A method for positioning and speed measurement of urban rail trains based on machine vision

Abstract

The present invention discloses a method for positioning and measuring the speed of urban rail trains based on machine vision. The method uses a high-frame-rate industrial camera to shoot the track line and the visual beacon set at a fixed point, and uses an improved YOLO algorithm to process the image in real time to obtain the sleeper count, thereby realizing the positioning and speed measurement of the urban rail train. The present invention realizes the real-time processing of the captured high-frame-rate image, continuously tracks the position information and speed information of the train, and the positioning accuracy is less than 0.1 meters, and the speed measurement accuracy reaches 0.1km / h, which can realize the sleeper count, positioning and speed measurement of the urban rail train when it is running at 120km / h.

Description

Technical Field

[0001] The invention belongs to the field of rail transportation. Background Art

[0002] In recent years, my country has continuously increased its investment in railway passenger dedicated lines and urban rail transit, and the status of rail transit has been increasingly improved. The location information of rail vehicle operation is the basic data for the operation, detection and maintenance of rail transit vehicles, and its research is becoming increasingly important. With the rise of artificial intelligence technology, the application of new technologies such as machine vision, fault detection and diagnosis, and unmanned driving in the rail transit industry has flourished. In order to ensure the safe operation of trains and better serve vehicle marshaling, scheduling, detection, maintenance and other tasks, it is of great theoretical and practical significance to study an independent device with high-precision, real-time, continuous and autonomous positioning on board.

[0003] The basic function of train operation positioning technology is to automatically obtain the position information, speed information and section occupancy information of the train on the track through sensing technology. It is generally believed that the train moves in one dimension along the radial direction of the track. By measuring the distance between the train and the reference point of the track, the train can be positioned and tracked.

[0004] The current train positioning technology mainly adopts the following methods, each of which has certain advantages and disadvantages and cannot meet the requirements of high-precision, real-time, continuous, and autonomous positioning and tracking.

[0005] (1) Train positioning method based on speed measurement

[0006] This method belongs to autonomous inertial navigation by continuously acquiring the real-time running speed of the train and integrating it to get the mileage of the train. Currently, there are two main speed measurement methods commonly used, one is the wheel speed method and the other is the Doppler speed meter.

[0007] When the train is running, the wheelset rotates and drives the encoder to rotate. The wheel speed method calculates the speed and mileage by measuring the encoder pulse frequency and pulse count. The working principle is simple and easy to implement. The calculation is based on the pure rolling motion between the wheelset and the track and the geometric dimensions of the wheelset. The wheel speed method has obvious defects in speed measurement: [1] When the wheelset idles, creeps, or slips during the operation of the train (for example, when braking), large measurement errors will occur; [2] Wear will change the geometric dimensions of the wheelset, and measuring the train mileage based on the geometric dimensions of the wheelset will result in large measurement errors; [3] The angle of rotation of the wheelset is too small, resulting in the inability to generate pulses, resulting in cumulative errors.

[0008] The Doppler speed meter is installed at the bottom of the train. The direction and speed of the train are obtained by measuring the frequency changes of the transmitted and reflected signals, and the train's mileage is then obtained by integration. This method is based on independent speed measuring equipment and does not involve the movement and wear of the wheelset, but it also has obvious disadvantages: [1] Doppler speed measurement has high requirements for the environment. If the environment is not flat, it will cause radio wave scattering, resulting in large measurement errors; [2] Doppler speed meters are usually installed in a top-down manner, and the measurement errors caused by the installation angle need to be corrected; [3] Obtaining the train's mileage through integration will cause a large cumulative error; [4] Experiments show that the Doppler speed meter has a large error when the train is running at low speed.

[0009] (2) Train positioning method based on query transponder

[0010] The train’s point positioning is achieved by reading the built-in location information of the ground transponder (beacon) through the on-board interrogator. Its advantages are mature technology (widely used in urban rail transit), high positioning accuracy, and the ability to correctly distinguish different tracks on double-track tracks, which is particularly suitable for eliminating the cumulative errors generated during train positioning. The train positioning method based on the query transponder also has obvious defects: [1] It can only perform point positioning and cannot achieve real-time and continuous positioning; [2] The cost of densely deploying transponders along the track is too high, and when the transponder is located in an unmanned area, it is difficult to maintain and maintain. Usually, transponders are set up near urban rail stations, and two sets of transponders may be more than a few kilometers apart.

