CN107067029A - The image classification method that a kind of ELM and DE based on multi-channel feature are combined - Google Patents

Abstract

The invention provides the image classification method that a kind of ELM based on multi-channel feature and DE is combined, it is characterised in that including training process and prediction process, the training process includes：Step a1：Take under different illumination scenes, the face such as all ages and classes, different sexes as positive sample, remove other body part fritters outside face as negative sample；Step a2：Above-mentioned size is uniformly transformed into given size, and carries out gaussian filtering process；Step a3：The integrating channel feature of above-mentioned sample image is extracted, including：Greyscale color channel characteristics, gradient orientation histogram channel characteristics and gradient magnitude channel characteristics；Step a4：Using the integrating channel feature extracted in step 3 is as ELM input and carries out image classification training, meanwhile, optimization is improved to ELM using DE differential evolution algorithms, is optimal ELM classifying quality, so as to obtain the grader trained；The prediction process is classified using the grader trained to image.

Description

The image classification method that a kind of ELM and DE based on multi-channel feature are combined

Technical field

The present invention relates to image classification field, more particularly to a kind of ELM (extreme based on multi-channel featureLearning machine ExtremeLearningMachines) and the figures that are combined of DE (differential evolution differential evolution algorithms)As sorting technique.

Background technology

Image classification problem is the underlying issue of computer nowadays vision and many important research fields of image processing field.Good Image Classfication Technology can efficiently solve the problem of other scientific research fields, such as field of image search, remote sensing images are ledDomain, three-dimensional reconstruction field etc..The purpose of classification is exactly to set up a grader according to existing characteristics of image, can be to unknownImage type be predicted.For example, obtaining land-use map, vegetative coverage figure and some other by remote sensing image classificationMap, and then these maps are subjected to environment, land use as the basic map of next step；Medically can be by XLine graph classification carries out the diagnosis of breast cancer bump.

Therefore, many image classification methods are suggested in the last few years.There is the figure based on texture, shape and color spaceAs sorting technique, grader typically all uses SVM (Support Vector Machine, SVMs).But these sidesThe problem of method all has different degrees of.Expression is not enough to firstly, for these textures, shape and the color space characteristic of useThe feature of image.Secondly, SVM classifier needs feedback regulation parameter, and time-consuming, it is impossible to reach Fast Classification effect.

The content of the invention

The present invention is directed to above-mentioned technical problem, proposes the image point that a kind of ELM and DE based on multi-channel feature are combinedClass method.The inventive method improves nicety of grading by extracting the characteristics of image of high resolution, and by using time-consuming fewGrader realize rapid image categorization.

The image classification method that the ELM based on multi-channel feature and DE of the present invention is combined, including：Training process and pre-Survey process.The training process includes：

Step a1：Obtain positive negative sample；

Step a2：Above-mentioned size is uniformly transformed into given size, and carries out gaussian filtering process；

Step a3：The integrating channel feature of above-mentioned sample image is extracted, including：Greyscale color channel characteristics, gradient directionHistogram channel characteristics and gradient magnitude channel characteristics；

Step a4：Using the integrating channel feature extracted in step 3 is as ELM input and carries out image classification training, togetherWhen, optimization is improved to ELM using DE differential evolution algorithms, is optimal ELM classifying quality, so as to be trainedGrader.

The prediction process includes：

Step b1：The size of image to be predicted is reduced to given size, and carries out gaussian filtering process；

Step b2：Extract the integrating channel feature of above-mentioned image to be predicted, including greyscale color passage, gradient direction NogataFigure passage and gradient magnitude passage；

Step b3：Using the integrating channel feature of said extracted as the above-mentioned grader trained input, by describedGrader is classified, and classification results are exported.

The technical solution adopted in the present invention is the image classification side that a kind of ELM based on multi-channel feature and DE is combinedMethod, it is possible to achieve quick, high accuracy, the image classification with very high practical value.

Brief description of the drawings

Fig. 1 is the flow chart of the image classification method of the present invention.

