CN113255837A - Improved CenterNet network-based target detection method in industrial environment - Google Patents
Improved CenterNet network-based target detection method in industrial environment Download PDFInfo
- Publication number
- CN113255837A CN113255837A CN202110723531.3A CN202110723531A CN113255837A CN 113255837 A CN113255837 A CN 113255837A CN 202110723531 A CN202110723531 A CN 202110723531A CN 113255837 A CN113255837 A CN 113255837A
- Authority
- CN
- China
- Prior art keywords
- network
- target
- industrial environment
- improved
- centernet
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0004—Industrial image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/60—Analysis of geometric attributes
- G06T7/66—Analysis of geometric attributes of image moments or centre of gravity
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/70—Determining position or orientation of objects or cameras
- G06T7/73—Determining position or orientation of objects or cameras using feature-based methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30108—Industrial image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Abstract
The invention discloses an improved CenterNet network-based target detection method in an industrial environment, which comprises the following steps: the method comprises the steps of collecting a plurality of images containing a target to be detected, and constructing a training data set of a CenterNet network; improving the CenterNet network, wherein the depth residual error neural network of ResNet-18 is replaced by ResNet-50, and the CenterNet network is trained by utilizing a training data set to obtain a target detection model; and detecting the target to be detected in the industrial environment by using the target detection model to obtain the position information and the classification information of the target. The invention not only ensures the target detection speed, but also improves the target detection precision by using the deep residual error network with deeper layers, particularly enhances the detection and identification capability of multiple targets and small target objects in the industrial environment, and has better universality compared with the traditional algorithm.
Description
Technical Field
The invention relates to the field of artificial intelligence, in particular to the field of target detection, and particularly relates to an improved CenterNet network-based target detection method in an industrial environment.
Background
Object detection is one of the important research directions in computer vision technology. In modern industrial production, the target detection technology has very wide application prospect. In actual production life, most of industrial target detection currently depends on human eye recognition, and many methods for detecting the industrial target through template matching are available, but the manual detection cost and the false detection rate are high, and the template matching is difficult to adapt to task scenes of complex illumination and multi-classification small targets. The working principle of the early target detection algorithm is to achieve the distinguishing effect by splicing strong classifiers, and the early target detection algorithm has the defect that series of non-rigid targets such as people cannot be detected. Based on the problem, some scholars propose an HOG + SVM structure, and the structural algorithm achieves great success in the early stage aiming at the directions of road detection and passing pedestrian detection, but cannot achieve the required real-time performance and accuracy under the industrial use environment.
With the rapid development of the deep learning technology, the target detection based on the deep neural network can be divided into an anchor-based detection algorithm and an anchor-free detection algorithm, the anchor-based detection algorithm clusters anchors in a data set before training, then performs feature extraction, and then performs classification and regression, so that a large amount of time is consumed.
Compared with RCNN and YOLO series in deep learning, the method needs to generate the anchor and then classify and regress, and through the self-contained structure, the method can not generate the anchor, so that the training and reasoning prediction time can be reduced, but the characteristic detection effect on small targets and objects in the vicinity is poor.
Disclosure of Invention
The invention aims to provide a target detection method in an industrial environment, which has low detection cost, high precision and good universality aiming at the defects and the defects of the prior art.
In order to achieve the purpose, the technical scheme of the invention based on the improved CenterNet network target detection method comprises the following steps:
step S1: collecting a plurality of images containing a target to be detected;
step S2: constructing a training data set of the CenterNet network;
step S3: modifying the CenterNet network to include a deep residual neural network that replaces ResNet-18 with ResNet-50;
step S4: training the CenterNet network by using the training data set to obtain a target detection model;
step S5: and detecting the target to be detected in the industrial environment by using the target detection model to obtain the position information and the classification information of the target.
The step S2 of constructing the training data set of the centeret network includes adding a plurality of images of the target to be detected and a plurality of images added by data enhancement, and the data enhancement includes random cropping, flipping, and translation. Step S2, the constructing the training data set further includes adding labeling information corresponding to each image, where the labeling information includes: whether the image contains the target to be detected, the position information and the classification information of the target to be detected.