[0011] (3) Train positioning method based on track circuit

[0012] Axle counters are set at the start and end of each section of the track circuit. By obtaining the status of the axle counters, the track circuit section the train is in can be determined. The advantages of this method are mature technology and strong reliability, but the positioning error is large and it is impossible to accurately give the specific position and speed information of the train in the section in real time.

[0013] (4) Train positioning method based on GPS / BDS

[0014] The principle of satellite positioning is simple and can work around the clock. However, it is easily affected by the natural environment. Trains running in mountainous areas or tunnels have positioning blind spots, and the current positioning accuracy cannot distinguish between double-track tracks. Urban rail transit mainly runs in underground tunnels, and the accuracy, reliability and availability of satellite navigation are questioned. Summary of the invention

[0015] Purpose of the invention: To solve the problems existing in the above-mentioned background technology, the present invention provides a method for positioning and measuring the speed of urban rail trains based on machine vision.

[0016] Technical solution: The present invention provides a method for positioning and measuring the speed of an urban rail train based on machine vision, which specifically includes the following steps:

[0017] Step 1: Setting visual beacons on the track line near the subway station, and performing graphic coding on each visual beacon, wherein each graphic coding includes accurate location information of the corresponding visual beacon;

[0018] Step 2: Install an image acquisition integrated machine at the bottom of the train, the image acquisition integrated machine collects track images when the train is running in real time, and transmits them to the in-vehicle visual processing system;

[0019] Step 3: The train is tested on the line, and the image acquisition integrated device collects the track line images in real time during operation; the in-vehicle visual processing system pre-processes the collected images, and uses the processed images as training samples, and uses the improved YOLO algorithm for training to obtain the weight file for sleeper target detection;

[0020] Step 4: When the train is officially running, the image acquisition integrated machine collects track line images in real time, and uses the improved YOLO algorithm and the weight file in step 3 to process the images in real time, identify the sleepers and train travel direction in the images, and count the sleepers;

[0021] Step 5: Calculate the position and speed of the train by interpolation based on the number of identified continuous sleepers and the spacing between adjacent sleepers; when the acquired image contains a visual beacon, calculate the deviation between the current train position information obtained through the sleepers and the accurate position information in the graphic code of the visual beacon, correct the train position at this time, and eliminate the accumulated error of the system.

[0022] Furthermore, the visual beacon is a passive visual beacon, which uses an aluminum alloy panel to spray-paint a barcode, and the beacon is installed on a track line close to a subway station through a heavy base.

[0023] Furthermore, the image acquisition integrated machine includes: an upper L-shaped plate, a lower L-shaped plate, a connecting plate, and an LED light source bracket; the upper L-shaped plate is fixedly installed on the bottom of the train, one end of the lower L-shaped plate is fixedly connected to the upper L-shaped plate, and the other end is fixedly connected to one side of the connecting plate, a high-speed industrial camera is arranged at the center of the other side of the connecting plate, and the LED light source bracket is arranged around the high-speed industrial camera and fixedly connected to the connecting plate.

[0024] Furthermore, the image acquisition integrated machine also includes a protective cover, which has a camera obstruction alarm function; the protective cover is a cylindrical protective cover, the column body of the cylindrical protective cover is made of aluminum alloy, and a dust-proof glass is provided at the bottom. The protective cover wraps the high-speed industrial camera and the LED light source bracket, and is fixedly connected to the connecting plate; the protective cover determines whether there are foreign objects or dust and fog blocking the dust-proof glass through the image taken by the camera, and if so, an alarm is issued.

[0025] Furthermore, the improved YOLO algorithm network structure is 21 layers, including a convolution layer, a maximum pooling layer, a pass-through layer and an output layer; the output layer outputs the offset of the center position of the sleeper, the height and width offset, the confidence and the category probability; the convolution layer includes convolution, batch normalization and activation function, and the activation function adopts the Leaky-ReLU function, which is expressed as follows:

[0026] f(x)=max(ax,x)

[0027] Among them: a is a constant, x is the input quantity, and max(.) is the maximum value function.