Fig. 2 (a) is the schematic diagram of the positive sample collection (face) of the present invention；Fig. 2 (b) is the positive sample collection (face) after reducingSchematic diagram.

Fig. 3 (a) is the schematic diagram of the negative sample collection (people for being capped face) of the present invention；Fig. 3 (b) is negative after reducingThe schematic diagram of sample set (face).

Embodiment

Below by embodiment, the invention will be further described, and its purpose is only that the research for more fully understanding the present inventionThe protection domain that content is not intended to limit the present invention.

As shown in figure 1, the image classification method that a kind of ELM and DE based on multi-channel feature of the present invention are combined, bagTraining process and prediction process are included, the training process comprises the following steps：

Step a1：Obtain in positive negative sample, the present embodiment using facial image as training sample, take under different illumination scenes,The faces such as all ages and classes, different sexes remove other body part fritters outside face as negative sample, such as positive sampleShown in Fig. 2 (a) and Fig. 3 (a).Any image without face can serve as negative sample, in order to reduce flase drop, and the present invention, which is used, to be coveredThe image that lid falls face is used as negative sample collection.In the present invention, positive sample quantity is 50,000, negative sample quantity 100,000.

Step a2：Above-mentioned size is uniformly scaled or be cut to given size, and carries out gaussian filtering process.ThisIn, given size can be 20*20 (resolution ratio), as shown in Fig. 2 (b) and Fig. 3 (b).Can certainly be according to the actual requirementsIt is other sizes.Used here as gaussian filtering process, weaken noise, strengthen contrast.

Step a3：The integrating channel feature of above-mentioned sample image is extracted, including：Greyscale color channel characteristics, gradient directionHistogram channel characteristics and gradient magnitude channel characteristics.Here, greyscale color channel characteristics be 1, gradient orientation histogram lead toRoad is characterized as 6, and gradient magnitude channel characteristics are 1.Certainly, according to actual demand, different characteristics can also be extractedAmount.

Step a4：Using the integrating channel feature extracted in step a3 is as ELM input and carries out image classification training, togetherWhen, optimization is improved to ELM using DE differential evolution algorithms, is optimal ELM classifying quality, so as to be trainedGrader.

Further, optimization is improved to ELM using DE differential evolution algorithms in step a4 to further comprise：

Step a41：The input weight w i that ELM is randomly generated and hidden layer weights b_iAs DE initial population θ and inputDE；

In above-mentioned formula (1), k represents node numbers of hidden layers, and n represents to input neuron number.To each individual in population, meterCalculate corresponding generalized inverse matrix, calculate the result of generalized inverse matrix root-mean-square error on checking collection, and using the error asThe fitness of each individual, the calculation formula of the root-mean-square error is as follows：

In above-mentioned formula (2), E represents root-mean-square error, (x_j,t_j) be ELM training data, N be input sample number, i tablesShow i-th of hidden layer node, j represents j-th of input data, and m represents input data x dimension (i.e. characteristic), rule of thumb setIt is 50 to put the maximum evolutionary generations of DE, chooses θ during minimal error, is used as input weights optimal ELM and hidden layer weights；

Step a42：Using ELM classification error rate as threshold value, the error rate is the image and total picture number of misclassification classThe ratio of amount, the input weights of the ELM to being obtained in step a41 are carried out preferably with hidden layer weights.Specifically, an ELM is setMaximum receptible error rate δ, if in DE maximum evolutionary generation, obtained ELM error rates are less than δ, then now DEIt is output as ELM best initial weights；If DE reaches maximum evolutionary generation, and ELM error rate remains unchanged more than the threshold value set, then selectsDE when selecting ELM minimal error rates in maximum algebraically is output as best initial weights.

By above-mentioned steps, the optimal grader trained can be obtained.

Prediction process is described below, the prediction process comprises the following steps：

Step b1：By the scaled of image to be predicted to given size, and carry out gaussian filtering process；

Step b2：Extract the integrating channel feature of above-mentioned image to be predicted, including greyscale color passage, gradient direction NogataFigure passage and gradient magnitude passage；

Step b3：Using the integrating channel feature of said extracted as the above-mentioned grader trained input, by describedGrader is classified, and classification results are exported.

