Feed Forward Neural Network With Fuzzy C

Published Date: 02 Nov 2017

Abstract- Due to the rapid growth of image databases in almost every field like medical science, multimedia, geographical information system, photography, Journalism, etc., a effective and efficient technique is required to process the image .Content-based image retrieval is a technique which is used to extract the images on the basis of their content such as texture, color, shape and spatial layout. However it’s very difficult to retrieve the image due to the semantic gap between the user’s high level concepts and low-level features of the image. In this paper a neural network based approach is proposed for texture finding in CNIR. In our environment, user submit the image, then it’s the responsibility of the CBIR system to detects the texture parameters within the query image and search the query image using the feed forward Neural network approach. Basically it completely removes the gap between low-level and high-level features. CBIR system also models the human perception by providing the relevance feedback.

Keyword: Content Based Image Retrieval, Neural Network, Relevance Feedback, Texture, Feed Forward Neural Network.

INTRODUCTION

With the advancement of technology, computational power and reduction in the price of memory, there is a need to switch from the traditional approach to the new advancements. Maintenance of large database is very crucial and especially when large collection of digital images need to be maintained. In many areas where digitized images are required like Journalism, Hospitals, academia, multimedia, geographical information system, crime prevention, pattern recognition, statistics, fashion, architecture and many more. Thus over the decade the volume of digital image is increasing very exponentially.

Traditional approach of searching the image was by indexing or simply by browsing. Problem with the historical approach

leads to another way of accessing the image on the basis of their content or feature. Thus Content Based Image Retrieval (CBIR) is defined as a process of searching a digital image from the large database on the basis of their visual features like shape, color and texture. Now there is no need to apply the indexing and images can be fetched in an effective and efficient manner. This reduces the semantic gap between the low level visual features and high level semantic features. Image database is different from the traditional approach. It is a two-step process:

Image Extraction: In this step image is extracted based on the image feature, every image is based on its pixel values.

Matching Step: In this step, the features extracted from the query are matched with the features of the images stored in the database and identify the visually similar images.

Various algorithms were proposed for the extraction of similar images on the basis of features of the images using various techniques.

WHAT IS CBIR

Content Based Image Retrieval is an application of computer vision where digitally similar images are retrieved from the large database on the basis of their content. Content in this context refer to the Information that describes the image like color, texture, and shapes. The detailed survey on content based image retrieval can be referred [1, 2, 22].

2.1. Color Based Retrieval: The Basic technique which is used is based on the technique of color histogram. Color Histogram of each image is calculated and then stored in the database which represents the proportion of pixel of each color within the image. Then matching algorithm will extract those images from the databases whose color histogram matches with the required one. There are various types of histograms: normal, weighted, dominant, and fuzzy, various color spaces: HSV, grayscale, HSL, Lab, Luv, HMMD, and YCbCr.

Figure 1: Generic CBIR System

2.2 Texture Based Retrieval: This is a very important characteristic of an image because it is able to distinguish two images with same color and shape. Variety of techniques has been proposed for matching the texture similarity. Tamura et al. [12], proposed a texture representation on 6 statistical features, including, contrast, coarseness, line-likeness, directionality, regularity, and roughness. These features were considered to be the most visually meaningful. Various techniques designed for texture feature extractions are: statistical parameters, entropy measures, transformed spaces and Markov Hidden Fields algorithms.

2.3 Shape Based Retrieval: This is a well defined term, which refers to the shape of the image. This is feature which naturally distinguishes the images. There are two main features of the shape: Global feature (like aspect ratio) and local feature (like boundary segments). Shape of an image can be represented using area, perimeter, radiuses, skeleton, statistics moments, form signature, Fourier and Hough contour signature.

APPLICATIONS

Scientific Databases.

Art Museum.

Medical Science databases.

Collection of Photographs.

Retail Market Catalogs.

Crime Detection & Prevention.

The military.

Historical Research.

Publishing & advertising.

Fashion & Graphic Design.

Remote sensing.

Education & Training.

Geometrical Information Systems.

Architectural & engineering design.

World Wide Web, etc.