Wherein the ResNet-50 deep residual neural network in step S3 includes 1 input convolutional layer, a maximum pooling layer, 4 ConvBlock residual blocks and 12 identyblock residual blocks, ConvBlock being a residual block for changing input-output dimensions including 3 convolutional layers, the first part of which includes 1 × 1 convolutional layer for changing input dimensions; the second part comprises a 3 × 3 convolutional layer for feature extraction; the third section includes 1 × 1 convolutional layers for changing output dimensions; the fourth part includes 1 × 1 convolutional layers for changing the picture input dimension and then jumping to connect to the output; IdentityBlock is a residual block for a continuous series network containing 3 convolutional layers, the first part of which comprises 1 x 1 convolutional layers for changing the input dimensions; the second part comprises a 3 × 3 convolutional layer for feature extraction; the third section includes 1 × 1 convolutional layers for changing output dimensions; the fourth part is the input picture jumping connection to the output. Firstly, carrying out convolution on an input industrial target sample picture for once to obtain a downsampling characteristic diagram of 2 times; then, performing maximum pooling once, wherein downsampling is not performed; then convolving the depth residual block consisting of 1 ConvBlock layer and 2 IdentityBlock layers, and sampling the depth residual block by 2 times; performing convolution on the feature map by a depth residual block consisting of 1 ConvBlock layer and 3 IdentityBlock layers, wherein the feature map is sampled by 2 times; performing convolution on the feature map by a depth residual block consisting of 1 ConvBlock layer and 5 IdentityBlock layers, wherein the feature map is sampled by 2 times; finally, performing convolution by a depth residual block consisting of 1 ConvBlock layer and 2 IdentityBlock layers, wherein the feature map is sampled by 2 times; finally, ResNet-50 can obtain a feature map 32 times of down-sampling, compared with the original ResNet-18 network, the generalization capability, feature extraction capability and detection accuracy of the model are improved, and due to the fact that ConvBlock and IdentityBlock convolution structures are used, the ConvBlock and IdentityBlock convolution structures firstly reduce dimensionality and then extract features through a layer of 3 x 3 convolution layer, compared with the ResNet-18 deep residual error neural network in the prior art that dimensionality is not reduced and features are extracted through two layers of 3 x 3 convolution layers, the number of parameters of the neural network is reduced under the same depth, and therefore the training speed of the network is improved.
The step S3 is to improve the centret network, and further includes adding a transposed convolutional layer to the neck feature extraction part of the neural network. The feature map sampled by 32 times is input into the transposed convolutional layer, and the feature map with high resolution sampled by 16 times can be obtained by the added transposed convolutional layer.
Step S5, detecting the target to be detected in the industrial environment by using the target detection model, and sending the high-resolution feature map into the server after obtaining the high-resolution feature mapIn the prediction heads of the network prediction part, the thermodynamic diagram loss, the central point offset loss and the width and height loss of the central point are respectively three parts in the prediction heads. The total loss function is
，
Wherein
Is a loss of the thermodynamic diagram,
is the loss of the center point bias,
is the loss of width and height of the center point,
，
respectively, the weight of each different loss function.
The step S3 is to improve the centret network, and further includes a method of increasing NMS non-maximum suppression in the prediction part of the neural network, so that when the neural network predicts the target to be detected in step S5, redundant local frames can be removed, and a high-quality detection frame can be obtained. Since it is very important that the centret network establishes the central point, and for a large target, there are many local information, so that it is difficult to determine the central point of the same large target, and the local frame cannot be removed by using the maximum pooling method, so that an extra local frame is left in the final detection result, which not only makes the detection result inaccurate, but also increases the overall detection time, so that a method of adding NMS non-maximum suppression is considered here to remove the extra local frame, and obtain a good-quality detection frame.
And the labeling information is labeled by software labelImg.
Compared with the prior art, the invention has the beneficial effects that: 1) the deep residual error network with a deeper layer number is used as an algorithm main body network, the capability of the algorithm for extracting the characteristics of a detected object is enhanced, the detection precision is improved while the detection speed is ensured, and the ConvBlock and IdentityBlock convolution structures are used, so that the dimensionality of the ConvBlock and IdentityBlock convolution structures is reduced firstly, and then a layer of 3 x 3 convolution layer is used for extracting the characteristics, compared with the ResNet-18 deep residual error neural network in the prior art, the dimensionality is not reduced, and the characteristics are extracted by using two layers of 3 x 3 convolution layers, the parameter number of the neural network is reduced under the same depth, and the training speed of the network is improved; 2) the feature extraction part of the CenterNet network is improved, a layer of transposition convolution layer is added, a high-resolution feature map with resolution twice as high as that of the original feature map can be obtained, the target detection capability is improved, and the detection and identification capability of multiple targets and small target objects in an industrial environment is particularly enhanced; 3) by adding an NMS non-maximum value inhibition method in the neural network prediction part, a high-quality detection frame is extracted, and the label classification capability and precision of the CenterNet network on the detection object are further improved; 4) the invention has wide application range, high precision and good universality.
