A method for extracting small-size garbage from computer rooms based on semantic segmentation technology

Abstract

The present invention provides a method for extracting small-sized garbage in a computer room based on semantic segmentation technology, comprising the following steps: photographing all areas of a computer room to obtain image data of the entire computer room, collecting the image data, and then making it into a binary true value image file in a semantic segmentation format, using a SOG detection model to detect the binary true value image file, extracting target data, and sending the target data to a mobile terminal, so as to obtain whether there is small-sized garbage in the computer room. It is very good to directly increase the number of pixels of the small-sized garbage target by adding an upsampling layer in the image encoding part, thereby improving the extractable small-sized garbage features, and merging feature maps of different resolutions together by adding feature matrices, thereby improving the feature fusion capability of the model and improving the accuracy, that is, it is better to obtain specific target data and play a better detection role.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision, and in particular to a method for extracting small-size garbage from a machine room based on semantic segmentation technology. Background Art

[0002] With the continuous development of computer vision technology and artificial intelligence technology, many tasks are gradually becoming intelligent. Through the use of relevant technologies to assist or even replace manual intelligent detection, we can avoid the problems of heavy workload, strong subjectivity and low efficiency of manual quality inspection. When performing intelligent detection tasks, the garbage detection in the computer room is an important link. The operating environment of the computer room should be kept clean. If some computer room garbage is not cleaned up in time, it will affect the ventilation and heat dissipation of the equipment. Some garbage such as cigarette butts will cause fire risks.

[0003] A computer room is a room or area that is used to store computers and related equipment. It usually needs to have good ventilation systems, dust prevention facilities, and safety controls to ensure that computer equipment can operate normally and protect its safety. The reason why computer rooms need to be kept clean and free of garbage is that a large number of key equipment, including servers, network equipment, storage devices, etc., are placed in computer rooms. These devices are essential for the normal operation of an organization. The presence of garbage may cause damage to the equipment, such as garbage entering the interior of the equipment, blocking the vents or blocking the circuit boards, thus affecting the normal operation of the equipment.

[0004] In addition, the computer room is usually an environment with limited space. If a large amount of garbage accumulates in the computer room, it will occupy valuable space, resulting in unreasonable placement of equipment, poor ventilation, and even affecting the heat dissipation effect of the equipment. Keeping the computer room clean can maximize the use of space and improve the efficiency of equipment placement and ventilation quality. Secondly, garbage in the computer room will produce odor and bacterial growth, posing a threat to the computer room environment and personnel health. The accumulation of garbage will also attract insects and pests, further exacerbating hygiene problems. Keeping the computer room clean can provide a clean and healthy working environment, which is conducive to the work efficiency and health of employees. In addition, there are a large number of electrical equipment in the computer room, and garbage may increase the risk of fire. For example, garbage piled up around the equipment may block the cables or cause the wires to short-circuit, causing a fire. Keeping the computer room clean can reduce the risk of fire and improve the safety of the computer room.

[0005] Although most existing deep learning technologies can accurately extract most of the garbage in the computer room (garbage bags, cartons, foam boxes, etc.), the extraction effect of some small-sized garbage (such as cigarette butts, paper scraps, bottle caps, etc.) is obviously poor, and it is very easy to miss or miss. The reason is that some small-sized garbage has small size, small feature quantity, and unclear features, so the features that can be extracted by deep learning are relatively few. On the other hand, since classical deep learning needs to use a large number of convolution and pooling operations when extracting features, the convolution and pooling operations will quickly reduce the size of the feature map, and the pixel points where the small target garbage in the feature map is located will also be reduced. As a result, the small target garbage features will suffer more losses, resulting in garbage missing in the final result.

[0006] The existing methods for extracting garbage based on deep learning models commonly include semantic segmentation models (such as FCN, Unet) or target detection models (such as SSD, YOLO, mask-RCNN). The principle is to use convolutional neural networks to extract target features in images and finally extract the location of garbage. The above-mentioned deep learning-based methods can automatically learn meaningful features from raw data when extracting target features without manually designing features. This greatly reduces the manual workload. Deep neural networks have strong nonlinear expression capabilities and can capture complex patterns and high-order features, thereby improving the performance of target detection. After sufficient training, deep learning models can be applied to different targets and scenarios and have strong generalization capabilities. Deep learning methods can perform end-to-end learning and prediction directly from raw input data, such as images or videos, without the need for tedious manual feature engineering.

