Small target detection system in complex construction scene

Abstract

The invention discloses a small target detection system in a complex construction scene, which belongs to the technical field of target detection and is characterized in that a standardized image is acquired through an image acquisition and preprocessing module, and a target is detected by using an improved YOLO11 model. The improved YOLO11 model divides feature information into high and low frequency features through frequency domain decomposition and performs enhancement processing on the high and low frequency features; the edge and texture of the high frequency features are enhanced through expansion convolution and a channel attention mechanism, and semantic perception of the low frequency features is improved through multi-scale convolution and an adaptive gating mechanism. And the enhanced features are fused through an SE attention mechanism, so that the small target detection capability is improved. The system is also provided with a scale adaptive selection module which can dynamically adjust a detection strategy according to a target size to meet different detection requirements. According to the method, the problems of high false alarm and missing detection of small target detection in a complex construction scene are effectively solved, the detection precision and robustness are improved, and the method has important practical application value.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of target detection, and particularly relates to a small target detection system in a complex construction scene. BACKGROUND

[0002] In a complex construction scene, there are a large number of small targets such as safety helmets, tools, parts and the like. These targets are small in size, variable in shape, easily blocked and strongly interfered by the environment, which brings severe challenges to traditional target detection technology. Most of the existing monitoring systems are based on fixed rules or traditional algorithms, and it is difficult to realize accurate detection and continuous tracking of small targets in a complex construction environment. There are problems such as high false positives and many missed detections. Therefore, it has important practical application significance to realize intelligent detection of small targets in a construction scene by using advanced technologies such as deep learning, small target enhancement and multi-scale feature fusion. SUMMARY

[0003] The application provides a small target detection system in a complex construction scene to solve the problems in the prior art.

[0004] To achieve the above purpose, the application provides a small target detection system in a complex construction scene, comprising:

[0005] An image acquisition and preprocessing module is configured to acquire original video images of a construction scene and perform preprocessing to obtain standardized images.

[0006] A target detection module is configured to detect multiple targets in a complex scene by using an improved YOLO11 model to generate a bounding box and class information of the target. The improved YOLO11 model encodes the standardized images into feature information, and decomposes the feature information into high-frequency and low-frequency features for processing.

[0007] A scale self-adaptive selection module is configured to adaptively select targets of different scales according to user requirements and dynamically adjust the detection strategy.

[0008] Optionally, the image acquisition and preprocessing module comprises:

[0009] The original video images are acquired by a high-resolution camera or a drone, and frame sequences are extracted according to timestamps.

[0010] The original images are resized and converted to a gray space.

[0011] The converted images are pixel normalized to generate standardized images.

[0012] Optionally, the improved YOLO11 model comprises:

[0013] a separation unit configured to separate the feature information into high-frequency components and low-frequency components through frequency domain decomposition;

[0014] a high-frequency feature processing unit configured to perform enhancement processing on the high-frequency components in the feature information, and to strengthen edge and texture information of the target;

[0015] a low-frequency feature processing unit configured to perform semantic enhancement on the low-frequency components in the feature information, and to improve global perception ability of the scene;

[0016] a feature fusion unit configured to fuse the enhanced high-frequency feature and the low-frequency feature, and to generate comprehensive feature information for target detection.

[0017] Optionally, the high-frequency feature processing unit comprises:

[0018] a multi-scale dilated convolution module configured to expand the receptive field and capture detailed features of small targets;

[0019] a channel attention mechanism module configured to weight the channels of the high-frequency feature and highlight important feature information.

[0020] Optionally, the low-frequency feature processing unit comprises:

[0021] a multi-scale convolution module configured to extract multi-level semantic information of the low-frequency feature;

[0022] an adaptive gating mechanism module configured to dynamically adjust the semantic weight of the low-frequency feature and enhance global semantic expression.

[0023] Optionally, the step of generating a detection result by the target detection module comprises:

[0024] bounding box regression on the fused high-frequency and low-frequency features to predict the center coordinates, width and height of the target;

[0025] generating a class probability and a confidence score for each predicted target;

[0026] screening effective detection targets according to a confidence threshold.

