CN112101383B - A color shift image recognition method - Google Patents

Abstract

The invention provides a color-shift image recognition method; a neural network model is used to identify whether an input image is a color-shift image, and the training set used for training the neural network model adopts a normal picture set and a color-shift image generated on the basis of the normal picture set set build. Based on the method of generating a color shift picture set on the basis of the normal picture set, the present invention can effectively solve the problem of the data set for classifying the color shift pictures, and realize the effective identification of the color shift pictures.

Description

A color shift image recognition method

technical field

The invention relates to a color shift image recognition method.

Background technique

The convolutional neural network imitates the structure of the human cerebral cortex, processes the input data in layers, and then superimposes them layer by layer to mine the hidden features and laws in the data to achieve the purpose of distinction and recognition. In the process of hierarchical processing of data, the first few layers of processing extract shallow features such as image outline and color, and the latter layers focus on extracting more abstract features. Each layer includes a different number of neurons as the basic unit of data processing. The more layers and the more neurons in each layer, the greater the amount of information carried by the deep learning network and the stronger the processing function. In the process of superimposing the data, the extracted features are superimposed together, and then related functions such as softmax are applied to obtain a classification vector. The development of convolutional neural network has experienced the development of LeNet5, AlexNet, VGG-Net, GoogLeNet, and Inception series. The network depth is from shallow to deep, the number of neurons is from small to large, and the function is from simple to complex.

Convolutional neural networks have made great progress, but there are still some problems to be solved in color cast image classification. First, image training and validation sets are essential elements in machine training. However, there is currently no image training set and test set for color cast image distinction, which makes further machine learning work impossible.

SUMMARY OF THE INVENTION

In order to solve the above-mentioned technical problems, the present invention provides a color-shift image recognition method. The color-shift image recognition method is based on the method of generating a color-shift picture set on the basis of the normal picture set, which can effectively solve the data classification of color-shift pictures. Set the problem to achieve effective identification of color cast images.

The present invention is achieved through the following technical solutions.

The present invention provides a color-shift image recognition method; a neural network model is used to identify whether an input image is a color-shift image, and the training set used for training the neural network model adopts a normal picture set and a color-shift image generated on the basis of the normal picture set set build.

The color shift picture set is obtained by selecting a color channel by random numbers on the basis of a normal picture set, and then randomly shifting the selected color channel.

The random number ranges from 1 to 6, wherein 1 to 3 represent selecting a single color channel, and 4 to 6 represent selecting two color combination channels.

The neural network model adopts the VGG16 algorithm.

The neural network model is obtained by transferring the VGG16 pre-training model.

The neural network model adds a Dropout layer on the basis of the VGG16 pre-training model; the ratio of the Dropout layer is 0.6.

The neural network model is trained with a piecewise decaying learning rate.

The beneficial effects of the present invention are: based on the method of generating the color shift picture set based on the normal picture set, the problem of the data set for classifying the color shift pictures can be effectively solved, and the effective identification of the color shift pictures can be realized.

Detailed ways

The technical solutions of the present invention are further described below, but the claimed scope is not limited to the description.

The invention provides a color shift image recognition method; a neural network model is used to identify whether an input image is a color shift image, and the training set used for training the neural network model adopts a normal picture set and a color shift picture set generated on the basis of the normal picture set Construct.

The color shift picture set is based on the normal picture set, selects the color channel by random numbers, and then randomly shifts the selected color channel.

The random number ranges from 1 to 6, where 1 to 3 represent a single color channel, and 4 to 6 represent two color combination channels.

The neural network model adopts the VGG16 algorithm.

The neural network model is obtained by transferring the VGG16 pre-training model.

The neural network model adds a Dropout layer on the basis of the VGG16 pre-training model; the ratio of the Dropout layer is 0.6.

The neural network model is trained with a piecewise decaying learning rate.

Example 1

Build a training set: build a training set and a validation set for color cast images

(1) Copy the normal picture set

Download the Pascal VOC2012 picture set on the Internet (you can also download other picture sets, the number of pictures is recommended to be more than 5000), copy the JPEGImages folder to folder 0, as the normal picture set of the picture training set, there are 17125 pictures in this folder picture.

(2) Generate a color cast image set

Take out all the pictures in folder 0 one by one, and process a normal picture without color cast through the following ①-③ steps to generate the corresponding color cast picture, and store the processed pictures in folder 1. 1 will generate 17125 color cast images, corresponding to folder 0, which together constitute the color cast image training set. Some of the pictures are taken as the verification set, and it is recommended to take about 500 pictures for each category.

(3) Set the starting point and step size

For any normal input picture, set the grayscale change starting point as start=16 and step size step=2.

(4) Generate channel selection random numbers and offset random numbers

Generate channel selection random integer n1, range 1-6, generate grayscale offset random integer n2, n3, range 0-100. Among them, n1 determines which channel is selected for color shift processing, and 6 channel selection modes are set in the present invention.

