SSL-SSD network model-based coal miner safety helmet wearing identification and detection algorithm

Abstract

The invention relates to the technical field of safety helmet recognition and detection, in particular to a coal miner safety helmet wearing recognition and detection algorithm based on an SSL-SSD network model, and the algorithm comprises the steps: obtaining a data set; performing feature enhancement processing on the data set; an SSL-SSD network model is constructed; an optimal SSL-SSD network model is obtained through training; and inputting the image into the optimal SSL-SSD network model, and outputting an identification detection result. According to the method, the resolution of the image data can be improved by performing feature enhancement processing on the data sets of wearing and not wearing the safety helmet of the coal miner, the detection effect of a single-stage target can be improved by constructing the SSL-SSD network model, and meanwhile, the recognition and detection precision of a small target detection object can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of helmet recognition and detection, and in particular to an algorithm for identifying and detecting the wearing of helmets by coal miners based on an SSL-SSD network model. Background Art

[0002] With the improvement of coal mine safety awareness, ensuring that coal miners wear helmets has become an important task. Coal miners wear helmets to ensure that they can reduce the risk of injury as much as possible in a loaded and dangerous working environment. Therefore, every worker entering the mine should strictly abide by the regulations on wearing helmets to ensure their own safety. However, the traditional manual inspection method is inefficient and has subjective problems. Therefore, we need to provide a recognition and detection algorithm for coal miners' helmets to quickly and accurately identify whether coal miners are wearing helmets.

[0003] At present, in the field of coal mine target detection, the SSD algorithm is mainly used. The SSD algorithm combines the speed of the YOLO series algorithms and the accuracy advantages of Fastr R-CNN. By predicting on feature maps of different scales, it can achieve effective detection of targets of different sizes. However, since the SSD algorithm adopts the idea of ​​split learning and uses a low-level network to detect small target features, it will affect the detection accuracy of the SSD algorithm for smaller targets such as whether coal miners are wearing safety helmets. Summary of the invention

[0004] The technical problem to be solved by the present invention is: in order to solve the technical problem of low accuracy of the existing SSD algorithm in identifying and detecting whether coal miners are wearing safety helmets, the present invention provides an identification and detection algorithm for coal miners' safety helmet wearing based on the SSL-SSD network model. By improving the SSD algorithm, the identification and detection accuracy of whether coal miners are wearing safety helmets can be improved.

[0005] The technical solution adopted by the present invention to solve the technical problem is: an identification and detection algorithm for coal miners' helmet wearing based on the SSL-SSD network model, comprising the following steps:

[0006] S1. Obtain the data set of coal miners wearing safety helmets and not wearing safety helmets;

[0007] S2, obtaining the data set in step S1, and performing feature enhancement processing on the data set using a small target extraction algorithm, a fixed mask size, and a center point cutMix algorithm;

[0008] S3, constructing an SSL-SSD network model, integrating the auxiliary focusing branch network AFW module and the residual fusion branch network RFW into the SSD model to obtain the SSL-SSD network model;

[0009] S4, obtaining the data set after feature enhancement processing in step S2 and the SSL-SSD network model in step S3, and training the SSL-SSD network model with the data set to obtain an optimal SSL-SSD network model;

[0010] S5. Obtain the optimal SSL-SSD network model in step S4, input the coal miner image into the optimal SSL-SSD network model, and output the recognition detection result of whether the coal miner image is wearing a safety helmet.

[0011] Therefore, by enhancing the features of the datasets of coal miners wearing safety helmets and not wearing safety helmets, the resolution of image data can be improved. By constructing the SSL-SSD network model, the detection effect of single-stage targets can be improved. At the same time, the recognition and detection accuracy of small target detection objects can be effectively improved.

[0012] Furthermore, the step S2 comprises the following steps:

[0013] S2-1, obtaining the data set in step S1, and extracting small target images in the data set by a small target extraction algorithm to form a small target image data set;

[0014] S2-2, obtaining the small target image data set in step S2-1, and performing feature enhancement processing on the small target image using a fixed mask size and a center point cutMix algorithm.