FUZZY LOGIC BASED APPROACH

There are various algorithms was proposed to implement the content based image retrieval. Most of the designing of the system is based on the concept of fuzziness of the image. As human perception vary from one to another it’s very vague to describe the content of the image thus fuzzy is best suitable to implement CBIR. Even more, fuzzy takes the crisp decision at the end. Initially, in [6] 1976 William B. Thompson proposed the method for texture boundary analysis on the basis of dissimilarity of region based on various textural properties. It applies the textural boundary operator to various patterns and then constructs the edge map that identifies the boundaries within the image based on the textural difference.

After that in 1999, Swarup Medasani et.al, [8] designed a Java based system that represent the images using Fuzzy Attributed Relational Graphs. These FARG are stored in the database in groups (clusters) using graph clustering algorithm(LCA FARG) to increase the retrieval speed , where each cluster is represented by a leader FARG such that every element in a cluster is within a distance T of a leader. A total of 149 images are collected from the Vistex database of MIT Media Lab and 500 images are used from NETRA database and the test is conducted on these images. This method also introduced the relevance feedback concept. This method also has some limitations that the segmentation of images is done manually; further relevance feedback concept was only introduced but has not been integrated with it.

In [3], proposed the fuzzy logic approach using the Tamura features for texture feature based extraction of image. Tamura is based on Psychological studies of human perception. They defined six different meaningful properties of texture â€" contrast, coarseness, line-likeness, directionality, regularity, and roughness. Here, coarseness measures the texture scale; contrast measure vividness of the texture, directionality gives the main direction of the image texture.

With the help of fuzzy clustering, the term set is generated where each tamura feature is represented in linguistic terms. Here, query can be framed as a logic combination of the natural language terms & tamura feature values.

In [4], fuzzy logic based approach was used to reduce the semantic gap between low-level features to high level textures. It uses unsupervised fuzzy clustering approach to represent the tamura features in term set which is represented in natural language. User query is framed as a logic combination of the natural language terms & tamura feature values. Min-max composition rules are used to calculate the distance between user query and the images from the database.

Herba & Neamat [5] provides a new approach for graph matching that resembles the human thinking process. Each image is represented by Fuzzy Attributed Relational Graph (FRAG) that describes each object in the image by all its attribute and spatial relation. This approach used to extract features of texture and color both and successfully implemented the Human Vision System (HSV) model.

In 2008, Wu Kai-xing and Xu Qiang [16], proposed Fuzzy color histogram based on L*a*b* color space component. Here L* stands for luminance, a* stands for relative greenness-redness and b* represents relative blueness-yellowness.

Further, a* and b* was further divided into 5 different regions and L* was also divided into three regions:

a* = [Green, Greenish, Middle, Reddish, Red]

b* = [Blue, Bluish, Middle, Yellowish, Yellow]

L* = [Black, Grey, White]

Now, Fuzzification method is applied using membership function applied on L*a*b*. The output of these membership functions are combined through aggregation operator and then defuzzification is applied.

The similarity ratio is calculated for identifying the similarity between the query image and images stored in the feature database. The precision of this method comes out to be 95% i.e. more accurate and robust approach.

A new image retrieval approach was proposed by Dr. B. Prabhakara Rao. et.al [9]. Using the three features named dynamic dominant color (DDC), Motif co-occurrence matrix (MCM) and difference between pixels of scan pattern (DBPSP) a high precision technique was proposed.

Dynamic dominant color is used because in a given color image, the number of actual color occupies very small proportion of the total color image. Thus dominant color descriptor (DCD) increases the precision. The algorithm used for selection of color space is based on RGB approach which is effective and simple.

Motif Co-occurrence matrix (MCM) is used to calculate the probability of the occurrence of same pixel color between each pixel and its adjacent ones in each image. Using the proposed method, this method is able to consider total 49 number of MCM attributes.

Further MCM specifies the direction of texture but not the complexity of the texture, thus DBSP approach is integrated with MCM to improve texture description.

The proposed method performance was evaluated on the image set consists of 1000 images where 10 clusters are formed based on the similarity and around 90 % precision was achieved.

Another approach was proposed by Wang Xiaoling et.al [10], which extracts the color and shape of the image of around 500 images. The experiment includes two parts: Firstly, color feature extraction using Average Area Histogram (AAH); secondly, shape feature extraction using fuzzy image retrieval.

