CN110929577A - An improved target recognition method based on YOLOv3 lightweight framework - Google Patents
An improved target recognition method based on YOLOv3 lightweight framework Download PDFInfo
- Publication number
- CN110929577A CN110929577A CN201911013341.1A CN201911013341A CN110929577A CN 110929577 A CN110929577 A CN 110929577A CN 201911013341 A CN201911013341 A CN 201911013341A CN 110929577 A CN110929577 A CN 110929577A
- Authority
- CN
- China
- Prior art keywords
- yolov3
- tiny
- data set
- sample data
- network model
- 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
- 238000000034 method Methods 0.000 title claims abstract description 51
- 238000012549 training Methods 0.000 claims abstract description 32
- 238000001514 detection method Methods 0.000 claims abstract description 26
- 238000012360 testing method Methods 0.000 claims abstract description 19
- 238000007781 pre-processing Methods 0.000 claims abstract description 4
- 238000002372 labelling Methods 0.000 claims abstract description 3
- 238000011176 pooling Methods 0.000 claims description 19
- 238000012545 processing Methods 0.000 claims description 12
- 238000011056 performance test Methods 0.000 claims description 5
- 238000003064 k means clustering Methods 0.000 claims description 4
- 230000007613 environmental effect Effects 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 230000001629 suppression Effects 0.000 claims description 3
- 238000013519 translation Methods 0.000 claims description 3
- 230000006870 function Effects 0.000 description 35
- 230000004913 activation Effects 0.000 description 12
- 238000010586 diagram Methods 0.000 description 6
- 238000013527 convolutional neural network Methods 0.000 description 5
- 230000008569 process Effects 0.000 description 5
- 238000009432 framing Methods 0.000 description 4
- 230000003993 interaction Effects 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 230000006399 behavior Effects 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000011156 evaluation Methods 0.000 description 3
- 230000003213 activating effect Effects 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000002708 enhancing effect Effects 0.000 description 2
- 238000005286 illumination Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 210000004027 cell Anatomy 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000012886 linear function Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
-
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
本发明公开了一种基于YOLOv3的轻量级框架改进的目标识别方法，通过将YOLOv3的轻量级版本YOLOv3‑tiny和SENet相结合得到YOLOv3‑tiny‑SE来进行目标检测和识别。具体包括：在不同的路况、行车环境和天气条件下进行车辆、行人和交通环境图片的采集，对采集到的数据进行预处理以及数据增强，制作并完善目标识别样本集，对样本集进行标注，然后将样本集分为训练集和测试集两部分，在YOLOv3‑tiny中嵌入SENet结构，得到YOLOv3‑tiny‑SE，在训练集上训练YOLOv3‑tiny‑SE，在测试集上测试YOLOv3‑tiny‑SE，然后和YOLOv3‑tiny性能进行比较。本发明提出的目标识别方法泛化能力强，且可以加快目标检测速度、提高小目标检测的准确率、提高模型参数对噪声的鲁棒性。
The invention discloses an improved target recognition method based on a lightweight framework of YOLOv3, which is used for target detection and recognition by combining YOLOv3-tiny and SENet, a lightweight version of YOLOv3, to obtain YOLOv3-tiny-SE. Specifically, it includes: collecting pictures of vehicles, pedestrians and traffic environments under different road conditions, driving environments and weather conditions, preprocessing and data enhancement of the collected data, making and improving target recognition sample sets, and labeling the sample sets , and then divide the sample set into two parts: training set and test set, embed the SENet structure in YOLOv3‑tiny to obtain YOLOv3‑tiny‑SE, train YOLOv3‑tiny‑SE on the training set, and test YOLOv3‑tiny on the test set ‑SE, and then compare with YOLOv3‑tiny performance. The target recognition method proposed by the invention has strong generalization ability, and can speed up target detection, improve the accuracy of small target detection, and improve the robustness of model parameters to noise.
Description
Technical Field
The invention relates to the field of computer vision and deep learning, in particular to a lightweight frame improved target identification method based on YOLOv 3.
Background
The unmanned technology is characterized in that collected real-time road condition video images (including information of pedestrians, vehicles, traffic signs and the like) are subjected to modeling processing through camera tools such as a driving recorder and the like to obtain automobile drive-by-wire state parameters, and then the parameters are input into a decision and control network model of the vehicle to perform decision control on vehicle behaviors. The target detection is a precondition of behavior decision, and the target detection method based on deep learning not only ensures the accuracy of multi-target detection classification, but also meets the real-time processing requirement. Currently, mainstream target detection methods based on machine learning are mainly divided into two major methods based on Region nomination (Region probable) as an idea and a regression method.