[0015] Furthermore, the step S2-1 includes the following steps:

[0016] S2-1-1, obtaining the original image in the data set in step S1, the label file corresponding to the original image, and the resolution of the corresponding original image;

[0017] S2-1-2, obtaining the number of targets in the label file in step S2-1-1, traversing the targets in the image, and then obtaining the target resolution according to the four position information of the targets in the label;

[0018] S2-1-3, obtaining the resolution of the original image in step S2-1-1 and the target resolution in step S2-1-2, and judging the ratio between the two. If the ratio between the resolution of the original image and the target resolution is greater than 10, the image is cropped according to the label position information, and the cropped image is stored to obtain a small target image data set;

[0019] S2-1-4, repeat steps S2-1-1 to S2-1-3 multiple times until all the original image data in the data set of step S1 are traversed.

[0020] Further, in step S2-2, the center point cutMix algorithm generates new training samples The calculation formula is:

[0021]

[0022] The center point cutMix algorithm generates new corresponding labels The calculation formula is:

[0023]

[0024] Where: x A With x B are two different training samples, y A For x A The corresponding label, y B For x B The corresponding label, M∈{0,1} W×H Represents the binary mask for filling the cropped area. The ⊙ operator represents pixel-by-pixel multiplication. λ obeys the Beta distribution, λ~Beta(a, a). When a is 1, λ obeys the uniform distribution between (0, 1).

[0025] Further, in step S2-2, cutMix cuts the center coordinate r of the region x The expression is:

[0026] r x ~Unif(0,W);

[0027] cutMix cuts the area center horizontal coordinate r y The expression is:

[0028] r y ~Unif(0,H);

[0029] The length of the cropping area r w The expression is:

[0030]

[0031] The width of the cropping area r h The expression is:

[0032]

[0033] Where: W and H represent the length and width of the image respectively.

[0034] Further, in step S2-2, after the cropping area is determined, the cropping area in the mask Mask is set to 0, and other areas are set to 1, and then the cropping area in sample A is removed, and the cropping area in sample B is extracted and placed in the cropping area of ​​sample A.

[0035] Furthermore, the step S2-2 comprises the following steps:

[0036] S2-2-1, obtain the small target image data set in step S2-1-3, and set the length and width of the mask Mask to the length and width of the small target image;

[0037] S2-2-2. Slide the mask from the upper left corner of the image to be enhanced. Calculate the IOU overlap between the mask and the target label box in the image to be enhanced each time. If there is a position where the IOU with other targets is 0, then this position is used as the final position of the mask. If there is no position where the IOU with other targets is 0 when sliding to the lower left corner, then select the position where the intersection of the mask and the IOU with other targets during the sliding process is the smallest as the final position of the mask.

[0038] Furthermore, in step S2-2-1, the mask length r w The calculation formula is:

[0039] r w =s w ;

[0040] Mask length r h The calculation formula is:

[0041] r h =s h ;

[0042] Where: s w Indicates the length of the small image target, s h Indicates the width of the small image target;

[0043] In step S2-2-2, cutMix cuts the center coordinate r of the region x The calculation formula is:

[0044] r x =s x ;

[0045] cutMix cuts the center of the area to be cut. y The calculation formula is:

[0046] r y =s y ;

[0047] Among them: The horizontal coordinate of the center point of the selected bounding box is set to s x , the vertical axis is set to s y .

