CN108460772B - Advertisement harassment fax image detection system and method based on convolutional neural network - Google Patents
Advertisement harassment fax image detection system and method based on convolutional neural network Download PDFInfo
- Publication number
- CN108460772B CN108460772B CN201810150076.0A CN201810150076A CN108460772B CN 108460772 B CN108460772 B CN 108460772B CN 201810150076 A CN201810150076 A CN 201810150076A CN 108460772 B CN108460772 B CN 108460772B
- Authority
- CN
- China
- Prior art keywords
- neural network
- keyword
- module
- convolution
- convolution module
- 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
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/155—Segmentation; Edge detection involving morphological operators
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10004—Still image; Photographic image
- G06T2207/10008—Still image; Photographic image from scanner, fax or copier
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Abstract
The invention provides an advertisement harassment fax image detection system and method based on a convolutional neural network, which comprises a keyword area extraction module, wherein the keyword area extraction module is used for determining a keyword suspicious area of a fax image to be detected; the neural network confidence coefficient analysis module is connected with the keyword region extraction module and is used for identifying characters of the keyword suspicious region to realize classification of fax images. The method extracts the keyword suspicious region through the keyword region extraction module, automatically operates, and has high working efficiency; the characters of the keyword suspicious region are identified through the neural network confidence coefficient analysis module, classification judgment of the advertisement harassment fax is achieved, time is saved, and the control capability is strong, so that the method has the characteristics of high working efficiency and strong control capability.
Description
Technical Field
The invention belongs to the technical field of image classification, and particularly relates to an advertisement harassment fax image detection system and method based on a convolutional neural network.
Background
With the popularization of internet technology, the number of text images is increasing, and automatic processing of text images is an important subject in the field of computer application. The text images are of various types, and the layout structure is becoming more and more complex, and the text images not only contain character areas with different character sizes, but also often include elements such as images, tables, figures and the like, and the layout forms of the text images are also various, and not only have rectangular layouts, but also have non-rectangular layouts. The layout analysis is an important link for the automatic processing of text images. The layout analysis is to process the text image by using a computer to determine the physical structure of the text image, and divide the image into areas with different attributes such as text, image, graph, table and the like, so as to meet the requirements of different applications such as character recognition, table recognition, icon recognition and the like of the text image. Layout analysis is always a research hotspot in the field of text image processing, and as a preprocessing process of text images, the accuracy of subsequent processing is directly influenced by the result of layout analysis. Facsimile images are an important type of text images, and searching and classifying facsimile images using telephone numbers, e-mail addresses, and the like included in facsimile images has attracted much attention.
The facsimile image has the characteristics of less color information, lower spatial resolution and the like which are different from the characteristics of a common optical image. This results in the problem that the general optical image classification method is not good for directly classifying the fax image. Inspired by Hubel and Wiesel for electrophysiological studies of the visual cortex of cats, Convolutional Neural Networks (CNNs) were proposed, which were first used by Yann Lecun for handwritten digit recognition. In recent years, the convolutional neural network continuously exerts force in a plurality of directions, and has breakthroughs in speech recognition, face recognition, general object recognition, motion analysis, natural language processing and even electroencephalogram analysis. In the aspect of fax image identification, the application of the convolutional neural network is still blank.
With the gradual and deep influence of the internet technology on social life, the fax image is used as a character image, and the harassment of the fax image is on an increasing trend. In recent years, the influence of the advertisements disturbing the faxes is increasingly serious, and a real-time and high-automation fax image detection technology is urgently needed. According to the knowledge, the classification of the prior advertisement harassment fax images is basically completed manually, and the problems of low automation degree, time consumption, low working efficiency and the like exist, so that the industrial supervision requirement cannot be met.
From the above analysis, the existing method for detecting the fax image of the advertisement harassment has the following defects:
1. the existing method for detecting the image of the advertisement harassment fax has low automation degree and low working efficiency;
2. the existing method for detecting the fax image of the advertisement harassment consumes time, has poor prevention capability and can not meet the supervision requirement of the industry.
Disclosure of Invention
The invention provides an advertisement harassment fax image detection system and method based on a convolutional neural network, which can effectively solve the problem of low working efficiency of the existing advertisement harassment fax image detection method and can also solve the problem of poor control capability of the existing advertisement harassment fax image detection method.