[0028] Furthermore, the loss function L(O, o, C, c, l, g) used in the training of the improved YOLO algorithm in step 3 includes the target positioning offset loss function L loc (l, g), target confidence loss function L conf (o, c) and target classification loss function L cla (O, C);

[0029] L(O,o,C,c,l,g)=λ1L conf (o,c)+λ2L cla (O,C)+λ3L loc (l, g)

[0030] Among them, λ1, λ2 and λ3 are balance coefficients, and the target positioning offset loss function L loc (l, g) uses the sum of the squares of the differences between the true value and the predicted value, expressed as:

[0031]

[0032]

[0033]

[0034]

[0035] Among them, pos is all the real boxes, are the horizontal and vertical coordinates and width and height of the center of the i-th real frame respectively; are the horizontal and vertical coordinates and width and height of the center of the i-th prediction box, respectively, t x ,t y ,t w ,t h They are the predicted offsets of the center horizontal and vertical coordinates and the predicted offsets of width and height, respectively. are the horizontal and vertical coordinates and width and height of the center of the i-th real box, respectively. i ,y i 、w i 、h i are the horizontal and vertical coordinates and width and height of the center of the i-th prediction box, respectively, and σ is the Sigmiod function;

[0036] Target confidence loss function L conf The expression of (o, c) is:

[0037]

[0038]

[0039] o i ∈{0, 1} indicates whether the target actually exists in the predicted target bounding box i, 0 means no target exists, 1 means yes target exists. Obtained through the Sigmiod function, it indicates the probability of predicting whether there is a target in the target rectangular box i;

[0040] Target classification loss function L cla The expression of (O, C) is:

[0041]

[0042]

[0043] Where Pos is the bounding box of all predicted targets, cla is the Sigmoid probability of all target classes, O ij ∈{0, 1}, indicating whether the j-th target actually exists in the predicted target bounding box i, 0 means no target exists, 1 means yes target exists. The Sigmoid function is used to obtain the probability that the network predicts that there is a target of type j in the target bounding box i.

[0044] Beneficial effects: The present invention combines image feature extraction with convolutional neural networks for rapid detection, identification, positioning and tracking of sleepers; the present invention can process track line images with a frame rate of 300fps in real time, and the sleeper identification accuracy reaches 100%, and can realize sleeper counting, train positioning and speed measurement when urban rail trains run at 120km / h. The present invention can continuously update the position information and speed information of the train in a time period of 0.02 seconds, with a comprehensive positioning accuracy of less than 0.1 meter and an interval speed measurement accuracy of 0.10km / h. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a flow chart of the present invention;

[0046] Figure 2 : is a structural diagram of an image acquisition integrated machine; wherein (a) is a bottom view of the image acquisition integrated machine, and (b) is a top view of the image acquisition integrated machine;

[0047] Figure 3 The structure diagram of the protective cover; (a) is the front view of the protective cover, and (b) is the bottom view of the protective cover

[0048] Figure 4 This is the network structure diagram of the improved YOLO algorithm;

[0049] Figure 5 Schematic diagram of neurons and activation functions;

[0050] Figure 6 This is a comparison chart of the effects of the network model used in the present invention and the yolov3 model on sleeper detection. DETAILED DESCRIPTION

[0051] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0052] like Figure 1 As shown, this embodiment provides a method for locating and measuring the speed of urban rail trains based on machine vision, which uses a high-frame-rate industrial camera to shoot the track line and set visual beacons at fixed points, and uses an improved YOLO algorithm to process the image in real time to obtain the sleeper count, thereby realizing the positioning and speed measurement of urban rail trains. Specifically:

[0053] Step 1: Set up visual beacons on the track line near the subway station, and perform graphic coding on each visual beacon. Each graphic coding includes the accurate location information of the corresponding visual beacon. The accurate location information is the precise location information of the visual beacon, and the location information is accurate.

[0054] Step 2: Install the image acquisition integrated device at the bottom of the train to collect track images in real time while the train is running, and transmit them to the on-board visual processing system.

[0055] Step 3: The train is tested on the line, and the image acquisition integrated device collects the track line images in real time during operation; the in-vehicle visual processing system pre-processes the collected images and uses the processed images as training samples for deep learning. The improved YOLO algorithm is used for training to obtain the weight file for sleeper target detection.