The step of can also including statistics false drop rate after above-mentioned steps b3.

The image classification method being combined by the ELM based on multi-channel feature and DE of the present invention, it is possible to achieve quick,High accuracy, the image classification with very high practical value.

Obviously, those of ordinary skill in the art is it should be appreciated that the embodiment of the above is intended merely to explanation originallyInvention, and be not used as limitation of the invention, as long as in the spirit of the present invention, to embodiment described aboveChange, modification will all fall in claims of the present invention model.

Claims (6)

1. the image classification method that a kind of ELM and DE based on multi-channel feature are combined, it is characterised in that including training processWith prediction process, the training process includes：

Step a1：Obtain positive negative sample；

Step a2：Above-mentioned size is uniformly transformed into given size, and carries out gaussian filtering process；

Step a3：The integrating channel feature of above-mentioned sample image is extracted, including：Greyscale color channel characteristics, gradient direction NogataFigure channel characteristics and gradient magnitude channel characteristics；

Step a4：Using the integrating channel feature extracted in step 3 is as ELM input and carries out image classification training, meanwhile, makeOptimization is improved to ELM with DE differential evolution algorithms, is optimal ELM classifying quality, so as to obtain point trainedClass device；

The prediction process includes：

Step b1：The size of image to be predicted is reduced to given size, and carries out gaussian filtering process；

Step b2：The integrating channel feature of above-mentioned image to be predicted is extracted, including greyscale color passage, gradient orientation histogram lead toRoad and gradient magnitude passage；

Step b3：Using the integrating channel feature of said extracted as the input of the above-mentioned grader trained, pass through the classificationDevice is classified, and classification results are exported.

2. image classification method according to claim 1, it is characterised in that in step a1, takes under different illumination scenes, noThe faces such as same age, different sexes remove other body part fritters outside face as negative sample as positive sample.

3. image classification method according to claim 1, it is characterised in that in step a2, the given size is 20*20。

4. image classification method according to claim 1, it is characterised in that in step a3, the greyscale color passage is specialIt is 6 to levy as 1, gradient orientation histogram channel characteristics, and the gradient magnitude channel characteristics are 1.

5. image classification method according to claim 1, it is characterised in that calculated in the step a4 using DE differential evolutionsMethod is improved optimization to ELM and further comprised：

Step a41：The input weight w i that ELM is randomly generated and hidden layer weights b_iAs DE initial population θ and input DE；

In above-mentioned formula (1), k represents node numbers of hidden layers, and n represents to input neuron number；To each individual in population, phase is calculatedThe generalized inverse matrix answered, calculates the result of generalized inverse matrix root-mean-square error on checking collection, and using the error as eachThe fitness of individual, the calculation formula of the root-mean-square error is as follows：

In above-mentioned formula (2), E represents root-mean-square error, (x_j,t_j) be ELM training sample, N is input sample number, and i represents theI hidden layer node, j represents j-th of input data, and m represents input data x dimension i.e. characteristic, the evolution for setting DE maximumAlgebraically is empirical value 50, chooses θ during lowest mean square root error, is used as input weights optimal ELM and hidden layer weights；

Step a42：Using ELM error rate as threshold value, amount of images and participation classification of the error rate for classification errorTotal amount of images ratio, the input weights of the ELM to being obtained in step a41 carry out preferred, i.e. one ELM of setting with hidden layer weightsMaximum receptible error rate δ, if in DE maximum evolutionary generation, obtained ELM error rates are less than δ, then now DEIt is output as ELM best initial weights；If DE reaches maximum evolutionary generation, and ELM error rate remains unchanged more than the threshold value set, then selectsDE when selecting ELM minimal error rates in maximum algebraically is output as best initial weights.

6. image classification method according to claim 1, it is characterised in that statistics can also be included after the step b3The step of false drop rate.