[0048] Furthermore, the SSL-SSD network model includes: a backbone feature extraction network, an attention positioning fusion branch network, and a residual fusion branch network;

[0049] The step S3 comprises the following steps:

[0050] S3-1, input the input image into three convolution chains for extracting semantics;

[0051] S3-2, input the three convolution chains into the corresponding auxiliary focusing branch networks AFW1, AFW2, and AFW3. The AFW module inputs the input features into the spatial average pooling layer and the spatial maximum pooling layer respectively, and then fuses the output features of the spatial average pooling layer and the spatial maximum pooling layer in the channel dimension. Then, the fused features are reduced in dimension by a 1*1*2 convolution kernel, and a network feature with a channel of 1 is output;

[0052] S3-3, input the output features of conv4_3 layer, fc layer, and conv8_2 layer into the residual fusion branch network RFW for skip connection. The residual fusion branch network RFW downsamples the shallow high-dimensional spatial features through pooling to keep the same dimension as the deep low-dimensional features, and fuses the features into the deep low-dimensional features. The pooled features are convolved with a 1*1 convolution kernel, and finally the reduced-dimensional features are fused with the low-dimensional spatial features through add fusion.

[0053] Among them: in step S3-1, the first convolution chain contains 3 convolution kernels of 3*3*512, the second convolution chain contains 2 convolution kernels of 3*3*258, and the third convolution chain contains 1 convolution kernel of 3*3*128; in step S3-2, the feature resolution of the AFW1 pooling layer is 60×60, the feature resolution of the AFW2 pooling layer is 120×120, and the feature resolution of the AFW3 pooling layer is 240×240, channel is its input feature chanel, and the strides of spatial pooling are all set with step, which represents the ratio of the resolution of the input image to the resolution of its corresponding backbone feature map; in step S3-3, the downsampling adopts the combination of maximum pooling (MaxPooling) and average pooling (AvgPooling).

[0054] Furthermore, in step S4, the three layers of shallow features fused with the focused branch network, the features fused with conv9_2 and RM1, the features fused with conv10_2 and RM2, and the output features of the conv11_2 layer are respectively input into the SSL-SSD network model, and the SSL-SSD network model is trained to obtain the optimal SSL-SSD network model.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] By enhancing the features of the datasets of coal miners wearing safety helmets and not wearing safety helmets, the resolution of image data can be improved. By constructing the SSL-SSD network model, the detection effect of single-stage targets can be improved. At the same time, the recognition and detection accuracy of small targets can be effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0058] Figure 1 It is a flow chart of the recognition and detection algorithm for coal miners' helmet wearing based on the SSL-SSD network model of the present invention;

[0059] Figure 2 is a flow chart of step S2 of the present invention;

[0060] Figure 3 is a flow chart of step S2-1 of the present invention;

[0061] Figure 4 is a flow chart of step S2-2 of the present invention;

[0062] Figure 5 is a flow chart of step S3 of the present invention;

[0063] Figure 6 It is the principle diagram of the small target algorithm of the present invention;

[0064] Figure 7 A model diagram of the SSL-SSD network model of the present invention;

[0065] Figure 8 The auxiliary focusing branch network AFW model and feature layer size diagram of the present invention;

[0066] Fig. 9 This is a model diagram of the residual fusion branch network RFW of the present invention. DETAILED DESCRIPTION

[0067] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0068] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", "axial", "radial", "circumferential" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In addition, features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more.

[0069] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0070] like Figures 1 to 9 As shown, it is the best embodiment of the present invention. The recognition and detection algorithm of the coal miner's helmet wearing based on the SSL-SSD network model in this embodiment includes the following steps:

[0071] S1. Obtain the data set of coal miners wearing safety helmets and not wearing safety helmets;

[0072] S2, obtaining the data set in step S1, and performing feature enhancement processing on the data set using a small target extraction algorithm, a fixed mask size, and a center point cutMix algorithm;

[0073] S3, constructing an SSL-SSD network model, integrating the auxiliary focusing branch network AFW module and the residual fusion branch network RFW into the SSD model to obtain the SSL-SSD network model;

[0074] S4, obtaining the data set after feature enhancement processing in step S2 and the SSL-SSD network model in step S3, and training the SSL-SSD network model with the data set to obtain an optimal SSL-SSD network model;

[0075] S5, obtaining the optimal SSL-SSD network model in step S4, and inputting the coal miner image into the optimal SSL-SSD network model, and outputting the recognition detection result of whether the coal miner image is wearing a safety helmet. Thus, by enhancing the features of the data set of coal miners wearing safety helmets and not wearing safety helmets, the resolution of the image data can be improved, and by constructing the SSL-SSD network model, the detection effect of single-stage targets can be improved, and at the same time, the recognition detection accuracy of small target detection objects can be effectively improved.