Average area Histogram is an improvement over traditional histogram approach as it reflects the real color distribution over traditional histogram. Basically, traditional histogram is able to identifies the total number of pixel of each color but unable to identify the variations within the same color. Thus, Average area histogram (AAH) is used which uses the area feature of the regions formed by each image color.

In second phase, shape of an image can be identified by various methods. Here Wang Xiaoling et.al uses the moment invariants to describe a region.

Further, image similarity based on fuzzy logic can be identified using the different fuzzy rules that will reflect around 41 % precision i.e., good robustness.

Refer to [17], used the concept of Fuzzy hamming distance (FHD) for retrieving the image based on its content. The proposed system consists of three modules:

Preprocessing module: It basically extracts the information from the image. Then it segments the image and finally represents them in color histogram.

Similarity module: It takes the query image and tries to identify the similarity between them with the help of FHD concept. This method is able to distinguish the images with same color histogram but different semantic content.

Ranking module: It returns the result on the basis of raking arranged in decreasing order.

NEURAL NETWORK BASED APPROACH

In [13], proposed content based image retrieval using feed forward neural network approach. To implement this concept, images are clustered together with the help of K-means clustering algorithm and heuristic approach known as differential evaluation. Differential evaluation is a heuristic approach used to identify the value of cluster size i.e., k.

To retrieve the query image from the database, in RBF neural network model Gaussian function is applied on the hidden layer. With the help of Genetic algorithm, weights are adjusted iteratively for optimization.

Ionut Mironica and Radu Dogaru in [14] investigated and compare the various clustering based methods used for Content Based Image Retrieval. On this basis, best method is selected in terms of performance.

It covered four different classification methods, Decision Trees, Naïve Bayes, Support Vector Machines (SVM) and Radial Basis Function (RBF-M).

For the implementation and evaluation three different types of databases was maintained: Medical images (320 images), texture database (900 images) and natural database (2700 images).

In overall study, Naïve Bayes classification algorithm is more appropriate than other methods.

NEURO-FUZZY BASED APPROACH

The Samuel et.al. [7], proposed a fuzzy combined short-term and long term learning method to construct relevance feedback-based content based image retrieval. Firstly, applies fuzzy support vector machine (FSVM) based short-term learning technique. Here it uses the MPEG-7 150-bin edge histogram descriptor (EHD) and the 64-bin HSV-based scaled color descriptor (SCD) to extract the global low level features with the help of Euclidean distance method. Further it has divided the images into 5 different blocks and then applies the algorithm to identify the relevant and irrelevant blocks. Secondly, in long-term based high level image retrieval, clustering technique to group the images of same semantics and finally, we merge the result to compactly store the memorized feedback information using predictive algorithm to improve the retrieval result.

In [11] proposed an efficient approach for CBIR in order to remove the semantic gap between the low level and high level semantic for image. Rushikesh Borse et.al designed a framework for the implementation of CBIR using the feed forward neural network approach. This framework has been divided into two phases: Firstly, in offline processing all the images are stored with their features and in second phase i.e., online processing query image is processed with user feedback.

It tries to extract the color, texture and shape of the image. The color feature of the image is extracted using color histogram. Texture is calculated using Co-occurrence matrix and Wavelet Transform. Finally description of shape of the image is extracted using the erosion morph feature extraction technique.

Further, in order to reduce the semantic gap, it uses neural network approach using linguistic expression based image description (LEBID) framework as in Fig 2.

Refer to [21], proposed a technique based on fuzzy logic and neural network. The paper has proposed a novel approach to retrieve the features of the image including texture, color and shape with the help of natural language queries, fuzzy mapping of image database and fuzzy similarity distance. The Neural network approach is proposed to learn the meaning of fuzzy queries.

For conducting the experiment, images from various websites and images and for texture images are extracted from Brodatz album. This proposed method achieved 82% of the recall and 87% precision.

OTHER TECHNIQUES

In [15], proposed a technique for CBIR texture feature extraction using statistical texture features. Firstly it converts the RGB color image into gray scale single component image in order to reduce the computation. Further, grayscale image is divided into blocks like 2*2, 4*4, 8*8, 16*16, and 32*32.

