CN105005791A - Half-tone image classification method - Google Patents

Abstract

The invention discloses a half-tone image classification method. The method is carried according to the following steps of: connecting a sparse self-encoding depth neural network with a Softmax classifier to serve as a classification model of a half-tone image; partitioning the half-tone image for training, and training the model by using the blocks; extracting valid blocks from the half-tone image to be classified, and classifying the valid blocks; and counting type numbers of each valid block of the half-tone image to be classified, and taking a type with maximum type number as the type of the half-tone image to be classified. By the half-tone image classification method, the problem that manual selection and extraction is required for the characteristics of the half-tone image in the prior art is solved, the half-tone image classification method can be applicable for various types of half-tone images, and the classification accuracy is improved.

Description

A kind of half tone image sorting technique

Technical field

The invention belongs to printing technology, be specifically related to a kind of half tone image sorting technique.

Background technology

Half-tone picture similarly is that a class only comprises two contrasts, but can present the image adjusting continuously visual effect during eye-observation, and it is widely used in the fields such as traditional printing industry, digital publishing system, the display of various binaryzation and print out equipment.For the half tone image that various halftoning method generates, if will reuse, must remove the visual noises such as the reticulate pattern produced in halftoning process, this technology is called as inverse halftone technique.Current various high-quality inverse halftoning method all requires first to know corresponding half tone image classification, and namely this kind of half-tone picture similarly is by which kind of halftoning method produced.Therefore, classifying to the method that half tone image produces according to it, is key one step in various inverse halftone technique.

Current half tone image sorting technique is mostly first by artificial selection feature extracting method, the spatial feature of statistics half tone image texture or spectrum signature, carry out the expression of half tone image shallow-layer external performance, as method, covariance matrix feature extraction etc. that one dimension auto-relativity function method, gray level co-occurrence matrixes and run-length matrix combine, carry out the expression of half tone image shallow-layer external performance; The methods such as recycling BP neural network, Bayes classifier are classified.The feature extracting method of artificial selection is wasted time and energy, and does not have versatility, the pattern of the various half tone image of more difficult Efficient Characterization, and its shortcoming is Limited Number that is lower to the classification accuracy rate of various half tone image, classification.

Summary of the invention

The object of this invention is to provide a kind of half tone image sorting technique, solve the problem that existing half tone image sorting technique accuracy is low.

The technical solution adopted in the present invention is, a kind of half tone image sorting technique, specifically implements according to following steps:

Step 1, adopt sparse own coding deep neural network as the Feature Selection Model of half tone image, sparse own coding deep neural network connects Softmax sorter;

Step 2, to training half tone image carry out piecemeal, train sparse own coding deep neural network and Softmax sorter with each block image;

Step 3, piecemeal is carried out to half tone image to be sorted, extract its active block;

Step 4, using the input of the active block extracted from half tone image to be sorted as the sparse own coding deep neural network trained, carry out feature extraction, determined the classification of the active block extracted from half tone image to be sorted by Softmax sorter;

Step 5, add up the class number of the active block of half tone image to be sorted, the classification that class number is maximum is the classification of half tone image to be sorted.

Feature of the present invention is also:

In step 1, sparse own coding deep neural network comprises 4 layers, and ground floor is input layer, and the second layer and third layer are hidden layers, and the 4th layer is output layer, and output layer is connected with Softmax sorter.

In step 1, the output valve of Softmax sorter is the classification number of half tone image to be sorted.

The concrete steps of step 2 are:

Step 2.1, be divided into size to be the image block of m × n each width half tone image of training, the classification of image block is the classification of divided half tone image;

Step 2.2, the image block using the half tone image of training to mark off train sparse own coding deep neural network;

Step 2.3, the image block using the half tone image of training to mark off and classification thereof, training Softmax sorter, and adopt the parameter of method to sparse own coding deep neural network of overall fine setting to finely tune.

In step 2.1, non-overlapped block division methods is adopted to the division of the half tone image of training.

Sparse own coding deep neural network is trained to adopt gradient descent algorithm in step 2.2.