Drawings
FIG. 1 is a flow diagram of target detection based on the modified CenterNet algorithm in an example industrial environment;
fig. 2 is a schematic structural diagram of a backbone network ConvBlock in the embodiment;
fig. 3 is a schematic structural diagram of an identity block of a backbone network in an embodiment;
fig. 4 is a schematic structural diagram of the improved centrnet network model in the embodiment.
Detailed Description
In order to make the objects, technical solutions and advantages of the present application more apparent, the present application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present application and are not intended to limit the present application.
In this embodiment, in conjunction with fig. 1, there is provided a target detection method based on an improved centrnet network in an industrial environment, the method includes the following steps:
step S1: collecting a plurality of images containing a target to be detected;
step S2: constructing a training data set of the CenterNet network;
step S3: the improvement of the CenterNet network comprises a deep residual error neural network which utilizes ResNet-50 to replace ResNet-18, a transposed convolution layer is added to a neck feature extraction part of the deep residual error neural network, and a method for suppressing NMS non-maximum value is added to a prediction part of the deep residual error neural network;
step S4: training the CenterNet network by using the training data set to obtain a target detection model;
step S5: and detecting the target to be detected in the industrial environment by using the target detection model to obtain the position information and the classification information of the target.
Further, in this embodiment, the step S2 of constructing the centret network training data set includes adding a plurality of images of the target to be detected and a plurality of images added by data enhancement, where the data enhancement includes random cropping, flipping, and translating. Step S2, the constructing the training data set further includes adding labeling information corresponding to each image, where the labeling information includes: whether the image contains the target to be detected, the position information and the classification information of the target to be detected. The labeling information can be labeled by software labelImg, and can form a labeling file with a format suitable for the CenterNet network training.
Further, in this embodiment, with reference to fig. 2 to 4, in the step S3, a cenet network is modified, where the ResNet-50 deep residual error neural network includes 1 input convolutional layer, 1 maximum pooling layer, 4 ConvBlock residual blocks and 12 identity block residual blocks, a ConvBlock is a residual block for changing input and output dimensions that includes 3 convolutional layers, and a first part of the ConvBlock includes a 1 × 1 convolutional layer for changing input dimensions; the second part comprises a 3 × 3 convolutional layer for feature extraction; the third section includes 1 × 1 convolutional layers for changing output dimensions; the fourth part comprises a 1 × 1 convolutional layer for changing the input dimension of the picture and then jumping and connecting to the output, and the output dimension of the 1 × 1 convolutional layer of the third part is smaller than the input dimension of a ConvBlock residual block through artificial setting, so that the purpose of reducing the dimension is achieved; IdentityBlock is a residual block for a continuous series network containing 3 convolutional layers, the first part of which comprises 1 x 1 convolutional layers for changing the input dimensions; the second part comprises a 3 × 3 convolutional layer for feature extraction; the third part comprises a 1 x 1 convolutional layer for changing the output dimension, wherein the output dimension of the 1 x 1 convolutional layer of the third part is unchanged compared to the identity block residual block input dimension, and the fourth part is input picture skip connected to the output. Firstly, carrying out convolution on an input industrial target sample picture for once to obtain a downsampling characteristic diagram of 2 times; then, performing maximum pooling once without down-sampling; then convolving the depth residual block consisting of 1 ConvBlock layer and 2 IdentityBlock layers, and sampling the depth residual block by 2 times; performing convolution on the feature map by a depth residual block consisting of 1 ConvBlock layer and 3 IdentityBlock layers, wherein the feature map is sampled by 2 times; performing convolution on the feature map by a depth residual block consisting of 1 ConvBlock layer and 5 IdentityBlock layers, wherein the feature map is sampled by 2 times; finally, performing convolution by a depth residual block consisting of 1 ConvBlock layer and 2 IdentityBlock layers, wherein the feature map is sampled by 2 times; finally, ResNet-50 can obtain a feature map 32 times of down-sampling, compared with the original ResNet-18 network, the generalization capability, feature extraction capability and detection accuracy of the model are improved, and due to the fact that ConvBlock and IdentityBlock convolution structures are used, the ConvBlock and IdentityBlock convolution structures firstly reduce dimensionality and then extract features through a layer of 3 x 3 convolution layer, compared with the ResNet-18 deep residual error neural network in the prior art that dimensionality is not reduced and features are extracted through two layers of 3 x 3 convolution layers, the number of parameters of the neural network is reduced under the same depth, and therefore the training speed of the network is improved.