In order to solve the problems, the invention provides an advertisement harassment fax image detection system and method based on a convolutional neural network, and the technical scheme is as follows:
a convolutional neural network-based advertisement harassment fax image detection system comprises a keyword area extraction module, wherein the keyword area extraction module is used for determining a keyword suspicious area of a fax image to be detected; the neural network confidence coefficient analysis module is connected with the keyword region extraction module and is used for classifying characters in the keyword suspicious region.
The above advertisement harassment fax image detection system based on the convolutional neural network is further preferably: the keyword region extraction module comprises a binarization module and a morphological corrosion module; the binarization module is used for judging the keyword suspicious region; the morphological corrosion module is connected with the binarization module and is used for corroding the judged keyword suspicious region.
The above advertisement harassment fax image detection system based on the convolutional neural network is further preferably: the neural network confidence coefficient analysis module comprises an input layer, a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, a fifth convolution module and a neural network feature classifier.
The above advertisement harassment fax image detection system based on the convolutional neural network is further preferably: the neural network confidence coefficient analysis module comprises 23 layers, and the input layer is the layer 1 of the neural network confidence coefficient analysis module; the first convolution module is layers 2 to 4 of the neural network confidence analysis module; the second convolution module is layers 5 to 7 of the neural network confidence analysis module; the third convolution module is the 8 th to 11 th layers of the neural network confidence coefficient analysis module; the fourth convolution module is layers 12 to 15 of the neural network confidence analysis module; the fifth convolution module is layers 16 to 19 of the neural network confidence analysis module; the neural network feature classifier is layers 20 to 23 of the neural network confidence analysis module.
The above advertisement harassment fax image detection system based on the convolutional neural network is further preferably: the first convolution module, the second convolution module, the third convolution module, the fourth convolution module, and the fifth convolution module include a convolution layer and a pooling layer, respectively.
The detection method of the advertisement harassment fax image detection system based on the convolutional neural network comprises the following steps:
the method comprises the following steps: training the convolutional neural network of the neural network confidence coefficient analysis module (training is completed only before the keyword suspicious region is extracted for the first time);
step two: extracting the keyword suspicious region;
step three: and using the trained neural network confidence coefficient analysis module to identify and judge the keyword suspicious region.
The detection method as described above, more preferably: the neural network feature classifier sets keyword confidence coefficients for the keyword suspicious regions, and when characters of the keyword suspicious regions are judged to be keyword sequences, the fax images to be detected are judged to be advertisement harassment images, and otherwise, the fax images are normal images.
The detection method as described above, more preferably: in the second step, the binarization module carries out image binarization operation based on OTSU, divides the fax image to be detected into a background and a target, and obtains an optimal binarization threshold value; and the morphological corrosion module is used for compressing a target based on morphological corrosion, compressing pixels of the target and judging the suspicious region of the keyword by adopting the optimal binarization threshold value.
The detection method as described above, more preferably: the color of the background is set to 255 and the color of the object is set to 0.
The detection method as described above, more preferably: in step three, the input layer inputs the keyword suspicious region, the first convolution module, the second convolution module, the third convolution module, the fourth convolution module and the fifth convolution module sequentially process the keyword suspicious region, and the neural network feature classifier judges the keyword suspicious region.
Analysis shows that compared with the prior art, the invention has the advantages and beneficial effects that:
1. the advertisement harassment fax image detection system based on the convolutional neural network has the advantages that through the arrangement of the keyword area extraction module, the automation degree is high, manual operation is not needed, and the working efficiency is high; by arranging the neural network confidence coefficient analysis module, the fax image to be detected can be quickly identified, time is saved, and the control capability is strong, so that the method has the characteristics of high working efficiency and strong control capability.
2. The advertisement harassment fax image detection system based on the convolutional neural network can effectively judge the keyword suspicious region by arranging the binarization module, can reduce the number of pixels to be matched by arranging the morphological corrosion module, has high processing speed, can sequentially process the keyword suspicious region by dividing the neural network confidence coefficient analysis module, and has strict logic, so that the advertisement harassment fax image detection system based on the convolutional neural network has the characteristics of high processing speed and strict logic.