[0027] Optionally, the scale adaptive selection module comprises:

[0028] a scale classification unit configured to divide the target into a small target, a medium target and a large target according to the pixel proportion of the target in the image;

[0029] a detection strategy adjustment unit configured to enable corresponding detection branch model parameters and dynamically adjust the feature map scale and Anchor settings of the detection layer according to the target scale range selected by the user.

[0030] Optionally, the expression of the scale classification is:

[0031]

[0032] In the formula, A represents the target area.

[0033] Compared with the prior art, the present application has the following advantages and technical effects:

[0034] 1. The present application proposes a method of feature enhancement for high-frequency features and low-frequency features, which can improve the detection accuracy of small targets in complex scenes and provide a method for safety detection of construction scenes, ensuring the safety of construction personnel.

[0035] 2. The YOLO-HLFPM model structure proposed in the present application has good detection effect on small targets, medium targets and large targets, and can quickly respond to the frequently changing scenes in the construction site. The system uses a single-stage detection framework, has fast inference speed, is suitable for real-time monitoring system, and meets the demand for efficient target detection in complex construction scenes.

[0036] 3. The present application introduces a scale self-adaptive selection mechanism, allowing users to customize the size range of the detection target according to the actual application requirements, and automatically matches the corresponding detection branch and receptive field configuration, thereby enhancing the flexibility and configurability of the system, and adapting to the differentiated requirements of different construction scenes for target detection granularity. DETAILED DESCRIPTION

[0037] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application and are incorporated in and constitute a part of this application. The embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0038] Figure 1 The system flowchart of the present application embodiment is shown in the figure;

[0039] Figure 2 The high-frequency feature fusion module schematic diagram of the present application embodiment is shown in the figure;

[0040] Figure 3 The multi-scale convolution module schematic diagram of the present application embodiment is shown in the figure. DETAILED DESCRIPTION

[0041] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0042] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a group of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in a different order from that shown here.

[0043] As Figure 1 shown, the embodiment provides a small target detection system in a complex construction scene, which includes:

[0044] An image acquisition and preprocessing module is configured to acquire original video images of the construction scene and perform preprocessing to obtain standardized images.

[0045] A target detection module is configured to detect multiple targets in a complex scene by using an improved YOLO11 model to generate the bounding box and class information of the target. The improved YOLO11 model encodes the standardized image into feature information and decomposes the feature information into high and low frequency features for separate processing.

[0046] A scale adaptive selection module is configured to adaptively select targets of different scales according to user requirements and dynamically adjust the detection strategy.

[0047] Further, the image acquisition and preprocessing module is configured to preliminarily process the original video images collected in the construction scene to meet the input requirements of the subsequent image denoising, deblurring, feature extraction and target detection modules, thereby improving the robustness and stability of the overall detection system.

[0048] A high-resolution camera or a drone is used to continuously collect video images of the construction scene, and frame sequences are extracted according to timestamps to form an image dataset I raw .

[0049] The collected original image I raw is resized to I resize with a fixed size to keep the input image size consistent and facilitate subsequent network processing.

[0050] I resize =Resize(I raw ,W,H) (1)

[0051] where I raw is the original image, I resize is the resized image, and Resize(.) represents an interpolation scaling operation.

[0052] The resized image I resize is converted from RGB space to grayscale space to highlight the features of the target object and reduce redundant information.

[0053] Finally, the image is standardized. Each pixel p is normalized to the interval [0, 1] to unify the pixel distribution and improve the adaptability of the model to images collected under different conditions.

[0054] Inorm (p) = (I resize (p) - I min ) / (I max -I min ) (2)

[0055] where I norm (p) is the normalized image pixel, I max and I min represent the maximum and minimum values of the image, respectively.

[0056] The above steps complete the preliminary preprocessing of the image, ensuring that the input image meets the subsequent processing requirements in terms of size, pixel distribution, color space, etc.

[0057] Further, the main task of the target detection module is to input the processed video data into the target detection network to detect multiple targets in complex scenes. By combining deep learning model algorithms, the system can effectively monitor the size and position of the target and update it according to each frame of video, ensuring real-time and accuracy of detection.

[0058] The target detection uses an improved YOLO11 model - YOLO-HLFPM (High-Low Frequency Feature Processing Module). This module separates the feature information into high and low frequency features for separate processing.