Further, in this embodiment, with reference to fig. 4, in step S3, the centeret network is improved, and a transposed convolutional layer is added to a neck feature extraction part of the neural network, that is, a transposed convolutional layer is added after an original three-layer transposed convolutional layer is added, and a feature map after down-sampling by 32 times is input into the transposed convolutional layer, so that a high-resolution feature map after up-sampling by 16 times can be obtained by the added transposed convolutional layer.
Further, in this embodiment, after obtaining the high resolution feature map, the high resolution feature map is fed into a prediction head of the depth residual error neural network prediction part, where the prediction head includes three parts, namely, thermodynamic diagram loss, center point bias loss, and center point width and height loss. The total loss function is
，
。
Among these are thermodynamic predicted losses:
wherein
Is the number of positive samples and is,
and
is a hyper-parameter, here taken as 2 and 4, respectively.
Is referred to as
Is detected
The true value of the class or classes,
the class is the number of kinds of the detection target,
is referred to as
Is detected
The predicted value of the class.
wherein
Is the number of positive samples and is,
is the deviation predicted by the model for the center point of each target.
Is the true center point of the image,
is a multiple of the scaling factor and is,
is the value of down-sampling and rounding the true target center point.
Further, in this embodiment, when the neural network predicts the target to be detected in step S5, the method of suppressing the non-maximum value of the NMS is added to the predicted portion of the neural network, so that the redundant local frame can be removed, and a high-quality detection frame can be obtained. Since it is very important that the centret network establishes the central point, and for a large target, there are many local information, so that it is difficult to determine the central point of the same large target, and the local frame cannot be removed by using the maximum pooling method, so that an extra local frame is left in the final detection result, which not only makes the detection result inaccurate, but also increases the overall detection time, so that a method of adding NMS non-maximum suppression is considered here to remove the extra local frame, and obtain a good-quality detection frame.
According to the invention, the deep residual error network ResNet-50 with a deeper layer number is used as an algorithm main body network, so that the characteristic extraction capability of the algorithm on a detected object is enhanced, the detection speed is ensured, and the detection precision is improved; the feature extraction part of the CenterNet network is improved, a layer of transposition convolution layer is added, a high-resolution feature map with resolution twice as high as that of the original feature map can be obtained, the target detection capability is improved, and the detection and identification capability of multiple targets and small target objects in an industrial environment is particularly enhanced; by adding an NMS non-maximum value inhibition method in the neural network prediction part, a high-quality detection frame is extracted, and the label classification capability and precision of a detection algorithm on a detected object are further improved. In conclusion, the algorithm has the advantages of higher detection precision, wide application range and good universality.
The foregoing shows and describes the general principles and features of the present invention, together with the advantages thereof. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are described in the specification and illustrated only to illustrate the principle of the present invention, but that various changes and modifications may be made therein without departing from the spirit and scope of the present invention, which fall within the scope of the invention as claimed. The scope of the invention is defined by the appended claims and equivalents thereof.
Claims (7)
1. Method for detecting targets in industrial environment based on improved CenterNet network, characterized in that it comprises the following steps:
step S1: collecting a plurality of images containing a target to be detected;
step S2: constructing a training data set of the CenterNet network;
step S3: the CenterNet network is improved, and comprises a depth residual error neural network which replaces ResNet-18 with ResNet-50 and adds a transposition convolutional layer to a neck feature extraction part of the neural network, wherein the ResNet-50 depth residual error neural network comprises 1 input convolutional layer, a maximum pooling layer, 4 ConvBlock residual blocks and 12 IdentityBlock residual blocks, the ConvBlock is a residual block which comprises 3 convolutional layers and is used for changing input and output dimensions, the IdentityBlock is a residual block which comprises 3 convolutional layers and is used for a continuous series network, and the transposition convolutional layer is used for converting a feature map which is 32 times of downsampling into a high-resolution feature map which is 16 times of upsampling;
step S4: training the CenterNet network by using the training data set to obtain a target detection model;
step S5: and detecting the target to be detected in the industrial environment by using the target detection model to obtain the position information and the classification information of the target.