[0007] However, the existing deep learning model will obviously miss small-sized garbage (such as cigarette butts, paper scraps, bottle caps, etc.) when extracting garbage from computer rooms. The reason is that on the one hand, the garbage targets are small and the features are insufficient, which makes it difficult for the model to learn enough features to make judgments. On the other hand, the existing deep learning model causes the loss of pixels where small targets are located when extracting features using convolution and pooling layers, resulting in target omissions in the final result. Summary of the invention

[0008] To solve the above problems, the present invention adopts the following technical solution: a method for extracting small-size garbage from a computer room based on semantic segmentation technology, comprising the following specific steps:

[0009] Step S1. Photograph all areas of the computer room to obtain image data of the entire computer room;

[0010] Step S2. Collect the image data and then make it into a binary classification true value image file in a semantic segmentation data export format;

[0011] Step S3. adopting a SOG detection model, wherein the SOG detection model detects the binary true value image file through convolution operation, upsampling operation and matrix addition fusion operation to extract target data;

[0012] Step S4: Send the target data to the mobile terminal to obtain whether there is small-sized garbage in the computer room.

[0013] Furthermore, in the step S1, the image data is obtained by shooting using a plurality of monitoring cameras in a computer room.

[0014] Furthermore, in the step S1, the image data is sent to a control system that controls a plurality of the surveillance cameras.

[0015] Furthermore, in the step S2, the image data is collected by using the control system, and a binary classification true value image file in a semantic segmentation data export format is produced in the control system.

[0016] Furthermore, the SOG detection model is a Small object garbage detection model.

[0017] Furthermore, in the step S3, by using a convolution operation, an upsampling operation and a matrix addition fusion operation, the binary classification true value image file is finally detected and the target data is extracted.

[0018] Furthermore, in the step S3, an upsampling layer is added to the image encoding part to increase the number of pixels of small-sized garbage targets, thereby improving the small-sized garbage features that can be extracted.

[0019] Furthermore, in step S3, feature maps of different resolutions are fused together by adding feature matrices to enhance the feature fusion capability of the model and improve accuracy.

[0020] Furthermore, in step S4, the staff receives the target data through the mobile terminal to obtain whether there is small-sized garbage in the computer room. If so, the obtained target data can accurately determine which area of ​​the computer room it is in, and then quickly clean it up. If not, no cleaning is required.

[0021] Furthermore, the mobile terminal includes a device capable of receiving data via a network.

[0022] The beneficial effects of the present invention are as follows: by using this computer room small-size garbage extraction method based on semantic segmentation technology, it is possible to directly increase the number of pixels of small-size garbage targets by adding an upsampling layer in the image encoding part, thereby improving the extractable small-size garbage features, and merging feature maps of different resolutions together by adding feature matrices, thereby improving the feature fusion capability of the model and improving the accuracy, that is, being able to better obtain specific target data and play a better detection role. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings further illustrate the present invention, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention.

[0024] Figure 1 is a flow chart of the present invention;

[0025] Figure 2 This is a structural diagram of the SOG detection model. DETAILED DESCRIPTION

[0026] The technical solution of the present invention will be further described below in conjunction with the accompanying drawings of the embodiments of the present invention, and the present invention is not limited to the following specific implementation methods. It should be noted that the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

[0027] like Figure 1 to Figure 2 As shown, a method for extracting small-size garbage from a computer room based on semantic segmentation technology includes the following specific steps:

[0028] Step S1. Photograph all areas of the computer room to obtain image data of the entire computer room.

[0029] That is to say, the image data is obtained by shooting with several monitoring cameras in the computer room, and then the image data is sent to a control system that controls the several monitoring cameras.

[0030] Step S2. Collect the image data and then make it into a binary classification true value image file in the semantic segmentation data export format.

[0031] That is to say, the picture data is collected by using the control system, and a binary classification true value image file in a semantic segmentation data export format is produced in the control system.

[0032] Step S3. Using the SOG detection model, the SOG detection model detects the binary true value image file through convolution operation, upsampling operation and matrix addition fusion operation to extract the target data.

[0033] That is to say, by using convolution operation, upsampling operation and matrix addition fusion operation, the binary classification true value image file is finally detected and the target data is extracted. By adding an upsampling layer in the image encoding part to increase the number of pixels of small-sized garbage targets, the small-sized garbage features that can be extracted are improved. Furthermore, feature maps of different resolutions are fused together by adding feature matrices to improve the feature fusion capability of the model and improve accuracy.

