CN109858497B - Improved residual error network and feature extraction method and device thereof - Google Patents
Improved residual error network and feature extraction method and device thereof Download PDFInfo
- Publication number
- CN109858497B CN109858497B CN201910049298.8A CN201910049298A CN109858497B CN 109858497 B CN109858497 B CN 109858497B CN 201910049298 A CN201910049298 A CN 201910049298A CN 109858497 B CN109858497 B CN 109858497B
- Authority
- CN
- China
- Prior art keywords
- feature extraction
- layer
- extraction block
- block
- convolution
- 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.)
- Active
Links
Abstract
The application discloses an improved residual error network and a feature extraction method and device thereof. The improved residual error network comprises a preprocessing block, a feature extraction block and a Softmax layer, wherein a BN layer is arranged in the feature extraction block to accelerate the learning speed of the network, and a Dropout layer is arranged in the third feature extraction block and the fourth feature extraction block to discard the neural network unit in the extraction. According to the application, the learning rate of the network is accelerated through the BN layer, the extraction speed is ensured, the complexity of the network is reduced through the Dropout layer, the feature transition fitting phenomenon when the feature extraction depth is deeper is effectively avoided, the extraction speed is ensured, and the extraction accuracy is improved.
Description
Technical Field
The application relates to the field of neural networks, in particular to an improved residual error network and a feature extraction method and device thereof.
Background
Currently, with the development of unmanned technologies, advanced driving assistance systems are an important direction of research. In advanced driving assistance systems, early warning of traffic signs during driving is an important part of ensuring driving safety and legitimacy. In order to realize the recognition of the rapid recognition traffic sign, the acquired input image needs to be subjected to feature extraction, so how to rapidly extract the image features with sufficient depth is important to the efficiency of traffic sign recognition. In the prior art, a residual network is generally adopted to extract the characteristics of an image, and a plurality of characteristic extraction blocks (blocks) are included in the traditional residual network, so that the extracted image characteristics can meet the recognition requirement in depth, but overfitting and improvement generalization easily occur, so that the extraction accuracy is reduced, and the subsequent recognition is influenced.
Disclosure of Invention
In order to overcome the defects of the prior art, the application aims to provide an improved residual error network and a characteristic extraction method and device thereof, which can prevent excessive fitting during characteristic extraction in practical application and improve the accuracy of characteristic extraction.
The application solves the problems by adopting the following technical scheme:
in a first aspect, the present application provides an improved residual network comprising: a preprocessing block, a feature extraction block and a Softmax layer;
the pretreatment block sequentially comprises a first BN layer, a first convolution layer, a first PReLU layer, a second convolution layer and a second PReLU layer;
the feature extraction block sequentially comprises a first feature extraction block, a second feature extraction block, a third feature extraction block and a fourth feature extraction block, and the output end of the preprocessing block is connected with the input end of the first feature extraction block;
the first feature extraction block, the second feature extraction block, the third feature extraction block and the fourth feature extraction block comprise a first convolution module and a maximum pooling layer, and the first convolution module comprises a third convolution layer, a second BN layer and a first ReLU layer in sequence; the output end of the first ReLU layer in the first feature extraction block and the second feature extraction block is connected with the maximum pooling layer, the third feature extraction block and the fourth feature extraction block further comprise Dropout layers for discarding the neural network units in the extraction process, and the output end of the Dropout layers is connected with the maximum pooling layer.
Further, a second convolution module is further included between the first convolution module and the Dropout layer in the fourth feature extraction block, and the second convolution module sequentially includes a fourth convolution layer, a third BN layer and a second ReLU layer.
Further, the convolution kernel sizes of the first and third convolution layers are 3×3; the convolution kernel sizes of the second convolution layer and the fourth convolution layer are 1×1.
In a second aspect, the present application provides a feature extraction method based on the improved residual network described above, comprising the steps of:
acquiring an original data set, and performing input pretreatment on the original data set to acquire an input data set;
the input data set is sent to the preprocessing block for preprocessing and then is input to the feature extraction block;
when the input data set is input to the Dropout layer, acquiring a preset discarding probability, and temporarily discarding the neural network unit in the feature extraction block according to the discarding probability;
acquiring data output of the fourth feature extraction block, and acquiring output features after aggregation with the data output of the third feature extraction block;
and inputting the output characteristics into a Softmax layer to calculate the probability of the prediction class corresponding to the output characteristics, and outputting the prediction class with the highest probability as a prediction result.