The methods based on region nomination mainly include methods such as R _ CNN, SPP _ Net, Fast R _ CNN and Fast R _ CNN. The Fast R _ CNN method greatly reduces the operation time by utilizing a method of sharing a characteristic layer, in addition, the classification and regression method is changed from using an SVM model to using SoftMax for classification, and classification and regression are simultaneously carried out in a multitask mode, so the operation time of target detection is reduced to a certain extent, but all candidate frames are searched in the selective search process, the process is very time-consuming, and the bottleneck problem of calculation speed exists. The fast R _ CNN method directly utilizes the RPN network to extract candidate frames, and the operations of Region nomination, classification, regression and the like share convolution characteristics together, so that the operation speed is further improved.
A representative method based on a regression method is a YOLO method, the YOLO method simplifies the whole process of target detection, video frame images are zoomed into images with uniform size, but only two boundary frames are predicted for each cell in the specific implementation process, and the two boundary frames belong to one category, so that the YOLO method is not high enough in small target detection accuracy, weak in generalization capability and incapable of meeting the requirements of unmanned driving on multi-target detection.
Disclosure of Invention
The invention aims to provide a lightweight frame improved target identification method based on YOLOv3, which is used for solving the problems in the prior art and aims to improve the robustness of model parameters to noise while ensuring the improvement of the target detection speed and the accuracy of small target detection.
In order to achieve the purpose, the invention provides the following scheme: the invention provides a lightweight frame improved target identification method based on YOLOv3, which comprises the following steps:
s1, collecting pictures of vehicles, pedestrians and traffic environments under different road conditions, driving environments and weather conditions, and making an initial sample data set; specifically, step S1 includes:
s11, starting a driving recorder or a high-definition camera installed by a vehicle, and shooting driving information in a road traffic environment in real time;
s12, performing framing processing on the obtained driving video, and extracting images of each frame to obtain driving image sequence sets in different driving environments;
s13, screening the driving image sequence set obtained in the step S12, and selecting driving images under different illumination conditions, traffic time periods and environmental backgrounds;
and S14, marking the selected driving image by using a marking tool, framing a target area, wherein the target area comprises vehicles, pedestrians and traffic signs, and then labeling the target area to make an initial sample data set.
S2, preprocessing and enhancing the picture data in the initial sample data set to obtain a target identification sample data set, specifically, the step S2 includes: and (5) processing the characteristic parameters of the target to be recognized through translation, rotation, saturation adjustment, exposure adjustment and noise addition operation on the initial sample data set obtained in the step (S1) to obtain a complete sample data set.
And S3, dividing the obtained target recognition sample data set into a training set and a test set.
S4, embedding a SEnet structure in a YOLOv3-tiny method frame to obtain a YOLOv3-tiny-SE network model, specifically, the step S4 comprises the following steps:
embedding a SEnet structure in a YOLOv3-tiny method, embedding the SEnet structure after each pooling layer and after a convolutional layer before a final output result, adding the SEnet structure after the pooling layers of the 2 nd, 4 th, 6 th, 8 th, 10 th and 12 th layers and the convolutional layers of the 13 th, 14 th, 15 th, 19 th and 22 th layers by modifying a YOLOv3-tiny.cfg file, and specifying the characteristic channel values 16, 32 th, 64 th, 128 th, 256 th, 512 th, 1024 th, 256 th, 512 th and 256 of the input global pooling layer of the SEnet structure as the number of the characteristic channels output by the embedding layers to obtain a YOLOv3-tiny-SE network model.
S5, training a YOLOv3-tiny-SE network model on a training set, specifically, the step S5 comprises:
s51, after the sample data set is enhanced and the parameters are marked in the step S2, recalculating the anchorbox value of the prepared complete sample data set; calculating an anchor value in a traffic environment by using a K-means clustering method, and comprising the following steps of: reading the marked data set, randomly taking out the width and height values of one picture as coordinate points and initial clustering centers, and performing iterative computation by using a K-means clustering method to obtain a specific anchor value;
s52, setting hyper-parameters and network parameters during training, inputting the training set into a YOLOv3-tiny-SE network model for multi-task training, and storing a trained network model weight file.
S6, testing the performance of YOLOv3-tiny-SE on the test set, specifically, the step S6 comprises the following steps:
s61, loading the trained network model weight file obtained in the step S52, inputting a test set into the trained YOLOv3-tiny-SE network model, and obtaining a multi-scale characteristic diagram through a convolutional layer, a pooling layer, a SENET structure and an upper sampling layer;
s62, activating the x, y, confidence coefficient and category probability of the network prediction by adopting a logistic function, and obtaining the coordinates, confidence coefficient and category probability of all prediction frames through threshold judgment;
and S63, removing redundant detection frames from the result obtained in the step S62 through non-maximum suppression processing, and generating a final target detection frame and a recognition result.
S7, comparing the performance test result of the YOLOv3-tiny-SE obtained in the step S6 on the test set with the performance of the YOLOv3-tiny to obtain the result of performance comparison.