2. The improved cenet network based target detection method in industrial environment as claimed in claim 1, wherein said step S2 of constructing the training data set of the cenet network includes adding a plurality of images of the target to be detected and a plurality of images added by data enhancement, the data enhancement includes random cropping, flipping, and shifting; and adding corresponding labeling information of each image, wherein the labeling information comprises whether the image contains a target to be detected, position information and classification information of the target to be detected.
3. The method for detecting the target based on the improved CenterNet network under the industrial environment according to claim 1, wherein the step S5 of detecting the target to be detected in the industrial environment by using the target detection model comprises the step of inputting the high-resolution feature map sampled by 16 times into a prediction head of a neural network prediction part, the prediction head comprises three parts, namely thermodynamic diagram loss, center point bias loss and center point width and height loss, and the total loss function is
，
Wherein
Is a thermodynamic diagram to predict the loss,
is the loss of the center point bias,
is the loss of width and height of the center point,
，
respectively, the weight of each different loss function.
4. The improved centret network based object detection method in industrial environment of claim 1, wherein said step S3 improves the centret network, further comprising a method of increasing NMS non-maxima suppression in the predicted part of the neural network.
5. The improved centret network based network object detection method in industrial environment of claim 2, characterized in that said labeling information is labeled by software labelImg.
6. The improved CenterNet-based network object detection method in an industrial environment of claim 1, wherein the ConvBlock residual block comprises four parts, a first part comprising a 1 x 1 convolutional layer for changing input dimensions; the second part comprises a 3 × 3 convolutional layer for feature extraction; the third section includes 1 × 1 convolutional layers for changing output dimensions; the fourth part includes 1 × 1 convolutional layers for changing the picture input dimension and then jumping-connecting to the output.
7. The improved CenterNet-based network object detection method in an industrial environment of claim 1, wherein the identyblock residual block comprises four parts, a first part comprising a 1 x 1 convolutional layer for changing input dimensions; the second part comprises a 3 × 3 convolutional layer for feature extraction; the third section includes 1 × 1 convolutional layers for changing output dimensions; the fourth part is the input picture jumping connection to the output.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110723531.3A CN113255837A (en) | 2021-06-29 | 2021-06-29 | Improved CenterNet network-based target detection method in industrial environment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110723531.3A CN113255837A (en) | 2021-06-29 | 2021-06-29 | Improved CenterNet network-based target detection method in industrial environment |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113255837A true CN113255837A (en) | 2021-08-13 |
Family
ID=77190035
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110723531.3A Pending CN113255837A (en) | 2021-06-29 | 2021-06-29 | Improved CenterNet network-based target detection method in industrial environment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113255837A (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113869246A (en) * | 2021-09-30 | 2021-12-31 | 安徽大学 | Wheat stripe rust germ summer spore microscopic image detection method based on improved CenterNet technology |
| CN114241589A (en) * | 2022-02-28 | 2022-03-25 | 深圳市城市交通规划设计研究中心股份有限公司 | Bus driver violation judgment method and device based on vehicle-mounted video |
| CN114596876A (en) * | 2022-01-21 | 2022-06-07 | 中国科学院自动化研究所 | Sound source separation method and device |
| CN115170923A (en) * | 2022-07-19 | 2022-10-11 | 哈尔滨市科佳通用机电股份有限公司 | Fault identification method for loss of railway wagon supporting plate nut |
| CN116309586A (en) * | 2023-05-22 | 2023-06-23 | 杭州百子尖科技股份有限公司 | Defect detection method, device, equipment and medium based on convolutional neural network |
| CN117690165A (en) * | 2024-02-02 | 2024-03-12 | 四川泓宝润业工程技术有限公司 | Method and device for detecting personnel passing between drill rod and hydraulic pliers |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111626120A (en) * | 2020-04-24 | 2020-09-04 | 南京理工大学 | Target detection method based on improved YOLO-6D algorithm in industrial environment |