Further, the pre-input processing of the raw data set includes the steps of:
acquiring an extraction example in the original data set, and rotating the extraction example, wherein the range of the rotation angle value is-30 degrees to 30 degrees, and the step length is 3;
after the rotation is completed, the RGB images in the resulting dataset are converted into YUV images and the set of YUV images is set as the input dataset.
Further, the output of the first feature extraction block is input into the second feature extraction block and the third feature extraction block, respectively; the output of the second feature extraction block is respectively input into the third feature extraction block and the fourth feature extraction block; the output of the third feature extraction block is input into the fourth feature extraction block and Softmax layer, respectively.
In a third aspect, the present application provides a feature extraction device based on the improved residual network described above, comprising:
an input data set acquisition unit, configured to acquire an original data set, and perform input preprocessing on the original data set to acquire an input data set;
the feature extraction block input unit is used for sending the input data set to the preprocessing block for preprocessing and inputting the input data set to the feature extraction block;
the neural network unit discarding unit is used for acquiring preset discarding probability when the input data set is input to the Dropout layer, and temporarily discarding the neural network unit in the feature extraction block according to the discarding probability;
an output feature acquisition unit, configured to acquire data output of the fourth feature extraction block, and aggregate the data output of the third feature extraction block to obtain an output feature;
and the prediction result acquisition unit is used for inputting the output characteristics into the Softmax layer to calculate the probability of the prediction class corresponding to the output characteristics, and outputting the prediction class with the highest probability as a prediction result.
Further, the device also comprises the following devices:
an extraction example rotating unit, configured to obtain an extraction example in the original data set, and rotate the extraction example, where the rotation angle value ranges from-30 degrees to 30 degrees, and the step length is 3;
and the image conversion unit is used for converting the RGB images in the obtained data set into YUV images after finishing rotation and setting the set of YUV images as an input data set.
In a fourth aspect, the present application provides an improved feature extraction apparatus for a residual network, comprising at least one control processor and a memory for communicative connection with the at least one control processor; the memory stores instructions executable by the at least one control processor to enable the at least one control processor to perform the improved method of feature extraction of a residual network as described above.
In a fifth aspect, the present application provides a computer-readable storage medium storing computer-executable instructions for causing a computer to perform the improved method of feature extraction of a residual network as described above.
In a sixth aspect, the present application also provides a computer program product comprising a computer program stored on a computer readable storage medium, the computer program comprising program instructions which, when executed by a computer, cause the computer to perform the improved method of feature extraction of a residual network as described above.
One or more technical solutions provided in the embodiments of the present application have at least the following beneficial effects: the application adopts an improved residual error network and a characteristic extraction method and device thereof. And performing feature extraction through the set preprocessing block, the feature extraction block and the Softmax layer, setting a BN layer in the feature extraction block to accelerate the learning speed of the network, and setting a Dropout layer in the third feature extraction block and the fourth feature extraction block to discard the neural network unit in the extraction. Compared with the prior art, the method and the device have the advantages that the learning rate of the network is increased through the BN layer, the complexity of the network is reduced through the Dropout layer, the feature transition fitting phenomenon when the feature extraction depth is deep is effectively avoided, the extraction speed is ensured, and meanwhile, the extraction accuracy is improved.
Drawings
The application is further described below with reference to the drawings and examples.
FIG. 1 is a block diagram of an improved residual network provided in accordance with a first embodiment of the present application;
FIG. 2 is a flowchart of a feature extraction method of an improved residual network according to a second embodiment of the present application;
FIG. 3 is a flowchart of the input preprocessing of the original data set in a feature extraction method of an improved residual network according to a second embodiment of the present application;
FIG. 4 is a complete step diagram of a feature extraction method for an improved residual network according to a second embodiment of the present application;
fig. 5 is a schematic device diagram of a feature extraction device of an improved residual network according to a third embodiment of the present application;
fig. 6 is a schematic structural diagram of a feature extraction device of an improved residual network according to a fourth embodiment of the present application.
Detailed Description
Currently, with the development of unmanned technologies, advanced driving assistance systems are an important direction of research. In advanced driving assistance systems, early warning of traffic signs during driving is an important part of ensuring driving safety and legitimacy. In order to realize the recognition of the rapid recognition traffic sign, the acquired input image needs to be subjected to feature extraction, so how to rapidly extract the image features with sufficient depth is important to the efficiency of traffic sign recognition. In the prior art, a residual network is generally adopted to extract the characteristics of an image, and a plurality of characteristic extraction blocks (blocks) are included in the traditional residual network, so that the image characteristics extracted by the method can meet the recognition requirement in depth, but overfitting and improvement generalization easily occur, so that the extraction accuracy is reduced, and the subsequent recognition is influenced.
