CN111626175A - Axial type identification method based on deep convolutional neural network - Google Patents

Abstract

The invention relates to an axle type recognition method based on a deep convolutional neural network. The shaft type identification method based on the deep convolutional neural network realizes non-contact identification by processing and identifying the video image collected by the camera arranged in the specific area, avoids damage to the road surface, and has simple equipment and easy maintenance.

Description

Axial type identification method based on deep convolutional neural network

Technical Field

The invention belongs to the technical field of image recognition, and particularly relates to an axis type recognition method based on a deep convolutional neural network.

Background

Axle type identification is the most important part of highway overrun detection. The traditional axle type classification is to identify the axle by a sensor laid on the road surface or a grating erected on the road surface, but the equipment maintenance difficulty and the cost are higher.

For example, shaft type recognition is realized by a dynamic weighing type detection system, the weighing system converts the pressure born by a sensor laid on the road surface into a voltage signal, and judges the number of shafts and the shaft type according to the output waveform, but the realization of the method needs to lay the sensor on the road surface, the equipment maintenance is not easy, and the damage can be caused to the road surface;

meanwhile, with rapid development of video image technology and artificial intelligence, vehicle type recognition based on video image processing and deep learning becomes a mainstream research direction more and more. Particularly, the deep convolutional neural network can extract deep features from the color image by virtue of the structural characteristics of local receptive fields, shared weights and downsampling, and has great superiority in image object classification.

Therefore, a new shaft type identification method is required.

Disclosure of Invention

The invention relates to a shaft type identification method based on a deep convolutional neural network, which solves the problems that in the prior art, a sensor paved on a road surface or an erected grating is used for identifying a wheel shaft, and the method is high in maintenance difficulty and high in cost.

In order to solve the problems, the technical scheme of the invention is as follows:

the axis type identification method based on the deep convolutional neural network comprises the following steps:

step 1, collecting a large number of vehicle side images, drawing a minimum rectangular frame containing an axle for the vehicle images, and cutting to form an axle detection image set;

step 2, training a Fast-RCNN axle detection model by using the axle image as a positive sample and the non-axle image as a negative sample, and outputting the trained axle detection model;

step 3, labeling the category labels of the corresponding axle types to the axle region pictures of various types to form an axle type identification image set;

step 4, training a deep convolutional neural network for axis type recognition by using the axis type recognition image set, and outputting a trained network model;

and 5, sequentially passing the vehicle image to be detected through the axle detection model and the axle type identification network model, and outputting the corresponding axle type category.

Further, in the step 1, the cut axle region picture is subjected to random small-angle rotation fine adjustment to generate more data sets.

Further, in step 1, the image acquisition device is an industrial camera, the industrial camera is mounted on a support on the side of the road or the detection station, and the shot vehicle image is a color image.

Further, in the step 4, the trained network model includes a convolution layer, a maximum pooling layer, a Relu activation function layer, a full connection layer, and a Softmax loss output layer, which are connected in sequence;

the convolution layer is used for extracting the characteristics of the axle image;

the maximum pooling layer is used for reducing the dimension of the characteristics of the convolutional layer;

the full-connection layer is used for carrying out feature fusion on each feature extracted by the convolution layer and the maximum pooling layer to obtain an overall feature, and then carrying out feature expansion;

and the Softmax loss output layer carries out final classification and normalization.

Further, in step 4, the trained network model is 14 layers, and the distribution thereof is specifically:

a convolution first layer comprising 8 convolution kernels, each convolution kernel being 3 x 3 in size;

a batch standardization first layer, which makes network training easier to optimize through activation and gradient of network propagation to accelerate network training and reduce sensitivity to network initialization;

exciting the first layer and adopting a ReLU activation function;

pooling a first layer with a size of 2 × 2, cutting the feature map output by the previous layer into 2 × 2 small blocks, and taking the maximum feature value in each small block as the output feature of the small block;

a second layer of convolution comprising 16 convolution kernels, each convolution kernel being 3 x 3 in size;

a batch normalization second layer, identical to the batch normalization first layer;

exciting the second layer, the same as exciting the first layer;

pooling a second layer with the size of 2 multiplied by 2, cutting a feature map for exciting the output of the second layer into small blocks of 2 multiplied by 2, and taking the maximum feature value in each small block as the output feature of the small block;

a third layer of convolution comprising 32 convolution kernels, the convolution kernels having a size of 3 x 3;

a batch normalized third layer, identical to the batch normalized second layer;

exciting the third layer, the same as exciting the second layer;

the output class number of the full connection layer is M, and represents the M class axis type in the system;

a Softmax loss output layer, which normalizes the output of the fully-connected layer for use as a classification probability of the classification layer;

and a classification layer, which uses the probability returned by the Softmax activation function for each input, assigns the input to one of the mutually exclusive classes and calculates the loss.