The invention discloses the following technical effects: aiming at the problems that the detection speed of a target is low and the detection accuracy of a small target is not accurate enough in a complex environment in the prior art, a lightweight version YOLOv3-tiny of YOLOv3 is combined with a SEnet structure to obtain a YOLOv3-tiny-SE network model, and the obtained YOLOv3-tiny-SE network model is used for target detection and identification. The invention also provides an improved activation function: PSReLU function and is used to activate the model. By the target identification method, real-time road condition video images acquired by camera tools such as a vehicle data recorder can be processed quickly, accurately in real time, and scientific basis is provided for decision control of vehicle behaviors in automatic driving.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without inventive exercise.
FIG. 1 is a block diagram of the system of the present application;
FIG. 2 is a diagram of the improved PSReLU activation function of the present application;
FIG. 3 is a diagram of a YOLOv3-tiny-SE network model structure;
FIG. 4 is a SENET structure diagram.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.
Referring to fig. 1-4, the invention provides an improved target identification method based on a lightweight framework of YOLOv3, which specifically comprises the following steps:
s1, collecting pictures of vehicles, pedestrians and traffic environments under different roads, traffic environments and weather conditions, and making an initial sample data set; the method comprises the following steps:
s11, starting a driving recorder or a high-definition camera installed by a vehicle, and shooting driving videos in a road traffic environment in real time;
s12, performing framing processing on the obtained driving video, and extracting images of each frame to obtain driving image sequence sets in different driving environments;
s13, screening the driving image sequence set, and selecting the driving images under different illumination conditions, traffic periods and environmental backgrounds;
and S14, marking the selected driving image by using a marking tool, marking the target to be identified in the sample data set by using a Labelimg sample marking tool for parameters, framing a target area (specifically comprising three types of vehicles, pedestrians and traffic signs), and marking a label to manufacture an initial sample data set.
S2, preprocessing and enhancing the data of the manufactured initial sample data set, perfecting the initial sample data set and obtaining a target identification sample data set; the method comprises the following specific steps:
processing the initial sample data set, and performing the following operations through a writing program: on the basis of the existing initial sample data set, the characteristic parameters of the target to be recognized are processed through translation, rotation, saturation degree and exposure amount adjustment and noise adding operation, sample data are added, the target recognition sample data set is obtained to improve the initial sample data set, and the generalization capability of the neural network is improved.
And S3, dividing the obtained target recognition sample data set into a training set and a test set according to the proportion of 7:3, 8:2 or 8: 1.
S4, embedding a SENet structure in a YOLOv3-tiny method to obtain a YOLOv3-tiny-SE network model; comprises the following steps:
s41, improving the lightweight frame of YOLOv3 in the step shown in the table 1, and embedding a SENET structure into the YOLOv3-tiny frame to obtain an improved YOLOv3-tiny-SE network model shown in the figure 3;
s42 and YOLOv3-tiny as a lightweight framework of YOLOv3, wherein the overall network architecture is shown in Table 1, and specifically comprises 13 convolutional layers, 6 pooling layers, 2 fusion layers, 1 upsampling layer and 2 output layers with different scales, compared with YOLOv3, the overall architecture reduces residual layers, replaces a series of pooling layers, and simultaneously omits some convolutional layers and FPN networks for extracting features, thereby simplifying the network, reducing the operation complexity and improving the recognition speed;
the processing idea of S43 and Yolov3-tiny on target detection and recognition is the same as that of Yolov3, and Yolov3 carries out Batch Normalization (BN) operation after convolution of each convolutional layer so as to avoid the occurrence of network training overfitting phenomenon, and then uses a Leaky-Relu function as an activation function after Batch normalization;
s44 and YOLOv3 add an FPN structure on the basis of the previous two generations of methods to improve the recognition accuracy of the multiple scale target, and the specific steps are as follows:
firstly, an image pyramid is established for an image, image pyramids of different levels are input into corresponding networks, target detection is respectively carried out on feature maps of different depths, the feature maps of a future layer are up-sampled through the feature map of a current layer and are utilized, so that the current feature map can obtain information of the future layer, low-order semantic information and high-order semantic information are organically fused, the detection precision is improved, the defects of the former two versions of methods are improved, an FPN (field programmable gate network) is introduced into a YOLOv3 frame, the precision of small target identification is improved, and the identification of traffic signs is more effective;
s45 and YOLOv3-tiny-SE network model as shown in figure 3, the SEnet structure firstly performs global average pooling on input feature maps to obtain feature maps (c is the number of feature channels) with the size of cx1 × 1, then passes through two full connection layers, performs the processes of firstly reducing and then increasing dimensions, finally performs nonlinear processing by using a Sigmoid function to obtain weights with the size of cx1 × 1, and then performs multiplication operation on the weights and the original input feature maps at corresponding positions to obtain the final output result;
s46, embedding a SENet structure in a YOLOv3-tiny method; the method comprises the following specific steps:
embedding a SENet structure after each pooling layer and after the convolutional layer before final output, adding the SENet structure after the pooling layers of the 2 nd, 4 th, 6 th, 8 th, 10 th and 12 th layers and the convolutional layers of the 13 th, 14 th, 15 th, 19 th and 22 th layers by modifying a YOLOv3-tiny file, and specifying the characteristic channel values 16, 32 th, 64 th, 128 th, 256 th, 512 th, 1024 th, 256 th, 512 th and 256 th layers of the global pooling layer of the SENet structure as the number of the characteristic channels output by the embedding layers to obtain a YOLOv3-tiny-SE network model;
the network depth of S47 and YOLOV3-tiny originally is 24 layers, and becomes 35 layers after embedding SENet structure, the main purpose of embedding SENet network is to strengthen useful information and compress useless information, wherein the concrete steps of the embedded SENet structure take a second layer of pooling layer as an example, the feature map output by the pooling layer is 208 x 16, and is also the input feature map size of a Global pooling layer (Global posing), the feature map of 1 x 16 is obtained after Global averaging pooling, then the feature map of 1 x 1 is obtained after dimensionality reduction through a first Full connection (Full connected), the feature map of 1 x 16 is obtained after dimensionality increase through a second Full connection (Full connected), finally the weight value of 1 x 16 is obtained after activation of a Sigmoid function, and finally the weight value of 1 x 16 is obtained by multiplying the feature map with the input feature map 208 and the weight value of 16 is obtained.