[0076] In this embodiment, step S2 includes the following steps:

[0077] S2-1, obtaining the data set in step S1, and extracting small target images in the data set by a small target extraction algorithm to form a small target image data set;

[0078] S2-2, obtaining the small target image data set in step S2-1, and performing feature enhancement processing on the small target image using a fixed mask size and a center point cutMix algorithm;

[0079] Step S2-1 includes the following steps:

[0080] S2-1-1, obtaining the original image in the data set in step S1, the label file corresponding to the original image, and the resolution of the corresponding original image;

[0081] S2-1-2, obtaining the number of targets in the label file in step S2-1-1, traversing the targets in the image, and then obtaining the target resolution according to the four position information of the targets in the label;

[0082] S2-1-3, obtaining the resolution of the original image in step S2-1-1 and the target resolution in step S2-1-2, and judging the ratio between the two. If the ratio between the resolution of the original image and the target resolution is greater than 10, the image is cropped according to the label position information, and the cropped image is stored to obtain a small target image data set;

[0083] S2-1-4, repeat steps S2-1-1 to S2-1-3 multiple times until all the original image data in the data set of step S1 are traversed.

[0084] In this embodiment, in step S2-2, the center point cutMix algorithm generates new training samples The calculation formula is:

[0085]

[0086] The center point cutMix algorithm generates new corresponding labels The calculation formula is:

[0087]

[0088] Where: x A With x B are two different training samples, y A For x A The corresponding label, y B For x B The corresponding label, M∈{0,1} W×H Represents the binary mask for filling the cropped area. The ⊙ operator represents pixel-by-pixel multiplication. λ follows the Beta distribution. λ~Beta(a, a). When a is 1, λ follows a uniform distribution between (0, 1).

[0089] In step S2-2, cutMix cuts the center coordinate r of the region x The expression is:

[0090] r x ~Unif(0,W);

[0091] cutMix cuts the center of the area to be cut. y The expression is:

[0092] r y ~Unif(0,H);

[0093] The length of the cropping area r w The expression is:

[0094]

[0095] The width of the cropping area r h The expression is:

[0096]

[0097] Where: W and H represent the length and width of the image respectively;

[0098] In step S2-2, after the cropping area is determined, the cropping area in the mask Mask is set to 0, and other areas are set to 1, then the cropping area in sample A is removed, and the cropping area in sample B is extracted and placed in the cropping area of ​​sample A;

[0099] Step S2-2 includes the following steps:

[0100] S2-2-1, obtain the small target image data set in step S2-1-3, and set the length and width of the mask Mask to the length and width of the small target image;

[0101] S2-2-2, slide the mask from the upper left corner of the image to be enhanced, and calculate the IOU overlap between the mask and the target label box in the image to be enhanced each time. If there is a position where the IOU with other targets is 0, then this position is used as the final position of the mask. If there is no position where the IOU with other targets is 0 when sliding to the lower left corner, then the position where the intersection of the mask and other targets during the sliding process is the smallest is selected as the final position of the mask;

[0102] In step S2-2-1, the mask length r w The calculation formula is:

[0103] r w =s w ;

[0104] Mask length r h The calculation formula is:

[0105] r h =s h ;

[0106] Where: s w Indicates the length of the small image target, s h Indicates the width of the small image target;

[0107] In step S2-2-2, cutMix cuts the center coordinate r of the region x The calculation formula is:

[0108] r x =s x ;

[0109] cutMix cuts the area center horizontal coordinate r y The calculation formula is:

[0110] r y =s y ;

[0111] Among them: The horizontal coordinate of the center point of the selected bounding box is set to s x , the vertical axis is set to s y .