The concrete steps extracting active block in step 3 are:

Step 3.1, to each pixel in each piece, centered by pixel, choose the regional area of its M × M, calculate the local entropy H of pixel _ij:

$\begin{array}{cc}{H}_{ij}=\underset{k=0}{\overset{1}{\Σ}}{p}_k{\mathrm{logp}}_k& \left(\begin{array}{cc}i=1,2,\mathrm{...},m& j=1,2,\mathrm{...},n\end{array}\right)---\left(1\right)\end{array}$

P in formula (1) _kit is the probability that in the regional area of M × M, pixel value gets k;

Step 3.2, to each piece, calculate average value mu and the variances sigma of all local entropys in this block:

$\mathrm{\μ}=\frac{1}{m\×n}\underset{i=1}{\overset{m}{\Σ}}\underset{j}{\overset{n}{\Σ}}{H}_{ij}---\left(2\right)$

$\mathrm{\σ}=\sqrt{\frac{1}{m\×n}\underset{i=1}{\overset{m}{\Σ}}\underset{j}{\overset{n}{\Σ}}{(H_{ij}-\mathrm{\μ})}^2}---\left(3\right)$

Step 3.3, to each piece, if it meets μ >=0.5 and σ≤0.1, then judge that this block is as active block.

The invention has the beneficial effects as follows: a kind of half tone image sorting technique, half tone image feature is extracted by designed sparse own coding deep neural network, overcome existing artificial selection feature extracting method bad adaptability, the feature of extraction more can characterize different classes of half tone image feature; Extracted by image block and active block, improve the accuracy of half tone image classification, decrease the time required for classification.

Accompanying drawing explanation

Fig. 1 is the process flow diagram of a kind of half tone image sorting technique of the present invention.

Embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is described in detail.

A kind of half tone image sorting technique of the present invention, flow process as shown in Figure 1, is specifically implemented according to following steps:

Step 1, adopt sparse own coding deep neural network as the Feature Selection Model of half tone image, sparse own coding deep neural network connects Softmax sorter;

Wherein, sparse own coding deep neural network comprises 4 layers, and ground floor is input layer, and the neuron number of ground floor is the number of pixel in image block, and if tile size is 16 × 16pixel, i.e. the neuron number of ground floor input is 256; The second layer and third layer are hidden layers, and the neuron number of the second layer and third layer is respectively 200 and 100, and the 4th layer is output layer, and output layer is connected with Softmax sorter.

The output valve of Softmax sorter is the classification number of half tone image to be sorted, has 13 classifications in the present embodiment.

Step 2, to training half tone image carry out piecemeal, train sparse own coding deep neural network and Softmax sorter with each block image, concrete steps are:

Step 2.1, adopt non-overlapped block division methods to be divided into size to be the image block of m × n each width half tone image of training, the classification of image block is the classification of divided half tone image;

Step 2.2, the image block using the half tone image of training to mark off adopt gradient descent algorithm to train sparse own coding deep neural network;

Step 2.3, the image block using the half tone image of training to mark off and classification thereof, training Softmax sorter, and adopt the parameter of method to sparse own coding deep neural network of overall fine setting to finely tune.

Step 3, piecemeal is carried out to half tone image to be sorted, extract its active block;

The concrete steps wherein extracting active block are:

Step 3.1, to each pixel in each piece, centered by pixel, choose its regional area of 9 × 9, calculate the local entropy H of pixel _ij:

$\begin{array}{cc}{H}_{ij}=\underset{k=0}{\overset{1}{\Σ}}{p}_k{\mathrm{logp}}_k& \left(\begin{array}{cc}i=1,2,\mathrm{...},m& j=1,2,\mathrm{...},n\end{array}\right)---\left(1\right)\end{array}$

P in formula (1) _kit is the probability that in the regional area of 9 × 9, pixel value gets k;

Step 3.2, to each piece, calculate average value mu and the variances sigma of all local entropys in this block:

$\mathrm{\μ}=\frac{1}{m\×n}\underset{i=1}{\overset{m}{\Σ}}\underset{j}{\overset{n}{\Σ}}{H}_{ij}---\left(2\right)$

$\mathrm{\σ}=\sqrt{\frac{1}{m\×n}\underset{i=1}{\overset{m}{\Σ}}\underset{j}{\overset{n}{\Σ}}{(H_{ij}-\mathrm{\μ})}^2}---\left(3\right)$

Step 3.3, to each piece, if it meets μ >=0.5 and σ≤0.1, then judge that this block is as active block.