Based on the method, the application provides an improved residual error network and a characteristic extraction method and device thereof. And performing feature extraction through the set preprocessing block, the feature extraction block and the Softmax layer, setting a BN layer in the feature extraction block to accelerate the learning speed of the network, and setting a Dropout layer in the third feature extraction block and the fourth feature extraction block to discard the neural network unit in the extraction. Compared with the prior art, the method and the device have the advantages that the learning rate of the network is increased through the BN layer, the complexity of the network is reduced through the Dropout layer, the feature transition fitting phenomenon when the feature extraction depth is deep is effectively avoided, the extraction speed is ensured, and meanwhile, the extraction accuracy is improved.
The present application will be described in further detail with reference to the drawings and examples, in order to make the objects, technical solutions and advantages of the present application more apparent. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the application.
It should be noted that, if not in conflict, the features of the embodiments of the present application may be combined with each other, which is within the protection scope of the present application. In addition, while functional block division is performed in a device diagram and logical order is shown in a flowchart, in some cases, the steps shown or described may be performed in a different order than the block division in the device, or in the flowchart.
Referring to fig. 1, a first embodiment of the present application provides an improved residual network comprising: a preprocessing block, a feature extraction block and a Softmax layer;
the pretreatment block sequentially comprises a first BN layer, a first convolution layer, a first PReLU layer, a second convolution layer and a second PReLU layer;
the feature extraction block sequentially comprises a first feature extraction block, a second feature extraction block, a third feature extraction block and a fourth feature extraction block, and the output end of the preprocessing block is connected with the input end of the first feature extraction block;
the first feature extraction block, the second feature extraction block, the third feature extraction block and the fourth feature extraction block comprise a first convolution module and a maximum pooling layer, and the first convolution module comprises a third convolution layer, a second BN layer and a first ReLU layer in sequence; the output end of the first ReLU layer in the first feature extraction block and the second feature extraction block is connected with the maximum pooling layer, the third feature extraction block and the fourth feature extraction block further comprise Dropout layers for discarding the neural network units in the extraction process, and the output end of the Dropout layers is connected with the maximum pooling layer.
In this embodiment, the feature extraction block may include any type of network with any hierarchy, and in this embodiment, a 27-layer structure as shown in the figure is preferably adopted. The BN layer can realize batch standardization of features, and the BN layer is introduced after convolution is completed, so that the learning rate of a network can be effectively accelerated. Preferably, in this embodiment, a ReLU layer is preferably used as an excitation function, where the ReLU layer in the preprocessing speed is preferably a prerlu layer, which adds parameter correction, and further optimizes the preprocessed features, so that the features output by the preprocessing block are more accurate.
In this embodiment, a Dropout layer may be added to any feature extraction block to discard neurons, in this embodiment, the Dropout layer is preferably added to the third feature extraction block and the fourth feature extraction block, after calculation of the first two feature extraction blocks, the number of neural network units in the first convolution modules in the third feature extraction block and the fourth feature extraction block is more, so that the risk of excessive fitting is increased, and in this embodiment, the Dropout layer is added to the second two feature extraction blocks, so that part of the neural network units can be temporarily discarded, thereby achieving the effect of reducing network complexity.
In this embodiment, an arbitrary pooling layer may be used before the feature extraction module obtains the output to reduce parameters and features, such as mean-pooling, max-pooling, and storage-pooling. In this embodiment, max-pooling (maximum pooling layer) is preferably adopted, and since the maximum pooling layer maximizes feature points in the neighborhood, the estimated mean value caused by the parameter error of the convolution layer can be reduced well, texture information in the feature is reserved more, and the accuracy of feature extraction is ensured.
In this embodiment, the first feature extraction block and the second feature extraction block have the same structure, which is beneficial to extracting features in the same manner in the initial stage of extraction, so as to ensure that feature extraction is sufficient.
Further, in another embodiment of the present application, a second convolution module is further included between the first convolution module and the Dropout layer in the fourth feature extraction block, where the second convolution module sequentially includes a fourth convolution layer, a third BN layer, and a second ReLU layer.
In this embodiment, the fourth feature extraction block may be the same as the third feature extraction block, and other network layers related to feature extraction may be added. In this embodiment, it is preferable to add a second convolution module in the fourth feature extraction block, and further extract features through a second convolution layer including different convolution kernels, so that it is beneficial to fully extract image features in the final stage and improve the accuracy of feature extraction.