3. The advertisement harassment fax image detection system based on the convolutional neural network provided by the invention adopts image binarization operation based on OTSU and target compression based on morphological corrosion to perform preprocessing when extracting the keyword suspicious region, can improve the character recognition accuracy, adopts ReLU as an activation function to process the keyword suspicious region, and adopts the convolutional neural network to classify the fax images to be detected, so that the system has the characteristics of high efficiency and high accuracy.
Drawings
FIG. 1 is a schematic diagram of the advertisement harassment fax image detection system based on the convolutional neural network.
FIG. 2 is a diagram illustrating the effect of the OTSU-based image binarization operation of the present invention.
FIG. 3 is a schematic diagram of the morphological etching-based method of the present invention.
FIG. 4 is a schematic diagram of a neural network confidence analysis module according to the present invention.
FIG. 5 is a schematic diagram of a fax image to be detected before being subjected to morphological etching treatment.
FIG. 6 is a schematic view of the effect of the morphological etching treatment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
As shown in fig. 1, the invention provides an advertisement harassment fax image detection system based on a convolutional neural network, which comprises a keyword region extraction module, a keyword image detection module and a keyword image processing module, wherein the keyword region extraction module is used for determining a keyword suspicious region of a fax image to be detected; and the neural network confidence coefficient analysis module is connected with the keyword region extraction module and is used for classifying the characters in the keyword suspicious region.
Specifically, the invention can divide and extract the fax image to be detected by setting the keyword region extraction module to obtain the keyword suspicious region, and the neural network confidence degree analysis module can detect and judge the keyword suspicious region. By arranging the keyword area extraction module, the automation degree is high, manual operation is not needed, and the working efficiency is high; by arranging the neural network confidence coefficient analysis module, the fax image to be detected can be quickly identified without manual identification, the time is saved, and the control capability is strong, so that the method has the characteristics of high working efficiency and strong control capability.
In order to further improve the working efficiency of the invention, as shown in fig. 1, the keyword region extraction module of the invention comprises a binarization module and a morphological corrosion module; the binarization module is used for judging the suspicious region of the keyword; the morphological corrosion module is connected with the binarization module and is used for corroding the judged keyword suspicious region. Aiming at the characteristics of less color information, simple background and relatively important pixel geometric position of the fax image, the keyword suspicious region can be effectively judged by arranging the binarization module; by arranging the morphological corrosion module, the number of pixels to be matched can be reduced and the processing time can be shortened on the premise that the geometric skeleton of the keyword suspicious region is not changed, so that the method has the characteristic of high working efficiency.
In order to further improve the management and control capability of the present invention, as shown in fig. 1, the neural network confidence level analyzing module of the present invention includes an input layer, a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, a fifth convolution module, and a neural network feature classifier. According to the invention, the neural network confidence coefficient analysis module is divided, so that the keyword suspicious regions can be sequentially processed, the logic is strict, and the condition of misjudgment of the keyword suspicious regions can be effectively avoided, so that the method has the characteristic of strong control capability.
In order to distribute the work of the neural network confidence coefficient analysis module of the present invention, as shown in fig. 1, the neural network confidence coefficient analysis module of the present invention includes 23 layers, and the input layer is the 1 st layer of the neural network confidence coefficient analysis module; the first convolution module is the 2 nd to 4 th layers of the neural network confidence coefficient analysis module; the second convolution module is the 5 th to 7 th layers of the neural network confidence coefficient analysis module; the third convolution module is the 8 th to 11 th layers of the neural network confidence coefficient analysis module; the fourth convolution module is the 12 th to 15 th layers of the neural network confidence coefficient analysis module; the fifth convolution module is the 16 th layer to the 19 th layer of the neural network confidence analysis module; the neural network feature classifier is the 20 th to 23 th layers of the neural network confidence analysis module. According to the invention, the neural network confidence coefficient analysis modules are reasonably divided, so that the work of each layer of the neural network confidence coefficient analysis modules can be reasonably distributed, and the work efficiency is improved, so that the neural network confidence coefficient analysis system has the characteristic of high work efficiency.
In order to simplify the parameter operation of the present invention, as shown in fig. 1, the first convolution module, the second convolution module, the third convolution module, the fourth convolution module and the fifth convolution module of the present invention respectively include a convolution layer and a pooling layer. According to the invention, the convolution layer and the pooling layer are arranged on each convolution module, so that the complexity of the operation model is greatly simplified, and the parameters of the model are reduced, so that the method has the characteristic of simple and convenient parameter operation.