[0059] The high-low frequency feature processing module includes:

[0060] A separation unit for separating feature information into high and low frequency components through frequency domain decomposition;

[0061] A high frequency feature processing unit for enhancing the high frequency components in the feature information, strengthening the edge and texture information of the target;

[0062] A low frequency feature processing unit for performing semantic enhancement on the low frequency components in the feature information, improving the global perception ability of the scene;

[0063] A feature fusion unit for fusing the enhanced high and low frequency features to generate comprehensive feature information for target detection.

[0064] Further, the high frequency feature processing unit includes:

[0065] A multi-scale dilated convolution module for expanding the receptive field to capture the detailed features of small targets;

[0066] A channel attention mechanism module for weighting the channels of high frequency features to highlight important feature information.

[0067] It also includes a high frequency feature fusion module as shown in Figure 2 for fusing high frequency features.

[0068] Further, the low-frequency feature processing unit comprises:

[0069] a multi-scale convolution module for extracting multi-level semantic information of the low-frequency features;

[0070] The multi-scale convolution module is composed of multiple parallel convolution branches, each branch using different convolution kernel sizes (1, 3, 5) to model semantic context at different scales. The outputs of all branches are spliced in the channel dimension and compressed and fused through a 1x1 convolution to achieve a unified representation of information. To further improve the discriminability of feature representation, a Squeeze-and-Excitation (SE) attention mechanism is introduced after the fusion output to guide the network to adaptively focus on key channels in the low-frequency semantic. Figure 3

[0071] an adaptive gating mechanism module for dynamically adjusting the semantic weight of the low-frequency features to enhance global semantic expression.

[0072] The position of the pedestrian is identified in each video image, a bounding box is generated and the target is classified. The YOLO model is characterized by its single-stage detection, which performs well in terms of detection speed and accuracy. With the HLFPM module, the accuracy is improved, making it suitable for complex scenes in real-time environments.

[0073] The YOLO-HLFPM model extracts features through CNN and predicts the bounding box (x i ,y i ,w i ,h i ) of the target, where (x i ,y i ) is the coordinate of the target center, w i is the target width, and h i is the target height. For each detected target i, the predicted bounding box can be represented as: B i =(x i ,y i ,w i ,h i ), where x i and y i are the center positions of the target, and w i and h i are the height and width, respectively.

[0074] The YOLO-HLFPM model also generates a confidence score p th for each detected target, representing the probability that the target belongs to a certain class. For each target i, the confidence score is: p norm ​= P(class | B i ) only when p i exceeds a set threshold p th , the system will regard the target as valid detection.

[0075] When processing the high and low frequency features in the frequency domain:

[0076] The high and low frequency features of the preprocessed image are separated and enhanced, and through the decomposition and adaptive fusion of frequency domain information, the expression ability of the edge texture information and global semantic information of small targets is improved.

[0077] First, frequency domain decomposition is performed, and the image I norm is sampled and decomposed to separate the high frequency component F high and the low frequency component F low :

[0078] F high ,F low = FrequencyDecompose(I norm ) (3)

[0079] Where FrequencyDecompose(.) represents the frequency domain decomposition operation, which separates the high and low frequency components in combination with the filter.

[0080] When enhancing the high frequency feature, the high frequency component F high is enhanced through multi-scale dilated convolution and channel attention mechanism:

[0081] F highenhanced = CA(DilatedConv(F high )) (4)

[0082] Where DilatedConv(.) represents the dilated convolution operation, and CA(.) represents the channel attention mechanism, which is used to strengthen the edge and texture information of small targets.

[0083] When enhancing the low frequency feature, the low frequency component F low is enhanced through multi-scale convolution and adaptive gating mechanism:

[0084] F lowenhanced = Gating(MultiscaleConv(F low )) (5)

[0085] Where MultiscaleConv(.) represents multi-scale convolution, and Gating(.) represents adaptive gating mechanism, which is used to improve the global perception ability of the scene.

[0086] Finally, high and low frequency feature fusion is performed. The enhanced high frequency features and low frequency features are spliced in the channel, and are adaptively weighted through the SE attention mechanism:

[0087] F fused =SE(Concat(F highenhanced ,F lowenhanced )) (6)

[0088] Where, Concat(.) represents channel dimension splicing, SE(.) is a Squeeze-and-Excitation attention mechanism, which realizes adaptive adjustment of the importance of different frequency domain features.