[0034] Step S4: Send the target data to the mobile terminal to obtain whether there is small-sized garbage in the computer room.

[0035] That is, the staff receives the target data through the mobile terminal and obtains whether there is small-sized garbage in the computer room. If there is, the staff can accurately determine which area of ​​the computer room it is in through the obtained target data, and then quickly clean it up. If there is no small-sized garbage, it does not need to be cleaned up. It is worth mentioning that the mobile terminal includes a device that can receive data through the network.

[0036] Example 1

[0037] A method for extracting small-size garbage from a computer room based on semantic segmentation technology includes the following specific steps:

[0038] Step S1. Photograph all areas of the computer room to obtain image data of the entire computer room.

[0039] That is to say, the image data is obtained by shooting with several monitoring cameras in the computer room, and then the image data is sent to a control system that controls the several monitoring cameras.

[0040] Step S2. Collect the image data and then make it into a binary classification true value image file in the semantic segmentation data export format.

[0041] That is to say, the picture data is collected by using the control system, and a binary classification true value image file in a semantic segmentation data export format is produced in the control system.

[0042] Step S3. Using the SOG detection model, the SOG detection model detects the binary true value image file through convolution operation, upsampling operation and matrix addition fusion operation to extract the target data.

[0043] That is to say, by adding an upsampling layer to the image encoding part to increase the number of pixels of small-sized garbage targets, the small-sized garbage features that can be extracted are improved. Furthermore, feature maps of different resolutions are fused together by adding feature matrices to improve the feature fusion ability of the model and improve accuracy. Specifically, a binary true value image file with a resolution of 512×512 is first input into the SOG detection model. After performing an upsampling operation on the feature matrix F11, a feature matrix with a resolution of 1024×1024 is obtained, named F31. Subsequently, the feature matrix begins to be extracted by the first-stage small target enhancement module. After performing a convolution operation on the feature matrix F11, a feature matrix F12 is obtained. After performing an upsampling operation on the feature matrix, a feature matrix F21 is obtained. After performing a convolution operation on the feature matrix, a feature matrix F32 is obtained. After using matrix addition to fuse the feature matrix F21 and the feature matrix F32, the two are fused into a feature matrix F33. Thus, the first-stage small target enhancement module completes the feature extraction task. Then, the second-stage small target enhancement module and the third-stage small target enhancement module are executed in the same processing mode. At this point, the picture encoding link is completed. Furthermore, the picture decoding module needs to use the feature matrices F33, F35, and F37 in the encoding module as input. First, the feature matrix F37 is upsampled to obtain the feature matrix F44, and the feature matrix F35 and the feature matrix F44 are added and fused to obtain the feature matrix F43, and the resolution size of the feature matrix is ​​256×256. After performing an upsampling operation on the feature matrix F43, the feature matrix F42 is obtained, and the resolution size of the feature matrix is ​​512×512. After adding and fusion of the feature matrix F33 and the feature matrix F42, the feature matrix F41 is obtained, and then the feature matrix F41 is output as the final result picture.

[0044] Since the convolution operation, upsampling operation and matrix addition fusion operation used are common methods in the field of deep learning, they are not described in detail in this embodiment. That is to say, the convolution kernel size in the convolution operation used is 3×3, and the resolution of the feature matrix after convolution is only reduced by 1 / 2.

[0045] Step S4: Send the target data to the mobile terminal to obtain whether there is small-sized garbage in the computer room.

[0046] That is, the staff receives the target data through the mobile terminal and obtains whether there is small-sized garbage in the computer room. If there is, the staff can accurately determine which area of ​​the computer room it is in through the obtained target data, and then quickly clean it up. If there is no small-sized garbage, it does not need to be cleaned up. It is worth mentioning that the mobile terminal includes a device that can receive data through the network.

[0047] In other words, this method can directly increase the number of pixels of small-sized garbage targets by adding an upsampling layer in the image encoding part, thereby improving the extractable small-sized garbage features, and merging feature maps of different resolutions together by adding feature matrices, thereby improving the feature fusion capability of the model and improving accuracy, that is, being able to better obtain specific target data and play a better detection role.