| CN112163602A (en) * | 2020-09-14 | 2021-01-01 | 湖北工业大学 | Target detection method based on deep neural network |
-
2021
- 2021-06-29 CN CN202110723531.3A patent/CN113255837A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111626120A (en) * | 2020-04-24 | 2020-09-04 | 南京理工大学 | Target detection method based on improved YOLO-6D algorithm in industrial environment |
| CN112163602A (en) * | 2020-09-14 | 2021-01-01 | 湖北工业大学 | Target detection method based on deep neural network |
Non-Patent Citations (4)
| Title |
|---|
| 史亚秋: "机器视觉技术在模具检测与定位中的应用研究", 《中国优秀硕士学位论文全文数据库工程科技Ⅰ辑》 * |
| 罗清等: "基于FL-ResNet50的皮肤镜图像分类方法", 《激光与光电子学进展》 * |
| 陈慧岩等: "《智能车辆理论与应用》", 31 July 2018 * |
| 陈诗昆等: "基于 CenterNet的光伏阵列遮盖物检测", 《电视技术》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113869246A (en) * | 2021-09-30 | 2021-12-31 | 安徽大学 | Wheat stripe rust germ summer spore microscopic image detection method based on improved CenterNet technology |
| CN114596876A (en) * | 2022-01-21 | 2022-06-07 | 中国科学院自动化研究所 | Sound source separation method and device |
| CN114241589A (en) * | 2022-02-28 | 2022-03-25 | 深圳市城市交通规划设计研究中心股份有限公司 | Bus driver violation judgment method and device based on vehicle-mounted video |
| CN114241589B (en) * | 2022-02-28 | 2022-08-23 | 深圳市城市交通规划设计研究中心股份有限公司 | Bus driver violation behavior determination method and device based on vehicle-mounted video |
| CN115170923A (en) * | 2022-07-19 | 2022-10-11 | 哈尔滨市科佳通用机电股份有限公司 | Fault identification method for loss of railway wagon supporting plate nut |
| CN116309586A (en) * | 2023-05-22 | 2023-06-23 | 杭州百子尖科技股份有限公司 | Defect detection method, device, equipment and medium based on convolutional neural network |
| CN117690165A (en) * | 2024-02-02 | 2024-03-12 | 四川泓宝润业工程技术有限公司 | Method and device for detecting personnel passing between drill rod and hydraulic pliers |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107341517B (en) | Multi-scale small object detection method based on deep learning inter-level feature fusion | |
| CN108509978B (en) | Multi-class target detection method and model based on CNN (CNN) multi-level feature fusion | |
| CN111047551B (en) | Remote sensing image change detection method and system based on U-net improved algorithm | |
| CN113255837A (en) | Improved CenterNet network-based target detection method in industrial environment | |
| CN109949317B (en) | Semi-supervised image example segmentation method based on gradual confrontation learning | |
| CN111612807B (en) | Small target image segmentation method based on scale and edge information | |
| WO2019237646A1 (en) | Image retrieval method based on deep learning and semantic segmentation | |
| Liu et al. | An attention-based approach for single image super resolution | |
| CN107808376B (en) | Hand raising detection method based on deep learning | |
| CN112488025B (en) | Double-temporal remote sensing image semantic change detection method based on multi-modal feature fusion | |
| CN112183203A (en) | Real-time traffic sign detection method based on multi-scale pixel feature fusion | |
| CN111476133B (en) | Unmanned driving-oriented foreground and background codec network target extraction method | |
| CN112560865B (en) | Semantic segmentation method for point cloud under outdoor large scene | |
| CN113888547A (en) | Non-supervision domain self-adaptive remote sensing road semantic segmentation method based on GAN network | |
| CN111832453A (en) | Unmanned scene real-time semantic segmentation method based on double-path deep neural network | |
| CN115512103A (en) | Multi-scale fusion remote sensing image semantic segmentation method and system | |
| CN111353544A (en) | Improved Mixed Pooling-Yolov 3-based target detection method | |
| CN114048822A (en) | Attention mechanism feature fusion segmentation method for image | |
| CN108564582B (en) | MRI brain tumor image automatic optimization method based on deep neural network | |
| CN111626120A (en) | Target detection method based on improved YOLO-6D algorithm in industrial environment | |
| CN111881743A (en) | Human face feature point positioning method based on semantic segmentation | |
| CN113743505A (en) | Improved SSD target detection method based on self-attention and feature fusion | |
| CN116129291A (en) | Unmanned aerial vehicle animal husbandry-oriented image target recognition method and device | |
| CN112270366A (en) | Micro target detection method based on self-adaptive multi-feature fusion | |
| CN111462090A (en) | Multi-scale image target detection method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20210813 |