As shown in fig. 1 to 6, the present invention further provides a detection method of an advertisement harassment fax image detection system based on a convolutional neural network, comprising the following steps:
the method comprises the following steps: and training the convolutional neural network of the neural network confidence coefficient analysis module.
1.1, obtaining a preliminary training model:
selecting a data set containing 1000 Chinese characters, wherein each Chinese character has 256 images, the images are collected from different handwritten texts, and performing normalization processing to obtain a preliminary training model.
1.2 Secondary training of the model:
on the basis of 1.1, a transfer learning mode is adopted for secondary training. And manually cutting a plurality of sub-images A, B, C, D and the like from the title position of the artificially determined advertisement harassment fax image, wherein each sub-image is a character. For each sub-image, rotate 3 times clockwise 45 degrees and turn each position horizontally to get 8 × 2 images. Randomly choose a size between [224,386], and resize each image to this size, at which crop a region of size (224 ) is taken as input.
1.3 obtaining a complete convolutional neural network training set:
and (3) randomly disordering the training set consisting of the 1.2, calculating the mean value of images of the training set, and performing mean value removing operation, wherein the processed result is a complete training set of the convolutional neural network.
The convolutional neural network needs to perform mean removal processing on a training set in the training process, and does not need to perform mean removal processing in the testing process and the system deployment process. The size of the batch for network training is set to 64, and an SGD (Stochastic gradient descent) algorithm is adopted. Momentum was set to 0.9, maximum number of iterations was 10 ten thousand, initial learning rate was 0.05, and attenuation was 0.0005 every 1 ten thousand. The L2 regularization coefficient of L2 regularization is 0.1, a dropout mode is adopted to reduce overfitting, and the probability is set to be 0.5. The network description is written into a prototxt file, the training parameters are written into a solvent.
Step two: and extracting the keyword suspicious region.
2.1 image binarization operation based on OTSU:
the binarization module carries out image binarization operation based on OTSU, an optimal binarization threshold value is selected from the fax image to be detected by adopting a maximum inter-class variance method (OTSU), the fax image to be detected is divided into a background and a target, the target color is set to be 0 (black), and the background color is set to be 255 (white), so that binarization is achieved, the difference between the target and the background in the same class is minimum, and the difference between different classes is maximum. The optimal binarization threshold value solving method comprises the following steps:
assuming that f (x, y) represents the gray level at the pixel point (x, y), and g (x, y) represents the average of the gray levels of the 3 × 3 neighborhood of (x, y), g (x, y) can be expressed as:
setting the size of the fax image to be detected as M multiplied by N, x + M is more than or equal to 0 and less than or equal to M-1, y + N is more than or equal to 0 and less than or equal to N-1, making i ═ f (x, y) and j ═ g (x, y) form a vector (i, j), and setting CijRepresenting the number of occurrences of (i, j), the probability of occurrence of (i, j) is:
setting a binarization threshold (s, t) to segment the fax image to be detected into a target and a background, wherein the probability of the occurrence of the two parts is as follows:
the gray level mean vector is:
the mean vector on the two-dimensional histogram is:
the between-class variance is:
for the background and the target, the similarity degree of the two classes is described by using the inter-class variance delta, wherein the larger delta is, the larger the difference between the background and the target is, and the smaller the difference is. If the background region is classified as the target region or the target region is classified as the background region, the inter-class variance δ is reduced. Therefore, when the inter-class variance of the target and the background is the largest and the intra-class variance is the smallest, the obtained segmentation result has the best effect. And when the delta is the maximum value, the corresponding (s, t) value is the optimal binarization threshold value. The effect of the treatment is shown in figure 2.
By adopting the image binarization operation based on the OTSU (maximum inter-class variance method), the method can ensure that the difference between the same class and the background is minimum, the difference between different classes is maximum, and the method has the best effect when the inter-class variance in the fax image to be detected is maximum. Otsu is carried out on the basis of a histogram of a fax image to be detected, and the operation is simple and rapid, so that Otsu segmentation efficiency is high, and the application is wide.