[0089] This step effectively improves the model's ability to perceive details and recognize semantics of small targets in complex backgrounds through high and low frequency feature cooperative enhancement and fusion.

[0090] Further, the scale adaptive selection module is used to adaptively select targets of different scales according to user needs, so as to adapt to the detection needs of targets of various sizes in the construction scene. The system dynamically judges the scale category of the target by analyzing the pixel ratio of the target in the image, and uses different detection strategies for processing, thereby improving the overall detection performance and flexibility.

[0091] In the present application, the scale of the target is defined as follows: small target (Small Object): area less than 32x32 pixels; medium target (Medium Object): area between 32x32 and 96x96 pixels; large target (Large Object): area greater than 96x96 pixels.

[0092] According to the above definition, the detection system calculates the area A of each target in terms of pixel width w and height h:

[0093] A=w×h (7)

[0094] According to the area A and the set threshold, the target scale is classified:

[0095]

[0096] The system supports users to manually or automatically select the target scale range to be detected through the interface (for example, only small targets are detected, or medium and small targets are detected at the same time, etc.). When the user selects a specific scale, the system will enable the detection branch model parameters of the corresponding scale, and dynamically adjust the feature map scale and Anchor setting of the detection layer to match the receptive field requirements of targets of different sizes.

[0097] This scale adaptive selection strategy not only improves the detection accuracy, but also provides users with the flexibility of on-demand configuration, meeting the key detection needs of different types of targets in actual construction scenarios.

[0098] The above embodiments are only preferred embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A small target detection system for complex construction scenarios, characterized in that, include: The image acquisition and preprocessing module is used to acquire raw video images of the construction scene and preprocess them to obtain standardized images; The target detection module is used to detect multiple targets in complex scenes using an improved YOLO11 model, generating bounding boxes and category information for the targets. The improved YOLO11 model encodes the standardized image into feature information and decomposes the feature information into high-frequency and low-frequency features for separate processing. The scale-adaptive selection module is used to adaptively select targets of different scales for detection according to user needs and dynamically adjust the detection strategy.

2. The system according to claim 1, characterized in that, The image acquisition and preprocessing module includes: Raw video images are captured using high-resolution cameras or drones, and frame sequences are extracted by timestamps. Resize and convert the grayscale space of the original image; The converted image is pixel-normalized to generate a standardized image.

3. The system according to claim 1, characterized in that, The improved YOLO11 model includes: The separation unit is used to separate feature information into high-frequency and low-frequency components through frequency domain decomposition. The high-frequency feature processing unit is used to enhance the high-frequency components in the feature information, thereby strengthening the edge and texture information of the target. The low-frequency feature processing unit is used to semantically enhance the low-frequency components in the feature information, thereby improving the global perception capability of the scene. The feature fusion unit is used to fuse the enhanced high-frequency features with low-frequency features to generate comprehensive feature information for target detection.

4. The system according to claim 3, characterized in that, The high-frequency feature processing unit includes: Multi-scale dilated convolutional modules are used to expand the receptive field and capture detailed features of small objects; The channel attention mechanism module is used to weight the channels of high-frequency features to highlight important feature information.

5. The system according to claim 3, characterized in that, The low-frequency feature processing unit includes: A multi-scale convolution module is used to extract multi-level semantic information from low-frequency features; An adaptive gating mechanism module is used to dynamically adjust the semantic weights of low-frequency features to enhance global semantic expression.

6. The system according to claim 1, characterized in that, The steps for the target detection module to generate detection results include: Boundary box regression is performed on the fused high- and low-frequency features to predict the center coordinates, width, and height of the target. Generate a category probability and confidence score for each predicted target; Valid detection targets are selected based on confidence thresholds.

7. The system according to claim 1, characterized in that, The scale adaptive selection module includes: The scale classification unit is used to classify targets into small, medium, and large targets based on their pixel proportion in the image. The detection strategy adjustment unit is used to enable the corresponding detection branch model parameters according to the target scale range selected by the user, and dynamically adjust the feature map scale and Anchor settings of the detection layer.

8. The system according to claim 7, characterized in that, The expression for the scale classification is: In the formula, A represents the target area.

Images