Now statistical texture features including mean, standard deviation, skewness, flatness, energy, entropy and smoothness are calculated for each block and store these features into feature database. Now, the query image statistical features are calculated with the images stored in the feature database with the help of distance between them.

They used 1000 images, and average precision obtained as 61 and recall as 76.

Refer to [18], presents the concept of multiresolution multigrid framework. It extracts the local color and texture feature of the image by partioning the image into equal sized non-overlapping tiles. Now, features was drawn using conditional co-occurrence histograms computed using the image and its complement in RGB color space. An integrated matching scheme based on most significant highest priority (MSHP) principle, and adjacency matrix of a bipartite graph constructed between images tiles, is implemented for image similarity.

Shape of an image is extracted using the Gradient vector flow field (GVF). Invariant moments are used to represent the shape features. The experimental result shows 96% precision.

With reference to [19], proposed architecture to develop a distributed multimedia retrieval system named as CMRS. Timo Ojala et.al tries to develop an application which supports media independent platform and addition of new media type to the system. To achieve this, Timo Ojala et.al encapsulates the data, operation and user interface related to a particular media type or query entity into a single unit.

With the help of Self Organizing Map (SOM), we can efficiently visualize the database in a 2- dimensional view.

PROPOSED MODEL

In our CBIR system model, each texture image in the database is represented as six tamura features. Tamura feature represents the human perception. Tamura features can be easily interpreted using the statistical properties of textures. The six texture features are summarized as follows:

8.1) Coarseness: Coarseness is the most fundamental feature in texture analysis; it refers to texture granularity, that is, the size and number of texture primitives. A coarse texture contains a small number of large primitives, whereas a fine texture contains a large number of small primitives Coarseness (ð‘"𝑐𝑟𝑠) can be computed as follows:-

Where, n*n denotes the image size and k=1, 2, 3, 4 and 5.

8.2) Contrast: Contrast stands for image quality in the narrow sense; it refers the difference in intensity among neighboring pixels. A texture on high contrast has large difference in intensity among neighboring pixels, whereas a texture on low contrast has small difference. Contrast (ð‘"𝑐𝑜𝑛) can be computed as follows:-

Where, 𝜎 is the image standard deviation and 𝜇4 is fourth moment of the image.

8.3) Directionality: Directionality is a global property over a specific region; it refers the shape of texture primitives and their placement rule. A directional texture has one or more recognizable orientation of primitives, whereas an isotropic texture has no recognizable orientation of primitives.

In this, one can use histogram of local edge probabilities against their directional angle. This method utilizes the fact that gradient is a vector, so it has both magnitude and direction. In the discrete case, the magnitude Î"G and the local edge direction θ are approximated as follows:

Î"G = Î"H + Î"V

θ= tan−1 (Î"V/Î"H) + 𝜋/2

Where, Î"H and Î"V are the horizontal and vertical differences measured by following 3 x 3 operators, respectively,

[−1 0 1, −1 0 1, −1 0 1] [1 1 1, 0 0 0, −1 −1 −1]

The desired histogram can be obtained by quantizing θ and counting the points with the magnitude Î"G over the threshold t;

k= 0, 1, 2..........., n-1

Where, Nθ (k) is the number of points at which (2k-1) 𝜋 /2n ≤ 𝜃 < (2k+1) 𝜋 /2n and Î"G ≥ 𝑡

Where,

𝑛𝑝= Number of peaks

∅𝑝= pth peak position of 𝐻𝐷,

𝑤𝑝 = range of pth peak between valleys

8.4) Line-likeness: Line-likeness refers only the shape of texture primitives. A line-like texture has straight or wave-like primitives whose orientation may not be fixed. Often the line-like texture is simultaneously directional. In this, a direction co-occurrence Matrix is constructed whose element PDd (i , j) is defined as the relative frequency with which two neighboring cells separated by a distance d along the edge direction occur on the image, one with the direction code i and the other with the direction code j. This Line-likeness (Flin) can be computed as follows:-

Where, 𝑑(𝑖,𝑗) is the n * n local direction co-occurrence matrix of points at a distance d.