2.2 target compression based on morphological erosion:
the morphological erosion module is used for compressing the target based on the morphological erosion, and compressing the pixel of the target, wherein the mathematical expression of the morphological erosion is as follows:
i.e. moving structure B, if the intersection of structure B and structure a is completely within the area of structure a, the location point is saved, and all location points satisfying the condition constitute the result of the corrosion of structure a by structure B. When B is the structure shown in fig. 3, a compression effect is exerted on a, the processing principle is as shown in fig. 3, and the processing effect is as shown in fig. 6 before the morphological etching treatment is performed, as shown in fig. 5. On the premise that the geometric skeleton of the target is not changed, the morphological corrosion can extract the skeleton of the connected region of the target, and the number of pixels needing to be matched is reduced, so that the calculation amount of matching operation is reduced, and the calculation time is shortened.
2.3 extracting keyword suspicious regions:
setting a keyword set to be identified as { A, B, C and D }, randomly selecting a fax image to be detected from each set element as a region template, processing the fax image to be detected in step 1.1 and step 1.2, solving all the templates to obtain a keyword suspicious region, completing segmentation of the fax image to be detected, extracting the keyword suspicious region and sending the keyword suspicious region to a neural network confidence coefficient analysis module for classification. The facsimile image belongs to one of character images and has the characteristics of less color information, simple background and relatively important pixel geometric position information. The accuracy rate of character recognition can be improved by carrying out the step 1.1 and the step 1.2.
Step three: and identifying and judging the keyword suspicious region.
The neural network confidence degree analysis module adopts the ReLU as an activation function to process the keyword suspicious region, and adopts an SOFTMAX algorithm to classify the fax images to be detected. As shown in fig. 4, the identification determination process includes: INPUT → [ CONV × m → POOL ] × 5 → FC × 3 → SOFTMAX, where INPUT is the INPUT layer, CONV is the convolutional layer, ReLU is the nonlinear activation function, POOL is the pooling layer, and FC is the fully connected layer.
3.1 entering keyword suspicious region:
the layer 1 of the neural network confidence coefficient analysis module is an input layer, and the input layer is used for inputting keyword suspicious regions.
3.2, processing the keyword suspicious region:
the 2 nd, 3 rd and 4 th layers of the neural network confidence coefficient analysis module are first convolution modules, wherein the 2 nd and 3 th layers are convolution layers, each layer is provided with 64 convolution kernels with the size of 3 multiplied by 3, the minimum perception visual field size of the directional concepts in the upper, lower, left and right directions can be ensured, the stepping value is 1, and the convolution operation adopts a zero filling mode to ensure the spatial resolution. The 4 th layer is a pooling layer with the size of 2 multiplied by 2 and the step value of 2, and a mode of maximum down-sampling is adopted.
The 5 th, 6 th and 7 th layers of the neural network confidence coefficient analysis module are second convolution modules, the number of convolution kernels of the 5 th and 6 th layers is 128, the reduction of spatial resolution caused by down-sampling can be compensated, and the 7 th layer of pooling is the same as the previous layer.
The 8 th, 9 th, 10 th and 11 th layers of the neural network confidence coefficient analysis module are third convolution modules, the 8 th, 9 th and 10 th layers are convolution layers, the number of convolutions is increased to 256, and compared with the previous method, the 11 th layer of pooling layer is the same as the previous method.
The 12 th, 13 th, 14 th and 15 th layers of the neural network confidence analysis module are fourth convolution modules, and the 16 th, 17 th, 18 th and 19 th layers are fifth convolution modules. The fourth convolution module and the fifth convolution module are the same in structure as the third convolution module.
In the convolutional neural network of the neural network confidence analysis module, one neuron is connected with only part of neurons. In a convolution layer of the convolution neural network, a plurality of characteristic planes are contained, each characteristic plane is composed of a plurality of neurons arranged in a rectangular mode, and the neurons of the same characteristic plane share convolution kernels. The convolution kernel is initialized in the form of a random matrix, and the convolution kernel learns to obtain a reasonable weight in the training process of the network. The shared convolution kernel can reduce the connection between each layer of the network and reduce the risk of overfitting. The convolutional neural network reduces the number of parameters by adopting a mode of sharing a local perception field and a convolutional kernel, only needs to perceive the local part of the fax image to be detected, and then synthesizes the local information at a higher layer, so that the global information of the fax image to be detected is obtained, a certain part of characteristics of the fax image to be detected can be used on the other part of the characteristics, and the same learning characteristics can be used for all positions on the fax image to be detected.