Further, in another embodiment of the present application, the convolution kernel sizes of the first and third convolution layers are 3×3; the convolution kernel sizes of the second convolution layer and the fourth convolution layer are 1×1.
Wherein the convolution kernel may be of any size, and in this embodiment it is preferred to use both 3 x 3 and 1 x 1. And the feature extraction is carried out by arranging two convolution layers with different convolution kernels in the preprocessing block and the fourth feature extraction block, so that the feature extraction sufficiency can be improved.
Referring to fig. 2, in addition, a second embodiment of the present application further provides a feature extraction method based on the improved residual network described above, including the following steps:
step S1, an original data set is obtained, and input pretreatment is carried out on the original data set to obtain an input data set;
step S2, the input data set is sent to the preprocessing block for preprocessing and then is input to the feature extraction block;
step S3, when the input data set is input to the Dropout layer, acquiring preset discarding probability, and temporarily discarding the neural network unit in the feature extraction block according to the discarding probability;
s4, acquiring data output of the fourth feature extraction block, and acquiring output features after aggregation with the data output of the third feature extraction block;
and S5, inputting the output characteristics into a Softmax layer to calculate the probability of the prediction class corresponding to the output characteristics, and outputting the prediction class with the highest probability as a prediction result.
In this embodiment, the improved residual network may be trained by any data set, so that image extraction can be achieved, and in this embodiment, a german traffic sign recognition base data set (GTSRB) is preferably adopted, where the german traffic sign recognition base data set has 43 classes, has more samples, and can provide a better data set for training.
In this embodiment, the input data set performs feature extraction layer by layer according to the 27-layer network as shown in fig. 1, where Dropout layers are included in the third feature extraction block and the fourth feature extraction block, and the Dropout layers need to be preset with a discarding probability, for example, 50% and 70%, so after the Dropout layers are executed in the two feature extraction blocks, the preset discarding probability needs to be read, and a neural network list is temporarily discarded according to the discarding probability, thereby reducing complexity of calculation of the neural network, and avoiding overfitting.
In this embodiment, the Softmax layer is a classifier, the probability that the input features correspond to each pre-stored prediction class is calculated through the Softmax, the input features are ranked according to the probability, and the image with the highest probability represents the classification of the extracted features, so that image recognition is realized.
Referring to fig. 3, further, in another embodiment of the present application, the pre-input processing of the raw data set includes the steps of:
step S11, obtaining an extraction example in the original data set, and rotating the extraction example, wherein the range of the rotation angle value is-30 degrees to 30 degrees, and the step length is 3;
step S12, after the rotation is completed, converting the RGB images in the obtained data set into YUV images, and setting the set of YUV images as an input data set.
In this embodiment, the input preprocessing may be performed on the original data set in any manner, and in this embodiment, the data set addition and the YUV color space conversion are preferably used. The extraction example is rotated around the center of the collection, the rotation angle range is [ -30 ° -30 ° ], the step length is 3, and the size of the obtained data set can be increased by 21 times. For example, taking the above-mentioned GTSRB dataset as an example, the GTSRB dataset has 43 classes, and the input image size is 32×32, then the number of rotated samples is 35288×21, which effectively increases the number of datasets, is more beneficial to feature extraction of the model, and avoids overfitting and improvement generalization.
In this embodiment, the image may be input in any color format, such as RGB and YUV, and in this embodiment, an image in YUV color space is preferably used. Since the residual network can extract rich functions by itself, the data set conversion color space can help to improve the performance of residual network feature extraction. The YUV color space has more advantages in illumination change and shape detection, and can be better suitable for input images in different external environments. Wherein, when the input image is an RGB image, the following formula is referred to as the conversion formula:
Y＝0.2291R+0.5876G+0.114B；
U＝0.492(B-Y)；
V＝0.877(R-Y)。
further, in another embodiment of the present application, the outputs of the first feature extraction block are input into the second feature extraction block and the third feature extraction block, respectively; the output of the second feature extraction block is respectively input into the third feature extraction block and the fourth feature extraction block; the output of the third feature extraction block is input into the fourth feature extraction block and Softmax layer, respectively.
In this embodiment, the output of each feature extraction block may be provided to any subsequent feature extraction block, which is preferably provided in the manner shown in fig. 1 in this embodiment, so as to avoid excessive aggregation of input features, which results in excessively complex data calculation.