The invention has the following beneficial effects:

1. the shaft type identification method based on the deep convolutional neural network realizes non-contact identification by processing and identifying the video image collected by the camera arranged in the specific area, avoids damage to the road surface, and has simple equipment and easy maintenance.

2. The method utilizes the multiscale invariance of the deep convolutional neural network to perform operations such as turning, scale transformation, rotation and the like on the segmented axle region image in the sample production, so that a large amount of effective training data can be generated, a vehicle type recognition model with high speed and precision can be trained, the robustness of a vehicle type recognition system is enhanced, and the problem that the recognition result is influenced by the vehicle contour is effectively solved.

Drawings

Fig. 1 is a block diagram of the overall design of the present invention.

Fig. 2 is a flowchart of the axle detection according to the present invention.

FIG. 3 is a schematic diagram of the Fast-RCNN axle detection network structure of the present invention.

FIG. 4 is a flow chart of the shaft type identification of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the following examples, it being understood that the described examples are only a part of the examples of the present invention, and not all examples. 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.

Example (b):

FIG. 1 shows the overall design block diagram of the present invention, after training the network with the training set data to achieve the goal, testing the network with the test set data, and finally achieving the purpose of axis type identification.

Fig. 2 shows a flow chart of the axle detection model establishment of the present invention, and the detection process of the present invention includes:

step a1, image acquisition. The acquisition equipment is a common industrial-grade camera and respectively captures images of the side surface of the vehicle including the axle area. Typically cameras are mounted on the side of the road or inspection station to ensure that the most complete image information of the vehicle is obtained.

And step A2, making positive and negative samples. And determining an axle area through manual cutting or Hough circle detection, and using the axle area as a positive sample for training an axle detection model. Negative examples are background images that do not contain axles, such as roads, buildings, other types of vehicles, billboards, etc. In one image, the non-axle region takes up a considerable weight. Therefore, the number of negative samples to be excluded in axle detection is usually much larger than the number of positive samples to be detected, and the detection accuracy is often affected by the negative samples. In order to improve the detection accuracy, the non-axle area sample collected by the invention is used as a negative sample.

And step A3, detecting a network model based on Fast-RCNN axles. The input to the detection network is a whole picture and a series of candidate boxes that target axle areas where vehicles may be present, waiting for the network to verify whether axles are present in these candidate areas. The Fast-RCNN of the present invention uses the RO I layer, as shown in FIG. 3, and the input is the whole picture, and the red box represents a candidate box in the picture. The neural network first maps the entire image into a series of convolution signatures through a series of convolution layer operations and a maximum pooling layer operation. Then, the system determines the ROI mapping region of the candidate frame in the convolution feature map according to the position proportion of the candidate frame and the original image, wherein the ROI mapping region is called as an ROI layer, and the ROI layer is scaled to a fixed size through an RO I pooling operation. Finally, the second half of the network converts the ROI pooled data into a ROI feature vector through the fully connected layer, which characterizes all features of the image candidate region. And respectively converting the data into Softmax output and classification layer output through two different full connection layers.

After the axle region is determined, the axle image is sent to an axle type recognition system based on a deep convolutional neural network, as shown in fig. 4, which is an axle type recognition flowchart.

Step B1, normalizing 5000 images for training in the axle type recognition database to the same pixel size of 100 x 50 pixels, and classifying and labeling the corresponding vehicle axle types;

and step B2, sequentially sending the processed images into a deep convolutional neural network, passing through two convolutional layers, a pooling layer and a full-link layer, obtaining the fused image characteristics at the last full-link layer, and comparing the fused image characteristics with the corresponding labeled images to obtain the prediction error. The convolution kernels of the C1 and C2 layers were 5 × 5 in size, the step size was 1, and the C1 layer contained 8 signatures 96 × 46 in size; step size 2 was obtained by pooling layers with pooling area 2 x 2, resulting in a signature containing 8S 1 layers of size 48 x 23. The C2 layer contained 16 signatures of 44 x 19 size. Similarly, the step size is 1, the number of the feature maps is unchanged, and the size of the feature maps is reduced to 43 × 18 through the pooling layer S2. Convolution operation is performed on 32 convolution kernels with the size of 3 × 3 and the step size of 1 and S2 to obtain a convolution layer C3 which comprises a characteristic diagram of 41 × 16. After passing through the pooling layer S3, 32 41 × 16 characteristic maps were finally obtained. Finally, vectorizing the characteristic diagrams obtained by the S2 and S3 layers to jointly form a final characteristic vector;