[0056] Step 4: When the train is officially running, the image acquisition integrated machine collects track line images in real time, and uses the improved YOLO algorithm and the weight file in step 3 to process the images in real time, identify the sleepers and train travel direction in the images, and count the sleepers.

[0057] Step 5: Calculate the position and speed of the train by interpolation based on the number of consecutive sleepers identified and the spacing between adjacent sleepers. When the acquired image contains a visual beacon, calculate the deviation between the train position obtained through the sleeper and the position information in the graphic code of the visual beacon, correct the train position at this time, and eliminate the accumulated error of the system.

[0058] In this embodiment, the visual beacon is a passive beacon, which uses an aluminum alloy panel to spray-paint a barcode. The visual beacon is installed on a heavy base and is set on a track line near a subway station.

[0059] In this embodiment, if Figure 2 As shown in Figures (a) and (b), the image acquisition integrated machine includes: an upper L-shaped plate, a lower L-shaped plate, a connecting plate, and an LED light source bracket; the upper L-shaped plate is fixedly installed at the bottom of the train, one end of the lower L-shaped plate is fixedly connected to the upper L-shaped plate, and the other end is fixedly connected to one side of the connecting plate, a high-speed industrial camera is arranged at the center of the other side of the connecting plate, and the LED light source bracket is arranged around the high-speed industrial camera and fixedly connected to the connecting plate. The high-speed industrial camera can capture track line images at a high frame rate and transmit them accurately and reliably to the in-vehicle visual processing system.

[0060] In this embodiment, if Figure 3 As shown, the image acquisition integrated machine includes a protective cover and a camera occlusion alarm. The protective cover is a cylindrical protective cover, the body of the cylindrical protective cover is made of aluminum alloy, and a dust-proof glass is provided at the bottom. The protective cover wraps the high-speed industrial camera and the LED light source bracket, and is fixedly connected to the connecting plate; the camera occlusion alarm determines whether there is foreign matter or dust and fog blocking the dust-proof glass through the image taken by the camera. The protective cover is waterproof, dustproof, and fog-proof, and can be washed with water to prevent the lens from being blocked by dust and fog. The protective cover is stimulated by the aerodynamics of the train when it is running, and a noise reduction design is performed;

[0061] In this embodiment, the target detection algorithm improves the YOLO (You only look once) algorithm to improve the ability of real-time image processing based on the characteristics of fixed camera installation position and focal length, unchanged image size and field of view, and relatively fixed target scale.

[0062] like Figure 4As shown, the improved YOLO algorithm network structure is 21 layers. In this embodiment, the size of the image collected is fixed to 720×540 and is a single-channel image, so 720×540×1 is used as the input of the network, and the output is a 4692-dimensional feature vector. The improved YOLO target detection algorithm includes a convolutional layer, a maximum pooling layer, a pass-through layer, and an output layer. The convolutional layer includes convolution, batch normalization, and activation function.

[0063] The padding and stride of the convolution layer are set to 1 by default. This layer mainly realizes feature extraction. The BN layer is an algorithm often used in deep neural network training. Its main function is to normalize the data, achieve rapid convergence, and increase the training speed. The essence of network training is to learn data distribution. Once the distribution of training data and test data is different, the generalization ability of the network will be greatly reduced. If the data is not normalized, it is easy to cause different distributions of each batch of data. The network must adapt to different distributions at each iteration, which greatly reduces the training speed of the network. In the process of training the network, the output data of each layer is constantly changing, so it is usually necessary to add a BN layer after the convolution layer. Normalizing the image can also reduce the impact of external environment such as lighting and contrast.

[0064] Leaky-ReLU activation function

[0065] like Figure 5 As shown in the figure, each neuron node in the neural network accepts the output value of the previous layer of neurons as the input value of this neuron, and passes the output value of this neuron to the next layer. By introducing nonlinear activation functions, the output of deep neural networks is no longer a linear combination of inputs, but can almost approximate any function.

[0066] Common activation functions include: sigmoid function, tanh function, Relu function, etc. In this embodiment, an improved Leaky-Relu function is used, and the expression is f(x)=max(ax, x) (a is a constant, x is an input, max(.) is a maximum value function, usually a is very small, and the common value is 0.01); the Leaky-Relu function retains the advantage that the ReLU function converges much faster than sigmoid and tanh, and avoids the Dead ReLU problem to a certain extent.