3.3 outputting the identification result of the fax image to be detected:
the 20 th layer, the 21 th layer, the 22 th layer and the 23 th layer of the neural network confidence coefficient analysis module are neural network feature classifiers, the 20 th layer and the 21 th layer both contain 4096 parameter units for placing the keyword suspicious regions processed in the step 2.2, the 22 th layer is provided with keyword confidence coefficients, and the 23 th layer adopts a softmax algorithm to output a fax image detection result to be detected. And when the characters of the keyword suspicious region are judged as a keyword sequence by the layer 22, the layer 23 judges that the fax image to be detected is an advertisement harassing image, and otherwise, the fax image to be detected is a normal image.
The neural network confidence analysis module replaces the single-layer large convolution kernel (1@7 x 7) with a multi-layer small convolution kernel (e.g., 3@3 x 3). Along with the increase of the depth, the recognition rate is obviously improved, the working efficiency is high, and the control capability is strong.
Analysis shows that compared with the prior art, the invention has the advantages and beneficial effects that:
1. the advertisement harassment fax image detection system based on the convolutional neural network has the advantages that through the arrangement of the keyword area extraction module, the automation degree is high, manual operation is not needed, and the working efficiency is high; by arranging the neural network confidence coefficient analysis module, the fax image to be detected can be quickly identified, time is saved, and the control capability is strong, so that the method has the characteristics of high working efficiency and strong control capability.
2. The advertisement harassment fax image detection system based on the convolutional neural network can effectively judge the keyword suspicious region by arranging the binarization module, can reduce the number of pixels to be matched by arranging the morphological corrosion module, has high processing speed, can sequentially process the keyword suspicious region by dividing the neural network confidence coefficient analysis module, and has strict logic, so that the advertisement harassment fax image detection system based on the convolutional neural network has the characteristics of high processing speed and strict logic.
3. The advertisement harassment fax image detection system based on the convolutional neural network provided by the invention adopts image binarization operation based on OTSU and target compression based on morphological corrosion to perform preprocessing when extracting the keyword suspicious region, can improve the character recognition accuracy, adopts ReLU as an activation function to process the keyword suspicious region, and adopts the convolutional neural network to classify the fax images to be detected, so that the system has the characteristics of high efficiency and high accuracy.
It will be appreciated by those skilled in the art that the invention may be embodied in other specific forms without departing from the spirit or essential characteristics thereof. The embodiments disclosed above are therefore to be considered in all respects as illustrative and not restrictive. All changes which come within the scope of or equivalence to the invention are intended to be embraced therein.
Claims (3)
1. An advertisement harassment fax image detection system based on a convolutional neural network, comprising:
the keyword area extraction module is used for determining a keyword suspicious area of the fax image to be detected;
the neural network confidence coefficient analysis module is connected with the keyword region extraction module and is used for classifying characters in the keyword suspicious region;
the keyword region extraction module comprises a binarization module and a morphological corrosion module; the binarization module is used for judging the keyword suspicious region; the morphological corrosion module is connected with the binarization module and is used for corroding the judged keyword suspicious region;
the neural network confidence coefficient analysis module comprises an input layer, a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, a fifth convolution module and a neural network feature classifier;
the neural network confidence coefficient analysis module comprises 23 layers, and the input layer is the layer 1 of the neural network confidence coefficient analysis module; the first convolution module is layers 2 to 4 of the neural network confidence analysis module; the second convolution module is layers 5 to 7 of the neural network confidence analysis module; the third convolution module is the 8 th to 11 th layers of the neural network confidence coefficient analysis module; the fourth convolution module is layers 12 to 15 of the neural network confidence analysis module; the fifth convolution module is layers 16 to 19 of the neural network confidence analysis module; the neural network feature classifier is layers 20 to 23 of the neural network confidence analysis module;