Referring to fig. 4, in addition, another embodiment of the present application further provides a feature extraction method based on the improved residual network described above, including the following steps:
step S110, after an original data set is input, an extraction example in the original data set is obtained, the extraction example is rotated, the range of the rotation angle value is-30 degrees to 30 degrees, and the step length is 3;
step S120, after the rotation is completed, converting the RGB image in the obtained data set into a YUV image with the size of 32 x 32, and setting the set of YUV images as an input data set;
step S210, the input data set is sent to the preprocessing block for preprocessing;
step S310, inputting the preprocessed input data set into the first feature extraction block, and respectively inputting the obtained output into the second feature extraction block and the third feature extraction block;
step S320, inputting the output of the second feature extraction block into a third feature extraction block and a fourth feature extraction block respectively;
step S331, when the input data set is input to the Dropout layer of the third feature extraction block, acquiring a preset discarding probability, and temporarily discarding the neural network unit in the feature extraction block according to the discarding probability;
step S332, inputting the output of the third feature extraction block into a fourth feature extraction block and a Softmax layer respectively;
step S341, after obtaining the output of the first convolution module in the fourth feature extraction block, inputting the output into a second convolution module, and inputting the obtained output into a Dropout layer to perform temporary discarding of a neural network unit, wherein the discarding probability is the same as that in step S331;
step S342, aggregating the output of the fourth feature extraction block and the output of the third feature extraction block, and inputting the aggregated output into a Softmax layer;
step S410, inputting the output feature into a Softmax layer to calculate the probability of the prediction class corresponding to the output feature, and outputting the prediction class with the highest probability as a prediction result.
In this embodiment, after the original data set is obtained, input preprocessing is performed first, which is favorable for ensuring that the image of the input data set is converted into a YUV image, and meets the input requirement of the improved residual error network of the present application; the preprocessing block preprocesses the input data set through two convolution modules, can perform preliminary processing on the input data set before feature extraction, and is beneficial to improving the feature extraction efficiency; the output of the feature extraction block in this embodiment is respectively input into the feature extraction blocks of the later 2 layers, which is beneficial to simplifying the complexity of the data; in the embodiment, the Dropout layer is introduced to temporarily discard the neural network, so that the calculated amount of the extraction network can be greatly reduced, and the data fitting is effectively avoided; in this embodiment, after the features are extracted, the probability corresponding to the prediction class is calculated through the Softmax layer, so that the image recognition can be more accurately completed.
Referring to fig. 5, a third embodiment of the present application further provides a feature extraction apparatus based on the improved residual network described above, in the feature extraction apparatus 1000 of the improved residual network, including but not limited to: an input data set acquisition unit 1100, a feature extraction block input unit 1200, a neural network unit discarding unit 1300, an output feature acquisition unit 1400, and a prediction result acquisition unit 1500.
The input data set obtaining unit 1100 is configured to obtain an original data set, perform input preprocessing on the original data set, and obtain an input data set;
the feature extraction block input unit 1200 is configured to send the input data set to the preprocessing block for preprocessing, and then input the input data set to the feature extraction block;
the neural network unit discarding unit 1300 is configured to obtain a preset discarding probability when the input data set is input to the Dropout layer, and temporarily discard the neural network unit in the feature extraction block according to the discarding probability;
the output feature acquiring unit 1400 is configured to acquire data output of the fourth feature extracting block, and aggregate the data output of the third feature extracting block to obtain an output feature;
the prediction result obtaining unit 1500 is configured to input the output feature into a Softmax layer to calculate a probability of a prediction class corresponding to the output feature, and output the prediction class with the highest probability as a prediction result.
Further, in another embodiment of the present application, there is also included, but is not limited to: an extraction instance rotation unit 1110 and an image conversion unit 1120.
The extraction instance rotation unit 1110 is configured to obtain an extraction instance in the original dataset, rotate the extraction instance, where the rotation angle value ranges from-30 degrees to 30 degrees, and the step size is 3;
the image conversion unit 1120 is configured to convert the RGB image in the obtained dataset into a YUV image after completing the rotation, and set the set of YUV images as the input dataset.
It should be noted that, since the feature extraction device of the improved residual network in the present embodiment and the feature extraction method of the improved residual network described above are based on the same inventive concept, the corresponding content in the method embodiment is also applicable to the embodiment of the present device, and will not be described in detail herein.
Referring to fig. 6, a fourth embodiment of the present application further provides a feature extraction device based on the improved residual network described above, where the feature extraction device 6000 of the improved residual network may be any type of intelligent terminal, such as a mobile phone, a tablet computer, a personal computer, etc.