Step 4, using the input of the active block extracted from half tone image to be sorted as the sparse own coding deep neural network trained, carry out feature extraction, determined the classification of the active block extracted from half tone image to be sorted by Softmax sorter;

Step 5, add up the class number of the active block of half tone image to be sorted, the classification that class number is maximum is the classification of half tone image to be sorted.

A kind of half tone image sorting technique of the present invention, is a kind of feature extracting method of unsupervised learning, adapts to the feature extraction of all kinds of half tone image; The active block sorting technique simultaneously proposed, decreases the interference of invalid block to classification, improves the accuracy of classification; By extracting active block, decreasing the number of to be sorted piece, also just decreasing the time required for classification.By the classification to 13 class half tone images, average correct classification rate can reach 99.62%, achieves very satisfied effect.

Claims (7)

1. a half tone image sorting technique, is characterized in that, specifically implements according to following steps:

Step 1, adopt sparse own coding deep neural network as the Feature Selection Model of half tone image, sparse own coding deep neural network connects Softmax sorter;

Step 2, to training half tone image carry out piecemeal, train sparse own coding deep neural network and Softmax sorter with each block image;

Step 3, piecemeal is carried out to half tone image to be sorted, extract its active block;

Step 4, using the input of the active block extracted from half tone image to be sorted as the sparse own coding deep neural network trained, carry out feature extraction, determined the classification of the active block extracted from half tone image to be sorted by Softmax sorter;

Step 5, add up the class number of the active block of half tone image to be sorted, the classification that class number is maximum is the classification of half tone image to be sorted.

2. a kind of half tone image sorting technique according to claim 1, is characterized in that, in described step 1, sparse own coding deep neural network comprises 4 layers, ground floor is input layer, the second layer and third layer are hidden layers, and the 4th layer is output layer, and output layer is connected with described Softmax sorter.

3. a kind of half tone image sorting technique according to claim 1, is characterized in that, in described step 1, the output valve of Softmax sorter is the classification number of half tone image to be sorted.

4. according to claim 1 in described a kind of half tone image sorting technique, it is characterized in that, the concrete steps of described step 2 are:

Step 2.1, be divided into size to be the image block of m × n each width half tone image of training, the classification of image block is the classification of divided half tone image;

Step 2.2, the image block using the half tone image of training to mark off train sparse own coding deep neural network;

Step 2.3, the image block using the half tone image of training to mark off and classification thereof, training Softmax sorter, and adopt the parameter of method to sparse own coding deep neural network of overall fine setting to finely tune.

5. according to claim 4 in described a kind of half tone image sorting technique, it is characterized in that, in described step 2.1, non-overlapped block division methods is adopted to the division of half tone image of training.

6. according to claim 4 in described a kind of half tone image sorting technique, it is characterized in that, train in described step 2.2 sparse own coding deep neural network adopt gradient descent algorithm.

7. according to claim 1 in described a kind of half tone image sorting technique, it is characterized in that, the concrete steps extracting active block in described step 3 are:

Step 3.1, to each pixel in each piece, centered by pixel, choose the regional area of its M × M, calculate the local entropy H of pixel _ij:

$H_{ij}=\underset{k=0}{\overset{1}{\Σ}}{p}_k{\mathrm{logp}}_k,(i=1,2,...,m,j=1,2,...,n)---\left(1\right)$

P in formula (1) _kit is the probability that in the regional area of M × M, pixel value gets k;

Step 3.2, to each piece, calculate average value mu and the variances sigma of all local entropys in this block:

$\mathrm{\μ}=\frac{1}{m\×n}\underset{i=1}{\overset{m}{\Σ}}\underset{j}{\overset{n}{\Σ}}{H}_{ij}---\left(2\right)$

$\mathrm{\σ}=\sqrt{\frac{1}{m\×n}\underset{i=1}{\overset{m}{\Σ}}\underset{j}{\overset{n}{\Σ}}{(H_{ij}-\mathrm{\μ})}^2}---\left(3\right)$

Step 3.3, to each piece, if it meets μ >=0.5 and σ≤0.1, then judge that this block is as active block.