[0112] In this embodiment, the SSL-SSD network model includes: a trunk feature extraction network, an attention positioning fusion branch network, and a residual fusion branch network;

[0113] Step S3 includes the following steps:

[0114] S3-1, input the input image into three convolution chains for extracting semantics;

[0115] S3-2, input the three convolution chains into the corresponding auxiliary focusing branch networks AFW1, AFW2, and AFW3. The AFW module inputs the input features into the spatial average pooling layer and the spatial maximum pooling layer respectively, and then fuses the output features of the spatial average pooling layer and the spatial maximum pooling layer in the channel dimension. Then, the fused features are reduced in dimension by a 1*1*2 convolution kernel, and a network feature with a channel of 1 is output;

[0116] S3-3, input the output features of conv4_3 layer, fc layer, and conv8_2 layer into the residual fusion branch network RFW for skip connection. The residual fusion branch network RFW downsamples the shallow high-dimensional spatial features through pooling to keep the same dimension as the deep low-dimensional features, and fuses the features into the deep low-dimensional features. The pooled features are convolved with a 1*1 convolution kernel, and finally the reduced-dimensional features are fused with the low-dimensional spatial features through add fusion.

[0117] Among them: in step S3-1, the first convolution chain contains 3 convolution kernels of 3*3*512, the second convolution chain contains 2 convolution kernels of 3*3*258, and the third convolution chain contains 1 convolution kernel of 3*3*128; in step S3-2, the feature resolution of the AFW1 pooling layer is 60×60, the feature resolution of the AFW2 pooling layer is 120×120, and the feature resolution of the AFW3 pooling layer is 240×240, channel is its input feature chanel, and the strides of spatial pooling are all set with step, which represents the ratio of the resolution of the input image to the resolution of its corresponding backbone feature map; in step S3-3, the downsampling adopts the combination of maximum pooling (MaxPooling) and average pooling (AvgPooling).

[0118] In this embodiment, in step S4, the three layers of shallow features fused with the focused branch network, the features fused with conv9_2 and RM1, the features fused with conv10_2 and RM2, and the output features of the conv11_2 layer are respectively input into the SSL-SSD network model, and the SSL-SSD network model is trained to obtain the optimal SSL-SSD network model.

[0119] It should be noted that the experimental data set is based on the open source helmet wearing detection data set SafetyHelmetWearing-Dataset (see Table 1). The experimental environment is NVIDIA GeForceRTX 2080Ti dual graphics card, the CPU processor model is Intel i7-9700K, the programming language is based on Python, the framework is keras2.2.5, the CUDA and CUDNN versions are both 10.0, the training batch is set to 4, and the exponential decay learning rate is used. The initial learning rate is 0.1 and the decay coefficient is 0.9. This experiment uses traditional SSD, YOLO v2 and the SSL-SSD network model proposed in this patent to perform detection on the SafetyHelmetWearing-Dataset dataset. Compared with YOLO, the SSL-SSD network model From the perspective of v2, the detection accuracy is improved by 12.9%, and the FPS is also increased by 30 images per second. Compared with the traditional SSD, the detection accuracy of the SSL-SSD network model is improved by 2.3%, and the detection speed is reduced by 7 images per unit time. Since the SSL-SSD network model increases a certain amount of calculation based on SSD and loses a certain amount of detection speed, it can still meet certain timeliness requirements.

[0120] Table 1 Safety Helmet Wearing-Dataset

[0121] Detector Extract Model Resolution map FPS SSD300 VGG16 300 86.2％ 75 YOLO v2 DarkNet19 448 75.6％ 38 SSL-SSD VGG16 300 88.5％ 68

[0122] In summary, the present invention can improve the resolution of image data by enhancing the features of the data sets of coal miners wearing safety helmets and not wearing safety helmets. By constructing an SSL-SSD network model, the detection effect of single-stage targets can be improved. At the same time, the recognition and detection accuracy of small target detection objects can be effectively improved.