Specifically, the feature extraction apparatus 6000 of the improved residual network includes: one or more control processors 6001 and memory 6002, one control processor 6001 being illustrated in fig. 6.
The control processor 6001 and memory 6002 may be connected by a bus or otherwise, for example in fig. 6.
The memory 6002 serves as a non-transitory computer readable storage medium, and is operable to store a non-transitory software program, a non-transitory computer-executable program, and modules, such as program instructions/modules corresponding to the feature extraction device of the improved residual network in the embodiment of the application, for example, the input data set acquisition unit 1100 and the feature extraction block input unit 1200 shown in fig. 6. The control processor 6001 executes various functional applications and data processing of the feature extraction device 1000 of the improved residual network by running non-transitory software programs, instructions, and modules stored in the memory 6002, that is, implements the feature extraction method of the improved residual network of the above-described method embodiment.
The memory 6002 may include a storage program area that may store an operating system, at least one application program required for functions, and a storage data area; the storage data area may store data created according to the use of the feature extraction apparatus 1000 of the modified residual network, and the like. In addition, memory 6002 may include high speed random access memory, and may include non-transitory memory, such as at least one magnetic disk storage device, flash memory device, or other non-transitory solid state storage device. In some embodiments, memory 6002 optionally includes memory remotely located with respect to control processor 6001, which may be connected to feature extraction device 6000 of the modified residual network through a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.
The one or more modules are stored in the memory 6002, which when executed by the one or more control processors 6001, perform the improved residual network feature extraction method of the method embodiments described above, e.g., perform method steps S1-S5 of fig. 2, and method steps S11-S12 of fig. 3 described above, to implement the functions of the apparatus 1100-1500 of fig. 5.
Embodiments of the present application also provide a computer-readable storage medium storing computer-executable instructions for execution by one or more control processors, e.g., one of control processors 6001 in fig. 6, which may cause the one or more control processors 6001 to perform the method of feature extraction of the improved residual network in the method embodiments described above, e.g., to perform method steps S1 through S5 in fig. 2, and method steps S11 through S12 in fig. 3 described above, to implement the functions of apparatus 1100-1500 in fig. 5.
The device embodiments described above are merely illustrative, in that the devices illustrated as separate components may or may not be physically separate, i.e., may be located in one place, or may be distributed over multiple network devices. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
From the above description of embodiments, it will be apparent to those skilled in the art that the embodiments may be implemented in software plus a general purpose hardware platform. Those skilled in the art will appreciate that all or part of the processes implementing the methods of the above embodiments may be implemented by a computer program for instructing relevant hardware, where the program may be stored in a computer readable storage medium, and where the program, when executed, may include the processes of the embodiments of the methods described above. The storage medium may be a magnetic disk, an optical disk, a Read Only Memory (ROM), a random access Memory (Random Access Memory, RAM), or the like.
While the preferred embodiment of the present application has been described in detail, the present application is not limited to the above embodiment, and various equivalent modifications and substitutions can be made by those skilled in the art without departing from the spirit of the present application, and these equivalent modifications and substitutions are intended to be included in the scope of the present application as defined in the appended claims.
Claims (5)
1. A method of feature extraction of an improved residual network, the improved residual network comprising: a preprocessing block, a feature extraction block and a Softmax layer; the pretreatment block sequentially comprises a first BN layer, a first convolution layer, a first PReLU layer, a second convolution layer and a second PReLU layer; the feature extraction block sequentially comprises a first feature extraction block, a second feature extraction block, a third feature extraction block and a fourth feature extraction block, and the output end of the preprocessing block is connected with the input end of the first feature extraction block; the first feature extraction block, the second feature extraction block, the third feature extraction block and the fourth feature extraction block comprise a first convolution module and a maximum pooling layer, and the first convolution module comprises a third convolution layer, a second BN layer and a first ReLU layer in sequence; the output end of a first ReLU layer in the first feature extraction block and the second feature extraction block is connected with the maximum pooling layer, the third feature extraction block and the fourth feature extraction block also comprise a Dropout layer for discarding the neural network unit in the extraction process, and the output end of the Dropout layer is connected with the maximum pooling layer; a second convolution module is further included between the first convolution module and the Dropout layer in the fourth feature extraction block, the second convolution module sequentially includes a fourth convolution layer, a third BN layer and a second ReLU layer, and convolution kernels of the first convolution layer and the third convolution layer are 3×3; the convolution kernel sizes of the second convolution layer and the fourth convolution layer are 1 multiplied by 1;
the improved feature extraction method of the residual error network comprises the following steps:
acquiring an original data set, acquiring an extraction example in the original data set, rotating the extraction example, wherein the rotation angle value ranges from-30 degrees to 30 degrees, the step length is 3, converting an RGB image in the obtained data set into a YUV image after completing rotation, and setting a set of the YUV image as an input data set, and the original data set is a GTSRB data set;
the input data set is sent to the preprocessing block for preprocessing and then is input to the feature extraction block;
when the input data set is input to the Dropout layer, acquiring a preset discarding probability, and temporarily discarding the neural network unit in the feature extraction block according to the discarding probability;
acquiring data output of the fourth feature extraction block, and acquiring output features after aggregation with the data output of the third feature extraction block;
and inputting the output characteristics into a Softmax layer to calculate the probability of the prediction class corresponding to the output characteristics, and outputting the prediction class with the highest probability as a prediction result.