[0067] Max Pooling Layer

[0068] For an image of a certain size area, the maximum eigenvalue is taken to replace the original image area. Pooling can increase the field of view; the pooling operation continuously abstracts regional features without caring about the position, which increases translation invariance to a certain extent; pooling calculation is convenient and can greatly reduce the number of model parameters and optimization workload.

[0069] Pass-through layer

[0070] Drawing on the ResNet network, the high-resolution feature map is subjected to a PassThrough operation, that is, the width and height are halved and the depth is quadrupled, and then connected with the low-resolution feature map to obtain a new feature map, thereby reducing the possibility of gradient vanishing, increasing the network training speed, and retaining some features of the high-resolution feature map.

[0071] Output Layer

[0072] The output layer is a 1x1x4692 convolutional layer connected to 23x17x1024 (excluding BN layer and activation function), where 4692 = (5+1) × 2 × 23 × 17, that is, the final 23x17 feature map. Each unit predicts two AnchorBoxes, and each Anchor Box contains position offset, width and height offset, confidence, and category probability.

[0073] In this implementation, the step 3 uses an improved YOLO target detection algorithm network structure for supervised learning. The loss function L(O, o, C, c, l, g) used in the training of the improved YOLO algorithm includes the target positioning offset loss function L loc (l, g), target confidence loss function L conf (o, c) and target classification loss function L cla (O, C);

[0074] L(O,o,C,c,l,g)=λ1L conf (o,c)+λ2L cla (O,C)+λ3L loc (l, g)

[0075] Where λ1, λ2, and λ3 are balance coefficients, and the target positioning offset loss function L loc (l, g) uses the sum of the squares of the differences between the true value and the predicted value, expressed as:

[0076]

[0077]

[0078]

[0079]

[0080] Among them, pos is all the real boxes, are the horizontal and vertical coordinates and width and height of the center of the i-th real frame respectively; are the horizontal and vertical coordinates and width and height of the center of the i-th prediction box, respectively, t x ,t y ,t w ,t h They are the predicted offsets of the center horizontal and vertical coordinates and the predicted offsets of width and height, respectively. are the horizontal and vertical coordinates and width and height of the center of the i-th real box, respectively. i ,y i 、w i 、h i are the horizontal and vertical coordinates and width and height of the center of the i-th prediction box, respectively, and σ is the Sigmiod function;

[0081] Target confidence loss function L conf The expression of (o, c) is:

[0082]

[0083]

[0084] o i ∈{0, 1} indicates whether the target actually exists in the predicted target bounding box i, 0 means no target exists, 1 means yes target exists. Obtained through the Sigmiod function, it indicates the probability of predicting whether there is a target in the target rectangular box i;

[0085] Target classification loss function L cla The expression of (O, C) is:

[0086]

[0087]

[0088] Where Pos is the bounding box of all predicted targets, cla is the Sigmoid probability of all target classes, O ij ∈{0, 1}, indicating whether the j-th target actually exists in the predicted target bounding box i, 0 means no target exists, 1 means yes target exists. The Sigmoid function is used to obtain the probability that the network predicts that there is a target of type j in the target bounding box i.

[0089] like Figure 6 As shown, compared with the yolov3 model, the method of this embodiment has a better detection effect on the sleeper target.

[0090] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of the present invention.