[0123] The above description is based on the ideal embodiment of the present invention. Through the above description, relevant personnel can make various changes and modifications without departing from the technical concept of the present invention. The technical scope of the present invention is not limited to the content in the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A recognition and detection algorithm for coal miners' helmet wearing based on SSL-SSD network model, characterized in that: The following steps are involved: S1. Obtain the data set of coal miners wearing safety helmets and not wearing safety helmets; S2, obtaining the data set in step S1, and performing feature enhancement processing on the data set using a small target extraction algorithm, a fixed mask size, and a center point cutMix algorithm; S3, constructing an SSL-SSD network model, integrating the auxiliary focusing branch network AFW module and the residual fusion branch network RFW into the SSD model to obtain the SSL-SSD network model; S4, obtaining the data set after feature enhancement processing in step S2 and the SSL-SSD network model in step S3, and training the SSL-SSD network model with the data set to obtain the optimal SSL-SSD network model; S5. Obtain the optimal SSL-SSD network model in step S4, input the coal miner image into the optimal SSL-SSD network model, and output the recognition detection result of whether the coal miner image is wearing a safety helmet.

2. According to the SSL-SSD network model-based recognition and detection algorithm for coal miners wearing safety helmets, the algorithm is characterized in that: The step S2 comprises the following steps: S2-1, obtaining the data set in step S1, and extracting small target images in the data set by a small target extraction algorithm to form a small target image data set; S2-2, obtaining the small target image data set in step S2-1, and performing feature enhancement processing on the small target image using a fixed mask size and a center point cutMix algorithm.

3. The identification and detection algorithm for coal miners' helmet wearing based on the SSL-SSD network model according to claim 2 is characterized in that: The step S2-1 comprises the following steps: S2-1-1, obtaining the original image in the data set in step S1, the label file corresponding to the original image, and the resolution of the corresponding original image; S2-1-2, obtaining the number of targets in the label file in step S2-1-1, traversing the targets in the image, and then obtaining the target resolution according to the four position information of the targets in the label; S2-1-3, obtaining the resolution of the original image in step S2-1-1 and the target resolution in step S2-1-2, and judging the ratio between the two. If the ratio between the resolution of the original image and the target resolution is greater than 10, the image is cropped according to the label position information, and the cropped image is stored to obtain a small target image data set; S2-1-4, repeat steps S2-1-1 to S2-1-3 multiple times until all the original image data in the data set of step S1 are traversed.

4. According to claim 2, the recognition and detection algorithm for coal miners' helmet wearing based on the SSL-SSD network model is characterized in that: In step S2-2, the center point cutMix algorithm generates new training samples The calculation formula is: The center point cutMix algorithm generates new corresponding labels The calculation formula is: Where: x A With x B are two different training samples, y A For x A The corresponding label, y B For x B The corresponding label, M∈{0,1} w×H Represents the binary mask for filling the cropped area. The ⊙ operator represents pixel-by-pixel multiplication. λ obeys the Beta distribution, λ~Beta(a, a). When a is 1, λ obeys the uniform distribution between (0, 1).

5. According to claim 4, the recognition and detection algorithm for coal miners' helmet wearing based on the SSL-SSD network model is characterized in that: In step S2-2, cutMix cuts the center coordinate r of the region x The expression is: r x ~Unif(0,W); cutMix cuts the area center horizontal coordinate r y The expression is: r y ~Unif(0,H); The length of the cropping area r w The expression is: The width of the cropping area r h The expression is: Where: W and H represent the length and width of the image respectively.