2. The method for feature extraction of an improved residual network of claim 1, wherein: the output of the first feature extraction block is respectively input into the second feature extraction block and the third feature extraction block; the output of the second feature extraction block is respectively input into the third feature extraction block and the fourth feature extraction block; the output of the third feature extraction block is input into the fourth feature extraction block and Softmax layer, respectively.
3. A feature extraction apparatus of an improved residual network, the improved residual network comprising: a preprocessing block, a feature extraction block and a Softmax layer; the pretreatment block sequentially comprises a first BN layer, a first convolution layer, a first PReLU layer, a second convolution layer and a second PReLU layer; the feature extraction block sequentially comprises a first feature extraction block, a second feature extraction block, a third feature extraction block and a fourth feature extraction block, and the output end of the preprocessing block is connected with the input end of the first feature extraction block; the first feature extraction block, the second feature extraction block, the third feature extraction block and the fourth feature extraction block comprise a first convolution module and a maximum pooling layer, and the first convolution module comprises a third convolution layer, a second BN layer and a first ReLU layer in sequence; the output end of a first ReLU layer in the first feature extraction block and the second feature extraction block is connected with the maximum pooling layer, the third feature extraction block and the fourth feature extraction block also comprise a Dropout layer for discarding the neural network unit in the extraction process, and the output end of the Dropout layer is connected with the maximum pooling layer; a second convolution module is further included between the first convolution module and the Dropout layer in the fourth feature extraction block, the second convolution module sequentially includes a fourth convolution layer, a third BN layer and a second ReLU layer, and convolution kernels of the first convolution layer and the third convolution layer are 3×3; the convolution kernel sizes of the second convolution layer and the fourth convolution layer are 1 multiplied by 1;
the feature extraction device of the improved residual error network comprises the following devices:
an input data set obtaining unit, configured to obtain an original data set, obtain an extraction instance in the original data set, rotate the extraction instance, where the rotation angle value ranges from-30 degrees to 30 degrees, and the step size is 3, convert an RGB image in the obtained data set into a YUV image after completing rotation, and set a set of YUV images as an input data set, where the original data set is a GTSRB data set;
the feature extraction block input unit is used for sending the input data set to the preprocessing block for preprocessing and inputting the input data set to the feature extraction block;
the neural network unit discarding unit is used for acquiring preset discarding probability when the input data set is input to the Dropout layer, and temporarily discarding the neural network unit in the feature extraction block according to the discarding probability;
an output feature acquisition unit, configured to acquire data output of the fourth feature extraction block, and aggregate the data output of the third feature extraction block to obtain an output feature;
and the prediction result acquisition unit is used for inputting the output characteristics into the Softmax layer to calculate the probability of the prediction class corresponding to the output characteristics, and outputting the prediction class with the highest probability as a prediction result.
4. An improved residual network feature extraction device, characterized by: comprising at least one control processor and a memory for communication connection with the at least one control processor; the memory stores instructions executable by the at least one control processor to enable the at least one control processor to perform an improved residual network feature extraction method according to any one of claims 1-2.