8.5) Regularity: Regularity refers to variations of the texture-primitive placement. A regular texture is composed of identical or similar primitives, which are regularly or almost regularly arranged. An irregular texture is composed of various primitives, which are irregularly or randomly arranged. Regularity (freg) can be computed as follows:-

Where, r is a normalizing factor and 𝜎𝑥𝑥𝑥 means the standard deviation of ð‘"𝑥𝑥𝑥. In this study, the normalizing factor r = 0.25.

8.6) Roughness: Roughness refers tactile variations of physical surface. A rough texture contains angular primitives, whereas a smooth texture contains rounded blurred primitives. Roughness (ð‘"𝑟ð‘") can be computed as follows:-

ð‘"𝑟ð‘"=ð‘"𝑐𝑟𝑠+ð‘"𝑐𝑜𝑛

Based on the different texture content of the image the grouping of images is done with the clustering technique i.e. Fuzzy C-means clustering. Fuzzy C-means clustering is a technique in which a dataset is clubbed into n clusters with every data element in the dataset belonging to every cluster to a certain degree. The FCM algorithm attempts to partition a finite collection of n elements X = {x1, …, xn}into a collection of c fuzzy clusters with respect to some given criterion. The algorithm returns a list of c cluster centres C = {c1, .., cc} and a partition matrix n, where each element wij tells the degree to which element xi belongs to cluster cj . The standard function is:

w_k(x) = \frac{1}{\sum_j \left(\frac{d(\mathrm{center}_k,x)}{d(\mathrm{center}_j,x)}\right)^{2/(m-1)}}.

The output given from grouping clusters is the initial input to our effective RBFNN. The main aim is to retrieve the much desired image on comparison with query image to that of database images.

FEED FORWARD NEURAL NETWORK APPROACH

RBFNNs, similarly to all Neural Networks, are associated with a set of parameters that need to be adjusted in order for the Neural Network to “learn” the correct mapping between inputs and outputs. The set of parameters of a Neural Network is directly dependent on the Neural Network’s architecture.

RBF networks have three layers:

Input layer â€" There is one neuron in the input layer for each predictor variable. In the case of categorical variables, N-1 neurons are used where N is the number of categories. The input neurons standardize the range of the values by subtracting the median and dividing by the interquartile range. The input neurons then feed the values to each of the neurons in the hidden layer.

Hidden layer â€" This layer has a variable number of neurons (optimal number is determined by the training process). Each neuron consists of a radial basis function centered on a point with as many dimensions as there are predictor variables. The spread (radius) of the RBF function may be different for each dimension. The centers and spreads are determined by the training process. When presented with the x vector of input values from the input layer, a hidden neuron computes the Euclidean distance of the test case from the neuron’s center point and then applies the RBF kernel function to this distance using the spread values. The resulting value is passed to the summation layer.

In the proposed RBFNN network model uses the activation function in hidden layer is Multiquadratics represented by:

Summation layer â€" The value coming out of a neuron in the hidden layer is multiplied by a weight associated with the neuron (W1, W2, ...,Wn) and passed to the summation which adds up the weighted values and presents this sum as the output of the network. In classification, there is one output (and a separate set of weights and summation unit) for each target category. The output value for a category is the probability that the case being evaluated has that category.

http://www.dtreg.com/RBFarchitecture.gif

Figure 2 : RBF-Neural Network Architecture.

Here we are going to filter images in the Fuzzy C-Clustering approach and then apply the clustered images to RBFNN network model, so that we get the better result.

Step-wise represented of proposed model:

Get the initial textured image from different conditions of image database.

Images are grouped according to clustering techniques.

Now the images got segmented images through filtering techniques.

Segmented image values are sent as input values to RBF Neural Network Model.

RBFNN provides the images that are very much more similar to the input image given as the retrieved output images.

CONCLUSION

Content Based Image retrieval System is immature and a research area. Initially it started with the indexing approach and then move towards the extraction of the image based on the content. This paper introduces new concept of Fuzzy C-means clustering approach with feed forward neural network approach which enhance the accuracy rate of the content based image retrieval system.