6. The identification and detection algorithm for coal miners' helmet wearing based on the SSL-SSD network model according to claim 5 is characterized in that: In step S2-2, after the cropping area is determined, the cropping area in the mask Mask is set to 0, and other areas are set to 1, then the cropping area in sample A is removed, and the cropping area in sample B is extracted and placed in the cropping area of ​​sample A.

7. The identification and detection algorithm for coal miners' helmet wearing based on the SSL-SSD network model according to claim 3 is characterized in that: The step S2-2 comprises the following steps: S2-2-1, obtain the small target image data set in step S2-1-3, and set the length and width of the mask Mask to the length and width of the small target image; S2-2-2, slide the mask from the upper left corner of the image to be enhanced, and calculate the I 0U overlap between the mask and the target label box in the image to be enhanced each time. If there is a position where the I 0U with other targets is 0, then this position is used as the final position of the mask. If there is no position where the I 0U with other targets is 0 when sliding to the lower left corner, then the position with the smallest I 0U intersection with other targets during the sliding process of the mask is selected as the final position of the mask.

8. The identification and detection algorithm for coal miners' helmet wearing based on the SSL-SSD network model according to claim 7 is characterized in that: In step S2-2-1, the mask length r w The calculation formula is: r w ＝s w ； Mask length r h The calculation formula is: r h ＝s h ； Where: s w Indicates the length of the small image target, s h Indicates the width of the small image target; In step S2-2-2, cutMix cuts the center coordinate r of the region x The calculation formula is: r x ＝s x ； cutMix cuts the area center horizontal coordinate r y The calculation formula is: r y ＝s y ； Among them: The horizontal coordinate of the center point of the selected bounding box is set to s x , the vertical axis is set to s y .

9. The identification and detection algorithm for coal miners' helmet wearing based on the SSL-SSD network model according to claim 1 is characterized in that: The SSL-SSD network model includes: Main feature extraction network, attention positioning fusion branch network and residual fusion branch network; The step S3 comprises the following steps: S3-1, input the input image into three convolution chains for extracting semantics; S3-2, input the three convolution chains into the corresponding auxiliary focusing branch networks AFW1, AFW2, and AFW3. The AFW module inputs the input features into the spatial average pooling layer and the spatial maximum pooling layer respectively, and then fuses the output features of the spatial average pooling layer and the spatial maximum pooling layer in the channel dimension. Then, the fused features are reduced in dimension by a 1*1*2 convolution kernel, and a network feature with a channel of 1 is output; S3-3, input the output features of conv4_3 layer, fc layer, and conv8_2 layer into the residual fusion branch network RFW for skip connection. The residual fusion branch network RFW downsamples the shallow high-dimensional spatial features through pooling to keep the same dimension as the deep low-dimensional features, and fuses the features into the deep low-dimensional features. The pooled features are convolved with a 1*1 convolution kernel, and finally the reduced-dimensional features are fused with the low-dimensional spatial features through add fusion. Wherein: in step S3-1, the first convolution chain contains three 3*3*512 convolution kernels, the second convolution chain contains two 3*3*258 convolution kernels, and the third convolution chain contains one 3*3*128 convolution kernel; In step S3-2, the feature resolution of the AFW1 pooling layer is 60×60, the feature resolution of the AFW2 pooling layer is 120×120, and the feature resolution of the AFW3 pooling layer is 240×240. The channel is the input feature chanel, and the strides of the spatial pooling are all set with step, which represents the ratio of the resolution of the input image to the resolution of its corresponding backbone feature map. In step S3-3, the downsampling adopts a combination of maximum pooling (MaxPooling) and average pooling (AvgPooling).

10. The recognition and detection algorithm for coal miners' helmet wearing based on SSL-SSD network model according to claim 1 is characterized in that: In step S4, the three layers of shallow features fused with the focused branch network, the features fused with conv9_2 and RM1, the features fused with conv10_2 and RM2, and the output features of the conv11_2 layer are respectively input into the SSL-SSD network model, and the SSL-SSD network model is trained to obtain the optimal SSL-SSD network model.

Images