(5) Generate color shift pictures according to random numbers

The above two steps have completed the setting of the starting point and step size of the color shift, as well as the generation of random numbers. In this step, the channel for color cast pollution will be determined according to the channel selection random number, n1 is 1-6, corresponding to R channel, G channel, B channel, RG combined channel, RB combined channel and GB combined channel.

Then determine the amount of grayscale change according to the offset random number, the formula is as follows:

Among them, r, g, b represent the original grayscale respectively, and r_, g_, b_ represent the processed grayscale respectively. When n1 is 1, only R channel is changed, when n1 is 2, only G channel is changed, when n1 is 3, only B channel is changed, when n1 is 4, R channel and G channel are changed, when n1 is 5, R channel and B channel are changed, n1 When it is 6, change the G channel and B channel. The amount of change is determined by random numbers n2 and n3 according to the above formula. Therefore, the selection of the channel is random, and the value of the change amount is random.

Model Training: Deep Convolutional Neural Network Model Training

The purpose of this step is to train the image training set and validation set, optimize the parameters of each layer, and obtain a high-precision classifier to achieve accurate distinction between normal images and color-shifted images. The present invention is constructed based on the open-source vgg16 model, importing the open-source xxx.h5 weight file for migration learning, adding a Dropout layer, setting the learning rate to decay in segments, and training for 100 cycles to obtain a classifier.

①Add Dropout layer

The function of Dropout is to discard some parameters according to a certain proportion, so that the parameters match the samples and prevent overfitting. On the basis of the original model, the present invention sets the Dropout layer according to the ratio of 0.6.

②Set the segmented learning rate decay

On the basis of the original vgg16 model, the learning rate is set to decay in segments. According to each machine learning experiment, the value of the learning rate of each node is summed up. This training consists of 100 cycles, and the learning rate is updated every 10 cycles. The formula is as follows:

Among them, LR represents the learning rate, and epoch represents the training period.

Implementation Effect

Using the process of the above-mentioned specific embodiment, import the training set and the verification set, start the set vgg16 convolutional neural network model, train for 100 cycles, and obtain the classifier file. With the scheme of the present invention, after testing 549 normal pictures and 576 color shift pictures, the accuracy of identifying normal pictures is 0.979964, and the accuracy of identifying color shift pictures is 0.843750. For pictures outside the library, the accuracy of identifying 300 normal pictures is 0.87, and the accuracy of identifying 300 color-shifted pictures is 0.910000. It is verified that the obtained classifier has a high degree of discrimination between normal pictures and color-shifted pictures.

Claims (7)

1. A color cast image recognition method is characterized in that: the method comprises the following steps of adopting a neural network model to identify whether an input image is a color cast image, constructing a training set used for training the neural network model by adopting a normal picture set and a color cast picture set generated on the basis of the normal picture set, and generating the color cast picture set, wherein the specific steps of:

processing a normal picture without color cast through the following steps (1) to (3) to generate a corresponding color cast picture, storing the processed picture in a folder 1, wherein 17125 color cast pictures are generated by the folder 1 and correspond to the folder 0 to jointly form a color cast picture training set;

(1) setting a starting point and a step length

Setting a gray scale change starting point as start =16 and step =2 for any normal input picture;

(2) generating channel selection random numbers and offset random numbers

Generating a channel selection random integer n1 in a range of 1-6, and generating gray scale offset random integers n2 and n3 in a range of 0-100;

(3) color cast picture generation from random numbers

Setting a color cast starting point and a step length and generating a random number, wherein the step is to select the random number according to a channel to determine the channel for color cast pollution, n1 is 1-6, and the R channel, the G channel, the B channel, the RG combined channel, the RB combined channel and the GB combined channel correspond to the channel respectively; then, the gray scale change amount is determined according to the offset random number, and the formula is as follows:

wherein R, G and B respectively represent original gray scales, R _, G _andb _ respectively represent processed gray scales, when n1 is 1, only the R channel is changed, when n1 is 2, only the G channel is changed, when n1 is 3, only the B channel is changed, when n1 is 4, the R channel and the G channel are changed, when n1 is 5, the R channel and the B channel are changed, and when n1 is 6, the G channel and the B channel are changed;

the amount of change is determined by the random numbers n2 and n3 according to the above formula, the selection of the channel is random, and the value of the amount of change is random.

2. The color cast image recognition method as claimed in claim 1, wherein: the color cast picture set is obtained by selecting a color channel by a random number on the basis of a normal picture set and then performing random offset on the selected color channel.

3. The color cast image recognition method as claimed in claim 2, wherein: the range of the random number is 1-6, wherein 1-3 represents selecting a single color channel, and 4-6 represents selecting two color combination channels.

4. The color cast image recognition method as claimed in claim 1, wherein: the neural network model adopts a VGG16 algorithm.

5. The color cast image recognition method as claimed in claim 1, wherein: the neural network model is obtained by performing transfer learning through a VGG16 pre-training model.

6. The color cast image recognition method as claimed in claim 5, wherein: the neural network model adds a Dropout layer on the basis of a VGG16 pre-training model; dropout layer ratio is 0.6.

7. The color cast image recognition method as claimed in claim 1, wherein: the neural network model is trained by adopting a learning rate of sectional attenuation.