Construction method of multi-scene coal dust recognition and detection lightweight model

Abstract

The invention discloses coal dust detection, and provides a construction method of a multi-scene coal dust identification and detection lightweight model, aiming at the problems that the existing coal dust detection research does not comprehensively consider the multi-scene demand and has insufficient segmentation and identification of semitransparent dust edges and backgrounds and model lightweight. According to the method, the model is constructed through the steps of coal dust image acquisition, data set construction, network model improvement, evaluation index selection, result evaluation and the like. In the aspect of network model improvement, based on a YOLOv10n model, a CAA attention mechanism module is introduced into a backbone network to enhance the feature extraction capability, and a C2fEMSCP module is introduced into a backbone and a neck part to improve the recognition precision and reduce the omission ratio. The improved PWD-YOLOv10n model structure covers a backbone network, a neck network and a detection head, and all the parts work cooperatively to achieve effective detection of coal dust. The high-performance PWD-YOLOv10n model trained by the method disclosed by the invention can be used for rapidly and accurately detecting the coal dust in a complex environment.

Description

Technical Field

[0001] The present invention relates to the technical field of coal dust detection, and in particular to a multi-scenario coal dust recognition and detection lightweight model and a construction method thereof. Background Art

[0002] Coal dust scenes are complex, and existing research has mostly focused on one of the complex scenes in the laboratory or natural environment. It has failed to fully consider the detection needs of multiple scenes, and there are still deficiencies in the segmentation and recognition of translucent dust edges and backgrounds and the lightweighting of the model. Summary of the Invention

[0003] This paper elaborates on the construction method of a lightweight model for multi-scenario coal dust recognition and detection, including steps such as video capture, dataset supplementation, image processing, annotation preparation, and model training; and explains the superior performance indicators of the improved YOLOv10n-DP model in coal dust recognition tasks. In order to achieve the above-mentioned invention objectives, the technical solutions adopted by the present invention are as follows:

[0004] The present invention provides a pine wood nematode diseased tree classification detection method based on UAV remote sensing, which comprises the following steps: (1) coal dust image acquisition; (2) data set construction; (3) network model improvement; (4) evaluation index selection; (5) result evaluation;

[0005] (1) Remote sensing image acquisition of diseased trees: Plan experiments in the selected research area, use high-definition industrial cameras to capture dynamic images, and save the image data;

[0006] (2) Dataset construction: The original images captured by the camera were cropped and enhanced, divided into training and validation sets in proportion, and the diseased trees were accurately labeled using annotation tools;

[0007] (3) Network model improvement: A high-precision, low-missing-detection PWD-YOLOv10 model was designed, and the CAA attention mechanism module was introduced into the backbone network of the YOLOv10n model to enhance the feature extraction capability; subsequently, the C2f_EMSCP module was introduced into the backbone and neck parts respectively to further improve the recognition accuracy of the model and effectively reduce the missing-detection rate.

[0008] The model is based on the YOLOv10n network, and the structure of the model includes:

[0009] Backbone

[0010] Layer 1: Use 3x3 convolution to downsample the input image and extract preliminary features.

[0011] Layer 2: 3x3 convolution is used again for downsampling to further extract features.

[0012] Layer 3: Multi-scale features are integrated through the C2f module to improve the model's ability to perceive multi-scale information.

[0013] Layer 4: Introducing the CAA attention mechanism module to enhance attention to key features

[0014] Layer 5: Use 3x3 convolution to downsample the feature map to 1 / 8 of the original size.

[0015] Layer 6: Use the C2f module to further fuse features and enhance feature expression.

[0016] Layer 7: The CAA module is applied again to strengthen the model’s recognition of key features.

[0017] Layer 8: Downsampling is performed through the SCDown module to obtain more in-depth information.

[0018] Layer 9: Use the C2f_EMSCP module and combine it with multi-scale convolution to improve the expressiveness of features.

[0019] Layer 10: Optimize feature channels through the CAA module to improve model accuracy.

[0020] Layer 11: Use the SCDown module to continue downsampling the feature map and extract deeper information.