the first convolution module, the second convolution module, the third convolution module, the fourth convolution module, and the fifth convolution module include a convolution layer and a pooling layer, respectively;
the system is realized by a method comprising the following steps:
the method comprises the following steps: training the convolutional neural network of the neural network confidence coefficient analysis module;
1.1, obtaining a preliminary training model;
selecting a data set comprising a preset number of Chinese characters, wherein each Chinese character comprises a preset number of images, the images are acquired from different handwritten texts, and normalization processing is carried out to obtain a preliminary training model;
1.2, performing secondary training on the model;
on the basis of 1.1, performing secondary training in a transfer learning mode; manually cutting a plurality of subimages A, B, C, D and the like from the title position of the artificially determined advertisement harassment fax image, wherein each subimage is a character; for each sub-image, rotating for 3 times at 45 degrees clockwise, and horizontally turning over each position to obtain 8 multiplied by 2 images; randomly selecting a size between [224,386] and resizing each image to that size, obtaining as input a region of size (224 );
the batch size of the convolutional neural network training is set to be 64, and a random gradient descent algorithm is adopted, wherein the momentum is set to be 0.9, the maximum iteration time is 10 ten thousand times, the initial learning rate is 0.05, and the attenuation is 0.0005 every 1 ten thousand times; the L2 regular coefficient of the L2 regularization is 0.1, a dropout mode is adopted for reducing overfitting, and the probability is set to be 0.5;
1.3 obtain the complete convolutional neural network training set:
randomly disorganizing the training set consisting of the 1.2, calculating the mean value of images of the training set, and performing mean value removing operation, wherein the processed result is a complete training set of the convolutional neural network;
step two: extracting the keyword suspicious region;
2.1 image binarization operation based on OTSU:
the binarization module carries out image binarization operation based on OTSU, selects an optimal binarization threshold value in the fax image to be detected by adopting a maximum inter-class variance method (OTSU), divides the fax image to be detected into a background and a target, sets the target color to be 0 (black) and the background color to be 255 (white), thereby achieving binarization, minimizing the difference of the same class between the target and the background and maximizing the difference between different classes;
2.2 target compression based on morphological erosion;
2.3 extracting keyword suspicious regions:
setting a keyword set to be identified as { A, B, C, D }, randomly selecting a fax image to be detected from each set element as a region template, and solving all templates to obtain a keyword suspicious region after processing in step 1.1 and step 1.2;
step three: and using the trained neural network confidence coefficient analysis module to identify and judge the keyword suspicious region.
2. The convolutional neural network-based advertisement harassment fax image detection system as claimed in claim 1, wherein:
the neural network feature classifier sets keyword confidence coefficients for the keyword suspicious regions, and when characters of the keyword suspicious regions are judged to be keyword sequences, the fax images to be detected are judged to be advertisement harassment images, and otherwise, the fax images are normal images.
3. The advertisement harassment fax image detection system based on the convolutional neural network as claimed in claim 2, characterized in that:
in step three, the input layer inputs the keyword suspicious region, the first convolution module, the second convolution module, the third convolution module, the fourth convolution module and the fifth convolution module sequentially process the keyword suspicious region, and the neural network feature classifier judges the keyword suspicious region.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810150076.0A CN108460772B (en) | 2018-02-13 | 2018-02-13 | Advertisement harassment fax image detection system and method based on convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810150076.0A CN108460772B (en) | 2018-02-13 | 2018-02-13 | Advertisement harassment fax image detection system and method based on convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108460772A CN108460772A (en) | 2018-08-28 |
| CN108460772B true CN108460772B (en) | 2022-05-17 |
Family
ID=63216529
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810150076.0A Active CN108460772B (en) | 2018-02-13 | 2018-02-13 | Advertisement harassment fax image detection system and method based on convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108460772B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111144399B (en) * | 2018-11-06 | 2024-03-05 | 富士通株式会社 | Apparatus and method for processing image |