TABLE 1
S5, training a YOLOv3-tiny-SE network model on a training set, and specifically comprises the following steps:
s51, clustering real target frames of the target to be recognized marked in the training set, obtaining an initial candidate frame of the target predicted in the training set by taking the area interaction ratio IOU as an evaluation index, and inputting the initial candidate frame as an initial parameter into a Yolov3-tiny-SE network model, wherein the method specifically comprises the following steps:
clustering the real target frame of the training data set by adopting a K-means method according to a distance formula dis (box, centroid) which is 1-IOU (box, centroid); the IOU (box, centroid) is the area interaction ratio of the predicted target frame and the real target frame, and when the IOU (box, centroid) is used as an evaluation standard and the value is not less than 0.5, the predicted candidate frame is used as an initial target frame;
the area interaction ratio IOU (box, centroid) formula is shown as follows:
wherein, boxpredAnd boxtruthRespectively representing the areas of the predicted target frame and the real target frame, wherein the proportion of the intersection and the union of the predicted target frame and the real target frame is the initial candidate of the real target frame and the predicted initial targetAverage interaction ratio of the bounding box;
s53, calling the initial weight of the YOLOv3-tiny network, setting the super parameters, the learning rate, the iteration step number N and the size of batch _ size, wherein the super parameters can be adjusted according to the obtained model data; then inputting the training data set into a Yolov3-tiny-SE network model for training until the loss value output by the training data set is smaller than a certain threshold Q1 or reaches the preset maximum iteration number N, and stopping training to obtain a well-trained Yolov3-tiny-SE network model; the method comprises the following specific steps:
calling initial network weight of YOLOv3-tiny, inputting a training data set into a YOLOv3-tiny network for training, outputting a loss function value, continuously training and adjusting the network weight and a bias value according to the loss function value until the loss function value output by the training set is smaller than a threshold value Q1 or the training is stopped after the maximum iteration number N is reached to obtain a trained YOLOv3-tiny-SE network model;
the loss function (object) is expressed by the following formula
Each term of the loss function corresponds to a loss of the prediction center coordinate, a loss of the prediction bounding box, a loss of the prediction confidence degree and a loss of the prediction category. The loss functions of the prediction center coordinates and the bounding boxes are expressed by error square sum, and the loss functions of the prediction categories and the confidence degrees are expressed by courtyard cross entropy loss functions;
in the above formula, λcoordError coefficients that are predicted coordinates; lambda [ alpha ]noobjectAn error coefficient that does not contain a confidence level when identifying an object; k2Indicating the number of meshes into which the input image is divided; m represents the predicted target frame number of each grid; x is the number ofi,yi,wi,hiRespectively representing the abscissa and ordinate of the predicted center point of the target and the width and height,
respectively representing the horizontal and vertical coordinates, the width and the height of the central point of a real target;
the ith grid representing the jth candidate box is responsible for detecting the object;
indicating that the ith grid in which the jth candidate box is positioned is not responsible for detecting the object; ciAnd
respectively representing the prediction confidence coefficient and the real confidence coefficient of the target to be detected in the ith grid; p is a radical ofi(c) And
respectively representing a prediction probability value and a real probability value of the target identification in the ith network belonging to a certain category;
the activation function of YOLOv3 after convolution layer adopts a Leaky-ReLU function, and the expression of the function is shown as the following formula:
the Leaky-ReLU function is evolved from the ReLU function, values obtained by the ReLU function when x is less than or equal to 0 are all 0, so that the problem that the neuron weight cannot be updated possibly occurs along with training, the problem is not large in influence on a deep neural network, but the problem is large in influence on a neural network with a shallow layer number, so that the output of the Leaky-ReLU function, which is 0 in a complex number field, is changed into a linear function with a small slope on the basis of ReLu, the output of a negative number field is reserved, but the parameter a is a proper parameter value determined through artificial prior and repeated training for many times, and the noise robustness in an inactive state cannot be ensured. Based on the above problem, the present embodiment proposes an improved activation function PSReLU (Parametric Soft plus-ReLU) function as shown in fig. 2, and the function expression is as follows:
in the positive value domain, the YOLOv3-tiny adopts an activation function, leak-ReLU, which is the same as a ReLU function, in the negative value domain, a Softplus function is adopted, log2 units are shifted downwards, and a parameter α is taken as a learnable parameter in the network, and back propagation training is carried out in the network to be jointly optimized with other network layers.