[0021] Layer 12: Introduce the C2f_EMSCP module to perform multi-scale fusion and enhance feature extraction capabilities.

[0022] Layer 13: Apply the CAA module to weight important feature channels and further improve feature extraction.

[0023] Layers 14-15: SPPF and PSA modules are added respectively to enhance the model's adaptability to features of different scales.

[0024] Neck (neck network)

[0025] Layers 16-17: Feature maps at different levels are fused through upsampling and feature concatenation.

[0026] Layer 18: The fused features are further processed through the C2f_EMSCP module to extract more effective information.

[0027] Layers 19-20: Upsampling and feature concatenation are used again to fuse multi-level feature maps.

[0028] Layer 21: Apply the C2f module to process the concatenated feature map and generate features suitable for small target detection.

[0029] Layer 22: Use 3x3 convolution to downsample the feature map to P4 scale.

[0030] Layer 23: Multi-scale information fusion is completed through feature splicing.

[0031] Layer 24: Generates a feature map suitable for medium-sized object detection through the C2f_EMSCP module.

[0032] Layer 25: Use the SCDown module to downsample the feature map to P5 scale.

[0033] Layer 26: Feature splicing module, further integrating multi-scale features.

[0034] Layer 27: Use the C2fCIB module to generate feature maps suitable for large object detection.

[0035] Head

[0036] Layer 28: Use the v10Detect target detection module to detect small, medium, and large targets on the feature maps of layers 21, 24, and 26, respectively, and output the final detection results.

[0037] Improvements to the model include:

[0038] The CAA (Context Anchor Attention) attention mechanism is used to enhance model feature extraction, highlight important contextual features, and suppress irrelevant background information.

[0039] The process is as follows:

[0040] First, global average pooling is performed on the input feature map to extract the global context information of each channel.

[0041] Capture the overall semantics of the image; then, use 1×1 convolution to fuse information between channels and generate preliminary context anchors; then, use 1×11 and 11×1 depth-separable convolutions to extract spatial context features in the horizontal and vertical directions respectively, enhancing the model's perception of horizontal and vertical structures; then use 1×1 convolution to further integrate features and extract key context anchor information; finally, generate an attention feature map after the Sigmoid activation function. Next, stack the convolution results of each subgroup to form a multi-scale feature map; finally, use 1×1 convolution to perform channel fusion on the stacked multi-scale feature map Y to obtain the final output feature map.

[0042] The beneficial effects of the present invention are:

[0043] The method disclosed in the present invention is a coal dust classification detection method based on an improved PWD-YOLOv10n algorithm. By improving the model recognition accuracy, recognition missed detection rate and other aspects, a high-performance PWD-YOLOv10n model is trained, thereby realizing rapid and accurate detection of coal dust in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is: a flow chart of the lightweight model detection of coal dust identification and detection according to the present invention;

[0045] Figure 2 Figure 1: Network structure diagram of the model of the present invention, with the red box being the improved module. DETAILED DESCRIPTION

[0046] The specific implementation process of this method is as follows:

[0047] Network model improvement. To address the shortcomings of existing models, this paper introduces a CAA attention mechanism module into the backbone network of the YOLOv10n model to enhance feature extraction capabilities. Subsequently, the C2f_EMSCP module is introduced into the backbone and neck regions, further improving the model's recognition accuracy and effectively reducing the missed detection rate.

[0048] The structure of the model includes:

[0049] Backbone

[0050] Layer 1: Use 3x3 convolution to downsample the input image and extract preliminary features;

[0051] Layer 2: Use 3x3 convolution again for downsampling to further extract features;

[0052] Layer 3: Fusion of multi-scale features through the C2f module to improve the model's ability to perceive multi-scale information;

[0053] Layer 4: Introduces the CAA attention mechanism module to enhance attention to key features;

[0054] Layer 5: Use 3x3 convolution to downsample the feature map to 1 / 8 of the original size;

[0055] Layer 6: Use the C2f module to further fuse features and enhance feature expression;

[0056] Layer 7: The CAA module is applied again to strengthen the model’s recognition of key features;