[0048] In summary, the above-mentioned embodiments are not limitative embodiments of the present invention, and any modifications or equivalent variations made by those skilled in the art on the basis of the essential contents of the present invention are within the technical scope of the present invention.

Claims

1. A method for extracting small-size garbage from a computer room based on semantic segmentation technology, characterized in that: The specific steps include: Step S1. Shoot all areas of the computer room to obtain image data of the entire computer room; Step S2. collecting the image data and then making it into a binary classification true value image file in a semantic segmentation data export format; Step S3. The SOG detection model is used. The SOG detection model detects the binary true value image file through convolution operation, upsampling operation and matrix addition fusion operation to extract the target data; the binary true value image file is input, the size is 512×512 resolution, named as feature matrix F11, after performing upsampling operation on feature matrix F11, a feature matrix with a resolution of 1024×1024 is obtained, named F31, then, the feature matrix begins to be extracted by the first stage small target enhancement module, and the feature matrix F12 is obtained after performing convolution operation on the feature matrix F11, and the feature matrix F12 is obtained after performing convolution operation on the feature matrix F11. After performing upsampling operation on the feature matrix, the feature matrix F21 is obtained. After performing convolution operation on the feature matrix F31, the feature matrix F32 is obtained. After the feature matrix F21 and the feature matrix F32 are fused by matrix addition, the two are fused into the feature matrix F33. The first-stage small target enhancement module completes the feature extraction task. The second-stage small target enhancement module and the third-stage small target enhancement module are executed in the same processing mode as the first-stage small target enhancement module to obtain feature matrices F35 and F37, completing the picture encoding link. Then the picture decoding module needs to use the feature matrices F33, F37 in the encoding module. F35 and F37 are used as input, and an upsampling operation is performed on the feature matrix F37 to obtain a feature matrix F44. The feature matrix F35 and the feature matrix F44 are added and fused to obtain a feature matrix F43, and the resolution size of the feature matrix is ​​256×256. The feature matrix F42 is obtained after an upsampling operation is performed on the feature matrix F43, and the resolution size of the feature matrix is ​​512×512. The feature matrix F33 and the feature matrix F42 are added and fused to obtain a feature matrix F41, and then the feature matrix F41 is output as the target data of the final result image; Step S4. The target data is sent to the mobile terminal, so as to obtain whether there is small-sized garbage in the computer room.

2. The method for extracting small-size garbage from a computer room based on semantic segmentation technology according to claim 1 is characterized by: In the step S1, the image data is obtained by shooting using a plurality of monitoring cameras in a computer room.

3. The method for extracting small-size garbage from a computer room based on semantic segmentation technology according to claim 2 is characterized by: In the step S1, the image data is sent to a control system that controls a plurality of the surveillance cameras.

4. The method for extracting small-size garbage from a computer room based on semantic segmentation technology according to claim 3 is characterized by: In the step S2, the image data is collected by using the control system, and a binary classification true value image file in a semantic segmentation data export format is produced in the control system.

5. The method for extracting small-size garbage from a computer room based on semantic segmentation technology according to claim 1 is characterized by: The SOG detection model is a Small object garbage detection model.

6. The method for extracting small-size garbage from a computer room based on semantic segmentation technology according to claim 1 is characterized by: In step S3, by using convolution operation, upsampling operation and matrix addition fusion operation, the binary classification true value image file is finally detected and the target data is extracted.

7. The method for extracting small-size garbage from a computer room based on semantic segmentation technology according to claim 6 is characterized by: In step S3, an upsampling layer is added to the image encoding part to increase the number of pixels of small-sized garbage targets, thereby improving the small-sized garbage features that can be extracted.

8. The method for extracting small-size garbage from a computer room based on semantic segmentation technology according to claim 7 is characterized by: In step S3, feature maps of different resolutions are fused together by adding feature matrices to improve the feature fusion capability of the model and improve accuracy.

9. The method for extracting small-size garbage from a computer room based on semantic segmentation technology according to claim 1, characterized in that: In step S4, the staff receives the target data through the mobile terminal to obtain whether there is small-sized garbage in the computer room. If so, the obtained target data can accurately determine which area of ​​the computer room it is in and then quickly clean it up. If not, no cleaning is required.

10. The method for extracting small-size garbage from a computer room based on semantic segmentation technology according to claim 9, characterized in that: The mobile terminal includes a device that can receive data through a network.

Images