Claims

1. A method for positioning and measuring speed of urban rail trains based on machine vision, characterized in that: The specific steps include: Step 1: visual beacons are set on the track line near the subway station, and each visual beacon is graphically encoded, each graphical code includes accurate location information of the corresponding visual beacon, and the visual beacon is a passive visual beacon; Step 2: Install an image acquisition integrated machine at the bottom of the train, the image acquisition integrated machine collects track images when the train is running in real time, and transmits them to the visual processing system in the train; Step 3: The train is tested on the line, and the image acquisition integrated device collects the track line images of the train in real time; the in-vehicle visual processing system pre-processes the received images, and uses the processed images as training samples, and uses the improved YOLO algorithm for training to obtain the weight file for sleeper target detection; Step 4: The train is officially running. The image acquisition integrated machine collects track line images in real time. The in-vehicle visual processing system pre-processes the received images and uses the improved YOLO algorithm and the weight file in step 3 to process the images in real time, identify the sleepers and train travel direction in the images, and count the sleepers. Step 5: According to the number of identified continuous sleepers and the spacing between adjacent sleepers, the position and speed of the train are calculated by interpolation; when the acquired image contains a visual beacon, the deviation between the current train position information obtained through the sleeper and the accurate position information in the graphic code of the visual beacon is calculated, and the position of the train at this time is corrected, thereby eliminating the accumulated error of the system; The loss function L(O,o,C,c,l,g) used by the improved YOLO algorithm in step 3 during training includes the target positioning offset loss function L loc (l,g), target confidence loss function L conf (o,c) and target classification loss function L cla (O,C); L(O,o,C,c,l,g)=λ1L conf (o,c)+λ2L cla (O,C)+λ3L loc (l,g) Among them, λ1, λ2 and λ3 are balance coefficients, and the target positioning offset loss function L loc (l,g) uses the sum of the squares of the differences between the true value and the predicted value, expressed as: Among them, pos is all the real boxes, are the horizontal and vertical coordinates and width and height of the center of the i-th real frame respectively; are the horizontal and vertical coordinates and width and height of the center of the i-th prediction box, respectively, t x ,t y ,t w ,t h They are the predicted offsets of the center horizontal and vertical coordinates and the predicted offsets of width and height, respectively. are the horizontal and vertical coordinates and width and height of the center of the i-th real box, respectively. i ,y i 、w i 、h i are the horizontal and vertical coordinates and width and height of the center of the i-th prediction box, respectively, and σ is the Sigmiod function; Target confidence loss function L conf The expression of (o,c) is: o i ∈{0,1} indicates whether there is a real target in the predicted target bounding box i, 0 means no target, 1 means yes target. Obtained through the Sigmiod function, it indicates the probability of predicting whether there is a target in the target rectangular box i; Target classification loss function L cla The expression of (O,C) is: Where Pos is the bounding box of all predicted targets, cla is the Sigmoid probability of all target classes, O ij ∈{0,1}, indicates whether the j-th target actually exists in the predicted target bounding box i, 0 means no target exists, 1 means yes target exists. The Sigmoid function is used to obtain the probability that the network predicts that there is a target of type j in the target bounding box i.

2. The method for positioning and measuring speed of urban rail trains based on machine vision according to claim 1, characterized in that: The visual beacon uses an aluminum alloy panel with a spray-painted barcode. The beacon is installed on the track line near the subway station via a heavy base.

3. The method for positioning and measuring speed of urban rail trains based on machine vision according to claim 1, characterized in that: The image acquisition integrated machine includes: an upper L-shaped plate, a lower L-shaped plate, a connecting plate, and an LED light source bracket; the upper L-shaped plate is fixedly installed on the bottom of the train, one end of the lower L-shaped plate is fixedly connected to the upper L-shaped plate, and the other end is fixedly connected to one side of the connecting plate, a high-speed industrial camera is arranged at the center of the other side of the connecting plate, and the LED light source bracket is arranged around the high-speed industrial camera and fixedly connected to the connecting plate.

4. The method for positioning and measuring speed of urban rail trains based on machine vision according to claim 3, characterized in that: The image acquisition integrated machine also includes a protective cover, which has a camera obstruction alarm function; the protective cover is a cylindrical protective cover, the column body of the cylindrical protective cover is made of aluminum alloy, and a dust-proof glass is provided at the bottom. The protective cover wraps the high-speed industrial camera and the LED light source bracket, and is fixedly connected to the connecting plate; the protective cover determines whether there is foreign matter or dust and fog obstructing the dust-proof glass through the image taken by the camera, and if so, an alarm is issued.

5. The method for positioning and measuring speed of urban rail trains based on machine vision according to claim 1, characterized in that: The improved YOLO target detection algorithm network structure is 21 layers, including convolution layer, maximum pooling layer, pass-through layer and output layer; the output layer outputs the offset of the center position of the sleeper, the height and width offset, the confidence and the category probability; the convolution layer includes convolution, batch normalization and activation function, and the activation function adopts the Leaky-ReLU function, and its expression f(x) is: f(x)=max(ax,x) Among them: a is a constant, x is the input quantity, and max(.) is the maximum value function.

Images