S6, testing the performance of the YOLOv3-tiny-SE network model on a test set, specifically, the step S6 comprises the following steps:
s61, loading the trained network weight, inputting the test set into the trained network, and obtaining a multi-scale characteristic diagram through a convolution layer, a pooling layer, a SENET structure and an up-sampling layer;
s62, activating the x, y, confidence coefficient and category probability of the network prediction by adopting a logistic function, and obtaining the coordinates, confidence coefficient and category probability of all prediction frames through threshold judgment;
s63, removing redundant detection frames from the result through non-maximum suppression processing (NMS) to generate a final target detection frame and an identification result;
s64, comparing the effect of the YOLOv3-tiny native model with the effect of the improved YOLOv3-tiny-SE network model by using the original activation function; respectively using an improved activation function and a primitive activation function to perform performance test on a Yolov3-tiny native model; respectively using an improved activation function and an original activation function to perform performance test on a YOLOv3-tiny-SE network model;
s65, respectively inputting the test sets obtained in the step S3 into the network corresponding to the step S61 for performance detection, and obtaining the final evaluation indexes for model performance, including Average accuracy mean mAP (mean Average precision), number of frames Per second detected FPS (frames Per second) and Recall rate (Recall).
S7, comparing the performance test result of the YOLOv3-tiny-SE network model obtained in the step S6 on the test set with the performance of YOLOv3-tiny to obtain a performance comparison result.
It will be evident to those skilled in the art that the invention is not limited to the details of the foregoing illustrative embodiments, and that the present invention may be embodied in other specific forms without departing from the spirit or essential attributes thereof. The present embodiments are therefore to be considered in all respects as illustrative and not restrictive, the scope of the invention being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. Various modifications and improvements of the technical solution of the present invention may be made by those skilled in the art without departing from the spirit of the present invention, and the technical solution of the present invention is to be covered by the protection scope defined by the claims.
Claims (6)
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911013341.1A CN110929577A (en) | 2019-10-23 | 2019-10-23 | An improved target recognition method based on YOLOv3 lightweight framework |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911013341.1A CN110929577A (en) | 2019-10-23 | 2019-10-23 | An improved target recognition method based on YOLOv3 lightweight framework |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110929577A true CN110929577A (en) | 2020-03-27 |
Family
ID=69849270
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911013341.1A Pending CN110929577A (en) | 2019-10-23 | 2019-10-23 | An improved target recognition method based on YOLOv3 lightweight framework |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110929577A (en) |
Cited By (28)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111709381A (en) * | 2020-06-19 | 2020-09-25 | 桂林电子科技大学 | Road environment target detection method based on YOLOv3-SPP |
| CN111723747A (en) * | 2020-06-22 | 2020-09-29 | 西安工业大学 | A lightweight and high-efficiency target detection method applied to embedded platforms |
| CN111754498A (en) * | 2020-06-29 | 2020-10-09 | 河南科技大学 | A Conveyor Belt Idler Detection Method Based on YOLOv3 |
| CN111814621A (en) * | 2020-06-29 | 2020-10-23 | 中国科学院合肥物质科学研究院 | A multi-scale vehicle pedestrian detection method and device based on attention mechanism |
| CN111882554A (en) * | 2020-08-06 | 2020-11-03 | 桂林电子科技大学 | An intelligent detection method for power line faults based on SK-YOLOv3 |
| CN111986240A (en) * | 2020-09-01 | 2020-11-24 | 交通运输部水运科学研究所 | Drowning person detection method and system based on visible light and thermal imaging data fusion |
| CN112070713A (en) * | 2020-07-03 | 2020-12-11 | 中山大学 | A Multi-scale Object Detection Method Introducing Attention Mechanism |
| CN112084870A (en) * | 2020-08-10 | 2020-12-15 | 同济大学 | YOLO-based multi-target detection method and device in traffic scene |