b3, reducing the prediction error by adopting a back propagation algorithm and a random gradient descent method, and training the neural network until the error does not descend any more to obtain a training model for recognizing the axis type;

and step B4, sending the vehicle image into the axle detection model trained in the step A3 for axle detection positioning, sending the normalized axle image into an axle type recognition model, and outputting an axle type recognition result.

The present invention has been described in terms of specific examples, which are provided to aid understanding of the invention and are not intended to be limiting. Any partial modification or replacement within the technical scope of the present disclosure by a person skilled in the art should be included in the scope of the present disclosure.

Claims (5)

1. An axis type identification method based on a deep convolutional neural network is characterized in that,

the method comprises the following steps:

step 1, collecting a large number of vehicle side images, drawing a minimum rectangular frame containing an axle for the vehicle images, and cutting to form an axle detection image set;

step 2, training a Fast-RCNN axle detection model by using the axle image as a positive sample and the non-axle image as a negative sample, and outputting the trained axle detection model;

step 3, labeling the category labels of the corresponding axle types to the axle region pictures of various types to form an axle type identification image set;

step 4, training a deep convolutional neural network for axis type recognition by using the axis type recognition image set, and outputting a trained network model;

and 5, sequentially passing the vehicle image to be detected through the axle detection model and the axle type identification network model, and outputting the corresponding axle type category.

2. The method for identifying the axle type of the local depth convolutional neural network as claimed in claim 1, wherein in the step 1, the trimmed axle region picture is subjected to random small-angle rotation fine tuning to generate more data sets.

3. The axis type identification method based on the deep convolutional neural network of claim 1 or 2, wherein in the step 1, the image acquisition device is an industrial camera, the industrial camera is installed on a support at the side of a road or a detection station, and the image of the vehicle is a color image.

4. The axis type identification method based on the deep convolutional neural network as claimed in claim 3, wherein in the step 4, the trained network model comprises a convolutional layer, a max pooling layer, a Relu activation function layer, a full connection layer, and a Softmax loss output layer which are connected in sequence;

the convolution layer is used for extracting the characteristics of the axle image;

the maximum pooling layer is used for reducing the dimension of the characteristics of the convolutional layer;

the full-connection layer is used for carrying out feature fusion on each feature extracted by the convolution layer and the maximum pooling layer to obtain an overall feature, and then carrying out feature expansion;

and the Softmax loss output layer carries out final classification and normalization.

5. The axis type identification method based on the deep convolutional neural network as claimed in claim 4, wherein in the step 4, the trained network model is 14 layers, and the distribution is specifically as follows:

a convolution first layer comprising 8 convolution kernels, each convolution kernel being 3 x 3 in size;

a batch standardization first layer, which makes network training easier to optimize through activation and gradient of network propagation to accelerate network training and reduce sensitivity to network initialization;

exciting the first layer and adopting a ReLU activation function;

pooling a first layer with a size of 2 × 2, cutting the feature map output by the previous layer into 2 × 2 small blocks, and taking the maximum feature value in each small block as the output feature of the small block;

a second layer of convolution comprising 16 convolution kernels, each convolution kernel being 3 x 3 in size;

a batch normalization second layer, identical to the batch normalization first layer;

exciting the second layer, the same as exciting the first layer;

pooling a second layer with the size of 2 multiplied by 2, cutting a feature map for exciting the output of the second layer into small blocks of 2 multiplied by 2, and taking the maximum feature value in each small block as the output feature of the small block;

a third layer of convolution comprising 32 convolution kernels, the convolution kernels having a size of 3 x 3;

a batch normalized third layer, identical to the batch normalized second layer;

exciting the third layer, the same as exciting the second layer;

the output class number of the full connection layer is M, and represents the M class axis type in the system;

a Softmax loss output layer, which normalizes the output of the fully-connected layer for use as a classification probability of the classification layer;

and a classification layer, which uses the probability returned by the Softmax activation function for each input, assigns the input to one of the mutually exclusive classes and calculates the loss.

Images