| CN112561055B (en) * | 2020-12-09 | 2023-11-07 | 北京交通大学 | Electromagnetic disturbance identification method based on bilinear time-frequency analysis and convolutional neural network |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105574524A (en) * | 2015-12-11 | 2016-05-11 | 北京大学 | Cartoon image page identification method and system based on dialogue and storyboard united identification |
| US9501724B1 (en) * | 2015-06-09 | 2016-11-22 | Adobe Systems Incorporated | Font recognition and font similarity learning using a deep neural network |
| CN106156777A (en) * | 2015-04-23 | 2016-11-23 | 华中科技大学 | Textual image detection method and device |
| CN106650721A (en) * | 2016-12-28 | 2017-05-10 | 吴晓军 | Industrial character identification method based on convolution neural network |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104346622A (en) * | 2013-07-31 | 2015-02-11 | 富士通株式会社 | Convolutional neural network classifier, and classifying method and training method thereof |
| US10762894B2 (en) * | 2015-03-27 | 2020-09-01 | Google Llc | Convolutional neural networks |
| US9858496B2 (en) * | 2016-01-20 | 2018-01-02 | Microsoft Technology Licensing, Llc | Object detection and classification in images |
| CN106096602A (en) * | 2016-06-21 | 2016-11-09 | 苏州大学 | A kind of Chinese licence plate recognition method based on convolutional neural networks |
| CN107622272A (en) * | 2016-07-13 | 2018-01-23 | 华为技术有限公司 | A kind of image classification method and device |
| CN107437100A (en) * | 2017-08-08 | 2017-12-05 | 重庆邮电大学 | A kind of picture position Forecasting Methodology based on the association study of cross-module state |
| CN107578060B (en) * | 2017-08-14 | 2020-12-29 | 电子科技大学 | Method for classifying dish images based on depth neural network capable of distinguishing areas |
-
2018
- 2018-02-13 CN CN201810150076.0A patent/CN108460772B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106156777A (en) * | 2015-04-23 | 2016-11-23 | 华中科技大学 | Textual image detection method and device |
| US9501724B1 (en) * | 2015-06-09 | 2016-11-22 | Adobe Systems Incorporated | Font recognition and font similarity learning using a deep neural network |
| CN105574524A (en) * | 2015-12-11 | 2016-05-11 | 北京大学 | Cartoon image page identification method and system based on dialogue and storyboard united identification |
| CN106650721A (en) * | 2016-12-28 | 2017-05-10 | 吴晓军 | Industrial character identification method based on convolution neural network |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108460772A (en) | 2018-08-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11783579B2 (en) | Hyperspectral remote sensing image classification method based on self-attention context network | |
| CN108562589B (en) | Method for detecting surface defects of magnetic circuit material | |
| CN111753828B (en) | Natural scene horizontal character detection method based on deep convolutional neural network | |
| CN112966691B (en) | Multi-scale text detection method and device based on semantic segmentation and electronic equipment | |
| Pastor-Pellicer et al. | Insights on the use of convolutional neural networks for document image binarization | |
| WO2017113232A1 (en) | Product classification method and apparatus based on deep learning | |
| WO2018052586A1 (en) | Method and system for multi-scale cell image segmentation using multiple parallel convolutional neural networks | |
| CN107609549A (en) | The Method for text detection of certificate image under a kind of natural scene | |
| Prasad et al. | An efficient classification of flower images with convolutional neural networks | |
| CN109685065B (en) | Layout analysis method and system for automatically classifying test paper contents | |
| Burie et al. | ICFHR2016 competition on the analysis of handwritten text in images of balinese palm leaf manuscripts | |
| CN109033978B (en) | Error correction strategy-based CNN-SVM hybrid model gesture recognition method | |
| CN110717497B (en) | Image similarity matching method, device and computer readable storage medium | |
| CN109740539B (en) | 3D object identification method based on ultralimit learning machine and fusion convolution network | |
| Shanta et al. | Bangla sign language detection using sift and cnn | |
| US20140079316A1 (en) | Segmentation co-clustering | |
| CN108460772B (en) | Advertisement harassment fax image detection system and method based on convolutional neural network | |
| Lee | Image classification with artificial intelligence: cats vs dogs | |
| Yeh et al. | Intelligent mango fruit grade classification using alexnet-spp with mask r-cnn-based segmentation algorithm | |
| Nandhini et al. | Sign language recognition using convolutional neural network | |
| CN111160372A (en) | Large target identification method based on high-speed convolutional neural network | |
| CN111666813A (en) | Subcutaneous sweat gland extraction method based on three-dimensional convolutional neural network of non-local information | |
| CN110795995A (en) | Data processing method, device and computer readable storage medium | |
| Varkentin et al. | Development of an application for car license plates recognition using neural network technologies | |
| CN115810106A (en) | Tea tender shoot grade accurate identification method in complex environment |
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 |