| CN112115880A (en) * | 2020-09-21 | 2020-12-22 | 成都数之联科技有限公司 | Ship pollution monitoring method, system, device and medium based on multi-label learning |
| CN112183361A (en) * | 2020-09-29 | 2021-01-05 | 中科人工智能创新技术研究院(青岛)有限公司 | Goal detection method and system combining target detection and dynamic difference |
| CN112233175A (en) * | 2020-09-24 | 2021-01-15 | 西安交通大学 | Chip positioning method and integrated positioning platform based on YOLOv3-tiny algorithm |
| CN112270252A (en) * | 2020-10-26 | 2021-01-26 | 西安工程大学 | Multi-vehicle target identification method for improving YOLOv2 model |
| CN112307921A (en) * | 2020-10-22 | 2021-02-02 | 桂林电子科技大学 | A vehicle-mounted multi-target recognition tracking and prediction method |
| CN112396002A (en) * | 2020-11-20 | 2021-02-23 | 重庆邮电大学 | Lightweight remote sensing target detection method based on SE-YOLOv3 |
| CN112464910A (en) * | 2020-12-18 | 2021-03-09 | 杭州电子科技大学 | Traffic sign identification method based on YOLO v4-tiny |
| CN112487969A (en) * | 2020-11-30 | 2021-03-12 | 苏州热工研究院有限公司 | Method for acquiring position of inspection target of inspection robot of steam generator |
| CN112561885A (en) * | 2020-12-17 | 2021-03-26 | 中国矿业大学 | YOLOv 4-tiny-based gate valve opening detection method |
| TWI723823B (en) * | 2020-03-30 | 2021-04-01 | 聚晶半導體股份有限公司 | Object detection device and object detection method based on neural network |
| CN112633308A (en) * | 2020-09-15 | 2021-04-09 | 北京华电天仁电力控制技术有限公司 | Detection method and detection system for whether power plant operating personnel wear safety belts |
| CN112686186A (en) * | 2021-01-05 | 2021-04-20 | 润联软件系统(深圳)有限公司 | High-altitude parabolic recognition method based on deep learning and related components thereof |
| CN112861919A (en) * | 2021-01-15 | 2021-05-28 | 西北工业大学 | Underwater sonar image target detection method based on improved YOLOv3-tiny |
| CN113128555A (en) * | 2021-03-09 | 2021-07-16 | 西南交通大学 | Method for detecting abnormality of train brake pad part |
| CN113537013A (en) * | 2021-07-06 | 2021-10-22 | 哈尔滨理工大学 | Multi-scale self-attention feature fusion pedestrian detection method |
| CN114399440A (en) * | 2022-01-13 | 2022-04-26 | 马上消费金融股份有限公司 | Image processing method, image processing network training method and device and electronic equipment |
| CN114463772A (en) * | 2022-01-13 | 2022-05-10 | 苏州大学 | Traffic sign detection and recognition method and system based on deep learning |
| CN114898152A (en) * | 2022-05-13 | 2022-08-12 | 电子科技大学 | Embedded elastic self-expanding general learning framework |
| CN115239949A (en) * | 2022-07-28 | 2022-10-25 | 湖南锐思华创科技有限公司 | Target detection and tracking method, storage medium and device based on yolo neural network |
| CN116030351A (en) * | 2023-03-28 | 2023-04-28 | 南京信息工程大学 | Cascade network-based aerial image ship segmentation method |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180101744A1 (en) * | 2016-10-09 | 2018-04-12 | Airspace Systems Inc. | Spatio-temporal awareness engine for priority tree based region selection across multiple input cameras and multimodal sensor empowered awareness engine for target recovery and object path prediction |
| US20180253613A1 (en) * | 2017-03-06 | 2018-09-06 | Honda Motor Co., Ltd. | System and method for vehicle control based on red color and green color detection |
| CN109214399A (en) * | 2018-10-12 | 2019-01-15 | 清华大学深圳研究生院 | A kind of improvement YOLOV3 Target Recognition Algorithms being embedded in SENet structure |
| CN109325418A (en) * | 2018-08-23 | 2019-02-12 | 华南理工大学 | Pedestrian recognition method in road traffic environment based on improved YOLOv3 |
| CN109447034A (en) * | 2018-11-14 | 2019-03-08 | 北京信息科技大学 | Traffic mark detection method in automatic Pilot based on YOLOv3 network |
| CN109934121A (en) * | 2019-02-21 | 2019-06-25 | 江苏大学 | A pedestrian detection method in orchard based on YOLOv3 algorithm |
| CN110222769A (en) * | 2019-06-06 | 2019-09-10 | 大连理工大学 | Improved target detection method based on YOLOV3-tiny |
-
2019
- 2019-10-23 CN CN201911013341.1A patent/CN110929577A/en active Pending
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180101744A1 (en) * | 2016-10-09 | 2018-04-12 | Airspace Systems Inc. | Spatio-temporal awareness engine for priority tree based region selection across multiple input cameras and multimodal sensor empowered awareness engine for target recovery and object path prediction |
| US20180253613A1 (en) * | 2017-03-06 | 2018-09-06 | Honda Motor Co., Ltd. | System and method for vehicle control based on red color and green color detection |
| CN109325418A (en) * | 2018-08-23 | 2019-02-12 | 华南理工大学 | Pedestrian recognition method in road traffic environment based on improved YOLOv3 |
