the output of the second CSP1 module is connected to the input of the cross-layer connection layer in the second FPN network, and the third CSP1 module is connected to the input of the cross-layer connection layer in the first FPN network. The CSP2 module includes a third branch structure and a fourth branch structure in parallel, a connection layer connecting the third branch structure and the fourth branch structure, and a CBL module connected to the connection layer; the third branch structure includes a CBL module and several second Bottleneck modules, and the second Bottleneck modules are composed of a CBL module and an add layer.

5. The small target detection method based on lattice-like superpixels as described in claim 4, characterized in that, The PAN structure includes a CBL module, a cross-layer connection layer, and a CSP2 module connected in sequence. The output of the CBL module in the first FPN network is connected to the input of the cross-layer connection layer in the third PAN structure. The output of the CBL module in the second FPN network is connected to the input of the cross-layer connection layer in the second PAN structure. The output of the CBL module in the third FPN network is connected to the input of the cross-layer connection layer in the first PAN structure.

6. The small target detection method based on lattice-like superpixels as described in claim 5, characterized in that, The detection head includes four convolutional layers (Conv). The input of the first convolutional layer is connected to the output of the CSP2 module in the third FPN network. The input of the second convolutional layer is connected to the output of the CSP2 module in the first PAN structure. The input of the third convolutional layer is connected to the output of the CSP2 module in the second PAN structure. The input of the fourth convolutional layer is connected to the output of the CSP2 module in the third PAN structure.

7. A small target detection system based on lattice-like superpixels, employing the small target detection method based on lattice-like superpixels as described in any one of claims 1-6, characterized in that, include: Model building module: Constructs an improved YOLOv5 detection model for small object detection. The improved YOLOv5 detection model includes a backbone feature network, a neck network, and a detection head. The backbone feature network includes a lattice superpixel generation module, several CBL modules, and a CSP1 module. The lattice superpixel generation module includes several convolutional layers and max pooling layers to generate image features of lattice superpixels. Small target detection module: Input the image of the small target to be detected into the improved YOLOv5 detection model for detection, and obtain the small target detection result.

8. A computer device, characterized in that, include: The system includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform a small target detection method based on lattice superpixels as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, performs a small target detection method based on lattice-like superpixels as described in any one of claims 1 to 6.

Images