[0057] Layer 8: Downsampling is performed through the SCDown module to obtain more in-depth information;

[0058] Layer 9: Use the C2f_EMSCP module and combine it with multi-scale convolution to improve the expressiveness of features;

[0059] Layer 10: Optimize feature channels through the CAA module to improve model accuracy;

[0060] Layer 11: Use the SCDown module to continue downsampling the feature map and extract deeper information;

[0061] Layer 12: Introducing the C2f_EMSCP module to perform multi-scale fusion and enhance feature extraction capabilities;

[0062] Layer 13: Apply the CAA module to weight important feature channels to further improve feature extraction.

[0063] Layers 14-15: SPPF and PSA modules are added to enhance the model's adaptability to features of different scales.

[0064] Neck (neck network)

[0065] Layers 16-17: Feature maps at different levels are fused through upsampling and feature concatenation;

[0066] Layer 18: The fused features are further processed through the C2f_EMSCP module to extract more effective information;

[0067] Layers 19-20: Upsampling and feature concatenation are used again to fuse multi-level feature maps;

[0068] Layer 21: Apply the C2f module to process the concatenated feature maps and generate features suitable for small target detection;

[0069] Layer 22: Use 3x3 convolution to downsample the feature map to P4 scale;

[0070] Layer 23: Multi-scale information fusion is completed through feature splicing;

[0071] Layer 24: Generates a feature map suitable for medium-sized target detection through the C2f_EMSCP module;

[0072] Layer 25: Use the SCDown module to downsample the feature map to P5 scale;

[0073] Layer 26: Feature splicing module, further integrating multi-scale features;

[0074] Layer 27: Use the C2fCIB module to generate feature maps suitable for large target detection; Head (detection head)

[0075] Layer 28: Use the v10Detect target detection module to detect small, medium, and large targets on the feature maps of layers 21, 24, and 26, respectively, and output the final detection results.

[0076] The improved PWD-YOLOv10n model is as follows Figure 2 shown.

[0077] ① Enhance model feature extraction through CAA (Context Anchor Attention) attention mechanism, highlight important contextual features, and suppress irrelevant background information.

[0078] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in the technical field, several improvements can be made without departing from the principles of the present invention. These improvements should also be regarded as the scope of protection of the present invention without paying creative labor.

Claims

1. A method for constructing a lightweight model for multi-scenario coal dust recognition and detection, characterized by: A high-resolution camera was used to capture the dust injection process in real time, and key frame images were captured from the video using a timed shooting method. A dust image database was constructed through field acquisition in multiple scenes, and explosion-proof industrial cameras and high-speed cameras were used to dynamically track and capture six types of high-dust risk scenes, such as underground coal mines, building demolition, and road construction. Gamma correction was performed on the images extracted from the video, and the visibility of dust particles was enhanced by adjusting the image brightness curve. The ROIS algorithm was used to process the images, and a threshold was set to automatically identify areas that may contain dust particles and perform local magnification or enhancement. The 5104 processed images were annotated using LabelImg software, and the annotated data was randomly divided into training, validation, and test sets in a ratio of 6:2:

2. Based on the annotated data, three mainstream target detection models, YOLOv10, EfficientDet, and SSD, were trained using the pytorch framework. Data enhancement was performed during training, and the adam optimizer was used with a learning rate decay strategy, combining cross entropy loss and Smooth L1 loss to optimize the model.

2. The method for constructing a lightweight model for multi-scenario coal dust recognition and detection according to claim 1 is characterized in that: The data enhancement includes rotating, translating, scaling, cropping, and color enhancement operations on the image.

3. A multi-scenario coal dust recognition and detection lightweight model constructed based on the method of claim 1, characterized in that: This model is an improved YOLOv10n-DP model, which is based on the improved YOLOv10 model. In the coal dust recognition task, the Precision reaches 98.9%, Recall reaches 99%, F1-Score reaches 97%, and mAP reaches 98.20%. The inference time for a single image is 8.0ms.

4. The model is used to detect dust targets in real-time video streams and output dust location, category, and concentration information.

Images