| CN109214399A (en) * | 2018-10-12 | 2019-01-15 | 清华大学深圳研究生院 | A kind of improvement YOLOV3 Target Recognition Algorithms being embedded in SENet structure |
| CN109447034A (en) * | 2018-11-14 | 2019-03-08 | 北京信息科技大学 | Traffic mark detection method in automatic Pilot based on YOLOv3 network |
| CN109934121A (en) * | 2019-02-21 | 2019-06-25 | 江苏大学 | A pedestrian detection method in orchard based on YOLOv3 algorithm |
| CN110222769A (en) * | 2019-06-06 | 2019-09-10 | 大连理工大学 | Improved target detection method based on YOLOV3-tiny |
Non-Patent Citations (1)
| Title |
|---|
| 刘军: "基于增强Tiny YOLOV3算法的车辆实时检测与跟踪", 《农业工程学报》 * |
Cited By (40)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| TWI723823B (en) * | 2020-03-30 | 2021-04-01 | 聚晶半導體股份有限公司 | Object detection device and object detection method based on neural network |
| CN111709381A (en) * | 2020-06-19 | 2020-09-25 | 桂林电子科技大学 | Road environment target detection method based on YOLOv3-SPP |
| CN111723747A (en) * | 2020-06-22 | 2020-09-29 | 西安工业大学 | A lightweight and high-efficiency target detection method applied to embedded platforms |
| CN111754498A (en) * | 2020-06-29 | 2020-10-09 | 河南科技大学 | A Conveyor Belt Idler Detection Method Based on YOLOv3 |
| CN111814621A (en) * | 2020-06-29 | 2020-10-23 | 中国科学院合肥物质科学研究院 | A multi-scale vehicle pedestrian detection method and device based on attention mechanism |
| CN111754498B (en) * | 2020-06-29 | 2023-11-21 | 河南科技大学 | Conveyor belt carrier roller detection method based on YOLOv3 |
| CN111814621B (en) * | 2020-06-29 | 2024-01-23 | 中国科学院合肥物质科学研究院 | Attention mechanism-based multi-scale vehicle pedestrian detection method and device |
| CN112070713A (en) * | 2020-07-03 | 2020-12-11 | 中山大学 | A Multi-scale Object Detection Method Introducing Attention Mechanism |
| CN111882554A (en) * | 2020-08-06 | 2020-11-03 | 桂林电子科技大学 | An intelligent detection method for power line faults based on SK-YOLOv3 |
| CN111882554B (en) * | 2020-08-06 | 2022-05-03 | 桂林电子科技大学 | An intelligent detection method for power line faults based on SK-YOLOv3 |
| CN112084870A (en) * | 2020-08-10 | 2020-12-15 | 同济大学 | YOLO-based multi-target detection method and device in traffic scene |
| CN111986240A (en) * | 2020-09-01 | 2020-11-24 | 交通运输部水运科学研究所 | Drowning person detection method and system based on visible light and thermal imaging data fusion |
| CN112633308A (en) * | 2020-09-15 | 2021-04-09 | 北京华电天仁电力控制技术有限公司 | Detection method and detection system for whether power plant operating personnel wear safety belts |
| CN112115880A (en) * | 2020-09-21 | 2020-12-22 | 成都数之联科技有限公司 | Ship pollution monitoring method, system, device and medium based on multi-label learning |
| CN112233175B (en) * | 2020-09-24 | 2023-10-24 | 西安交通大学 | Chip positioning method and integrated positioning platform based on YOLOv3-tiny algorithm |
| CN112233175A (en) * | 2020-09-24 | 2021-01-15 | 西安交通大学 | Chip positioning method and integrated positioning platform based on YOLOv3-tiny algorithm |
| CN112183361A (en) * | 2020-09-29 | 2021-01-05 | 中科人工智能创新技术研究院(青岛)有限公司 | Goal detection method and system combining target detection and dynamic difference |
| CN112307921A (en) * | 2020-10-22 | 2021-02-02 | 桂林电子科技大学 | A vehicle-mounted multi-target recognition tracking and prediction method |
| CN112307921B (en) * | 2020-10-22 | 2022-05-17 | 桂林电子科技大学 | Vehicle-mounted end multi-target identification tracking prediction method |
| CN112270252A (en) * | 2020-10-26 | 2021-01-26 | 西安工程大学 | Multi-vehicle target identification method for improving YOLOv2 model |
| CN112396002A (en) * | 2020-11-20 | 2021-02-23 | 重庆邮电大学 | Lightweight remote sensing target detection method based on SE-YOLOv3 |
| CN112396002B (en) * | 2020-11-20 | 2023-05-30 | 重庆邮电大学 | SE-YOLOv 3-based lightweight remote sensing target detection method |
| CN112487969A (en) * | 2020-11-30 | 2021-03-12 | 苏州热工研究院有限公司 | Method for acquiring position of inspection target of inspection robot of steam generator |
| CN112487969B (en) * | 2020-11-30 | 2023-06-30 | 苏州热工研究院有限公司 | A method for obtaining the position of a steam generator inspection robot inspection target |
| CN112561885B (en) * | 2020-12-17 | 2023-04-18 | 中国矿业大学 | YOLOv 4-tiny-based gate valve opening detection method |
| CN112561885A (en) * | 2020-12-17 | 2021-03-26 | 中国矿业大学 | YOLOv 4-tiny-based gate valve opening detection method |
| CN112464910A (en) * | 2020-12-18 | 2021-03-09 | 杭州电子科技大学 | Traffic sign identification method based on YOLO v4-tiny |
| CN112686186A (en) * | 2021-01-05 | 2021-04-20 | 润联软件系统(深圳)有限公司 | High-altitude parabolic recognition method based on deep learning and related components thereof |
| CN112861919A (en) * | 2021-01-15 | 2021-05-28 | 西北工业大学 | Underwater sonar image target detection method based on improved YOLOv3-tiny |
| CN113128555B (en) * | 2021-03-09 | 2022-05-31 | 西南交通大学 | A method for abnormal detection of train brake pad components |
| CN113128555A (en) * | 2021-03-09 | 2021-07-16 | 西南交通大学 | Method for detecting abnormality of train brake pad part |
| CN113537013A (en) * | 2021-07-06 | 2021-10-22 | 哈尔滨理工大学 | Multi-scale self-attention feature fusion pedestrian detection method |
| CN114463772B (en) * | 2022-01-13 | 2022-11-25 | 苏州大学 | Deep learning-based traffic sign detection and identification method and system |
| CN114399440B (en) * | 2022-01-13 | 2022-12-13 | 马上消费金融股份有限公司 | Image processing method, image processing network training method and device and electronic equipment |
| CN114463772A (en) * | 2022-01-13 | 2022-05-10 | 苏州大学 | Traffic sign detection and recognition method and system based on deep learning |
| CN114399440A (en) * | 2022-01-13 | 2022-04-26 | 马上消费金融股份有限公司 | Image processing method, image processing network training method and device and electronic equipment |
| CN114898152A (en) * | 2022-05-13 | 2022-08-12 | 电子科技大学 | Embedded elastic self-expanding general learning framework |
| CN114898152B (en) * | 2022-05-13 | 2023-05-30 | 电子科技大学 | Embedded elastic self-expanding general learning system |
| CN115239949A (en) * | 2022-07-28 | 2022-10-25 | 湖南锐思华创科技有限公司 | Target detection and tracking method, storage medium and device based on yolo neural network |
| CN116030351A (en) * | 2023-03-28 | 2023-04-28 | 南京信息工程大学 | Cascade network-based aerial image ship segmentation method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110929577A (en) | An improved target recognition method based on YOLOv3 lightweight framework | |
| Ma et al. | Automatic detection and counting system for pavement cracks based on PCGAN and YOLO-MF | |
| CN112052787B (en) | Target detection method and device based on artificial intelligence and electronic equipment | |
| CN109978893B (en) | Training method, device, equipment and storage medium of image semantic segmentation network | |
| CN107563372B (en) | License plate positioning method based on deep learning SSD frame | |
| CN112884742B (en) | A multi-target real-time detection, recognition and tracking method based on multi-algorithm fusion | |
| CN112541532B (en) | Target detection method based on dense connection structure | |
| CN108564097A (en) | A kind of multiscale target detection method based on depth convolutional neural networks | |
| CN113971764B (en) | Remote sensing image small target detection method based on improvement YOLOv3 | |
| JP2024513596A (en) | Image processing method and apparatus and computer readable storage medium | |
| CN109886147A (en) | A vehicle multi-attribute detection method based on single-network multi-task learning | |
| CN114627441B (en) | Unstructured road recognition network training method, application method and storage medium | |
| CN109871792B (en) | Pedestrian detection method and device | |
| CN110599459A (en) | Underground pipe network risk assessment cloud system based on deep learning | |
| CN117765409A (en) | Remote sensing graph semantic segmentation method based on multi-scale feature fusion and attention mechanism | |
| CN111881833B (en) | Vehicle detection method, device, equipment and storage medium | |
| CN112634174B (en) | Image representation learning method and system | |
| CN118941526A (en) | A road crack detection method, medium and product | |
| CN111160274A (en) | Pedestrian detection method based on binaryzation fast RCNN (radar cross-correlation neural network) | |
| CN114299328A (en) | Environment self-adaptive sensing small sample endangered animal detection method and system | |
| CN112132839B (en) | Multi-scale rapid face segmentation method based on deep convolution cascade network | |
| Zhao et al. | Recognition and Classification of Concrete Cracks under Strong Interference Based on Convolutional Neural Network. | |
| CN119206659A (en) | YOLOv8-based target detection method, device, storage medium, and equipment in driving scenarios | |
| CN116861262B (en) | Perception model training method and device, electronic equipment and storage medium | |
| CN117975018A (en) | Alignment module, decoder training method, image segmentation method, device and medium |
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 | ||
| AD01 | Patent right deemed abandoned | ||
| AD01 | Patent right deemed abandoned |
Effective date of abandoning: 20231013 |