5. A computer-readable storage medium, characterized by: the computer readable storage medium stores computer executable instructions for causing a computer to perform an improved method of feature extraction of a residual network according to any one of claims 1-2.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910049298.8A CN109858497B (en) | 2019-01-18 | 2019-01-18 | Improved residual error network and feature extraction method and device thereof |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910049298.8A CN109858497B (en) | 2019-01-18 | 2019-01-18 | Improved residual error network and feature extraction method and device thereof |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109858497A CN109858497A (en) | 2019-06-07 |
| CN109858497B true CN109858497B (en) | 2023-09-01 |
Family
ID=66895315
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910049298.8A Active CN109858497B (en) | 2019-01-18 | 2019-01-18 | Improved residual error network and feature extraction method and device thereof |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109858497B (en) |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107169504A (en) * | 2017-03-30 | 2017-09-15 | 湖北工业大学 | A kind of hand-written character recognition method based on extension Non-linear Kernel residual error network |
| WO2017166586A1 (en) * | 2016-03-30 | 2017-10-05 | 乐视控股（北京）有限公司 | Image identification method and system based on convolutional neural network, and electronic device |
| CN107945173A (en) * | 2017-12-11 | 2018-04-20 | 深圳市宜远智能科技有限公司 | A kind of skin disease detection method and system based on deep learning |
| CN108710826A (en) * | 2018-04-13 | 2018-10-26 | 燕山大学 | A kind of traffic sign deep learning mode identification method |
| CN109034035A (en) * | 2018-07-18 | 2018-12-18 | 电子科技大学 | Pedestrian's recognition methods again based on conspicuousness detection and Fusion Features |
-
2019
- 2019-01-18 CN CN201910049298.8A patent/CN109858497B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2017166586A1 (en) * | 2016-03-30 | 2017-10-05 | 乐视控股（北京）有限公司 | Image identification method and system based on convolutional neural network, and electronic device |
| CN107169504A (en) * | 2017-03-30 | 2017-09-15 | 湖北工业大学 | A kind of hand-written character recognition method based on extension Non-linear Kernel residual error network |
| CN107945173A (en) * | 2017-12-11 | 2018-04-20 | 深圳市宜远智能科技有限公司 | A kind of skin disease detection method and system based on deep learning |
| CN108710826A (en) * | 2018-04-13 | 2018-10-26 | 燕山大学 | A kind of traffic sign deep learning mode identification method |
| CN109034035A (en) * | 2018-07-18 | 2018-12-18 | 电子科技大学 | Pedestrian's recognition methods again based on conspicuousness detection and Fusion Features |
Non-Patent Citations (1)
| Title |
|---|
| "信道自适应主动块低频分量丢弃 数字图像传输技术";王坚等;《计算机工程与应用》;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109858497A (en) | 2019-06-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| EP4145353A1 (en) | Neural network construction method and apparatus | |
| EP4198820A1 (en) | Training method for semi-supervised learning model, image processing method, and device | |
| US11816570B2 (en) | Method for accelerated detection of object in videos, server, and non-transitory computer readable storage medium | |
| US11003895B2 (en) | Systems and methods for verifying authenticity of ID photo | |
| US10614339B2 (en) | Object detection with neural network | |
| CN110929569B (en) | Face recognition method, device, equipment and storage medium | |
| CN110163801B (en) | Image super-resolution and coloring method, system and electronic equipment | |
| US20200005151A1 (en) | Artificial neural network | |
| CN110830807B (en) | Image compression method, device and storage medium | |
| CN109359527B (en) | Hair region extraction method and system based on neural network | |
| CN110232418B (en) | Semantic recognition method, terminal and computer readable storage medium | |
| CN111178312B (en) | Face expression recognition method based on multi-task feature learning network | |
| CN110599554A (en) | Method and device for identifying face skin color, storage medium and electronic device | |
| CN112580643A (en) | License plate recognition method and device based on deep learning and storage medium | |
| CN112465709B (en) | Image enhancement method, device, storage medium and equipment | |
| CN114445651A (en) | Training set construction method and device of semantic segmentation model and electronic equipment | |
| WO2022194079A1 (en) | Sky region segmentation method and apparatus, computer device, and storage medium | |
| CN112052949A (en) | Target model obtaining method and device, electronic equipment and storage medium | |
| CN112560701B (en) | Face image extraction method and device and computer storage medium | |
| CN109858497B (en) | Improved residual error network and feature extraction method and device thereof | |
| CN113743277A (en) | Method, system, equipment and storage medium for short video frequency classification | |
| Li et al. | Object Detection for UAV Images Based on Improved YOLOv6 | |
| CN113095506A (en) | Machine learning method, system and medium based on end, edge and cloud cooperation | |
| CN110598650A (en) | License plate false alarm filtering method and device, electronic equipment and storage medium | |
| CN112464924A (en) | Method and device for constructing training set |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |