1. Introduction
Aontent Based Image Retrieval (CBIR) system effectively searches image database and retrieves particular images relevant to query image submitted by user. CBIR system focuses on visual content of images rather than manual annotation. Images are thus indexed by their individual visual features [1-2] such as color, texture and shape. An extensive variety of image descriptors are provided by MPEG7 [3] to extract these visual features. Selection of descriptors greatly affects performance of the CBIR system. From literature it is evident that extracting fixed set of features from images fails to give satisfactory results [4-5]. To overcome the problems in selecting feature sets, a CBIR system that adopts different feature sets for different classes was proposed in [6]. The work in [7] provides convincing retrieval results by using combined shape and color features with relevance feedback.
Retrieval efficiency can be improved by various methods [8-11]. Statistical model and radial basis function neural network has been presented in [8] for color image retrieval. In [9] a hybrid meta-heuristic swarm intelligence-based search method is used for image retrieval. Problems in conventional distance and classifier based retrieval approaches was addressed in [10] using a fuzzy class membership. Retrieval accuracy was improved by user oriented image retrieval system based on interactive genetic algorithm [11]. Fuzzy algorithms are extensively used for classification in pattern recognition applications like image retrieval. Adaptive Resonance Theory (ART) was developed by Grossberg [12]. This architecture possesses many desirable properties that can solve arbitrarily complex classification problems. Fuzzy Artmap (FAM) [13] which better classifies input patterns can be implemented as the matching machine of CBIR system. Extension and modification of FAM neural classifier proposed for face recognition systems in [14-15] has made effective feature classification.
Limitation faced by FAM is category proliferation problem reported in [16]. This refers to the case where FAM classifier creates unnecessarily large IF-THEN rules while dealing with large and noisy data sets. Distributed Artmap that combines computational advantages of Multi Layer Perceptron (MLP) and Adaptive Resonance Theory (ART) systems in real time neural network for supervised learning is presented in [17] and its performance was verified with a circle in square problem. A hybrid neural network classifier of FAM and Dynamic Decay Adjustment (DDA) algorithm is presented in [18]. The algorithm proves to decrease misclassification rate. Evidence in [19] explores that Boosted Artmap as a viable learning technique over FAM. Safe μArtmap described in [20] limits the growth of a category in response to a single pattern, so that large hyperboxes are not created. In [21], number of modifications is implemented in FAM to reduce the category proliferation problem caused due to overlapping classes. This paper proposes a modified FAM known as Non Proliferation Fuzzy Artmap (NPFAM) which focuses on avoiding proliferation problem. Once low level features are extracted using MPEG7 descriptors, they are fused into NPFAM classifier for training process. During retrieval the classifier matches a class for given query image. After system identifies the class of query image, similarity between images in the class and query image is calculated using Euclidean distance similarity metric. Images in the class with higher similarity are presented to user.
The organization of this paper is as follows: Section 2 presents details of descriptors used in feature set. Section 3 describes FAM classifier. Section 4 gives details about NPFAM classifier. Section 5 provides experimental data sets used and performance comparison between FAM and NPFAM. Finally Section 6 gives conclusion and future work.
2. Feature Space and Formation of Feature Space
Two prominent steps involved in CBIR systems are indexing and searching. Indexing computes essential features that describe the image to its best and stores them in database as feature space. Searching computes feature vectors for query image, compares them with feature vectors of images in database and images which are most similar to query image are presented to the user. This section provides brief details of MPEG7 visual content descriptors included in feature space.
2.1 Visual Descriptors
In literature, a variety of descriptors has been discussed for CBIR systems. MPEG7 have formed four types of visual descriptors namely i) Movement descriptor ii) Texture descriptor iii) Color descriptor and iv) Shape descriptor. Detailed studies of these descriptors are presented in [22]. Descriptors used for indexing in the proposed system are briefed below.
2.1.1 Color Descriptor
Color descriptors originating from histogram analysis have played a central role in the development of visual descriptors in MPEG7.
Color Layout Descriptor (CLD) is a spatial distribution of colors in compressed form of a segmented or complete image. Compactness of descriptor provides better retrieval results and supports sketch based retrieval. Representative colors on a sliding 8X8 grid are encoded by DCT and used in CLD. YCrCb color space is adopted and contributes 192 coefficients (64Y, 64Cr, 64Cb) in the feature set which is used to index images.
Scalable Color Descriptor (SCD) defined in HSV color space with a uniform quantization to 256 bins, addresses interoperability issue available in generic color histogram descriptor. SCD is a color histogram encoded by Haar transform. Full interoperability between different resolutions of color representation is achieved by ranging 16 bits/histogram to 1000 bits/histogram. From experiments it was found that 64 bits/histogram yield good retrieval results. SCD has 64 coefficients in the feature set representing the image.
Color Structure Descriptor (CSD) uses HMMD color space to describe local color structure of an image in the feature set. By scanning the image using an 8X8 structure, presence of a particular color in the structure is counted. A color histogram is constructed such that it provides the count of a particular color in an image. CSD contributes 32 coefficients in the feature set.
2.1.2 Shape Descriptor
This feature provides a strong visual hint for similarity matching and is not influenced by scaling, rotation and translation. Shape transformation can be 2D or 3D. In general 2D shape description can be region or contour-based.
Region-based descriptor-ART is a moment invariant descriptor constructed by a complex angular radial transformation. All pixels comprising the shape bounded within a frame are used. This feature has the ability to describe any shape such as connected or disjoint regions. It can sustain minimal distortions along the object's boundary. The descriptor is represented with 36 coefficients in the feature set.
2.1.3 Texture Based Descriptors
Texture Browsing Descriptor (TBD), Homogenous Texture Descriptor (HTD) and Local Edge Histogram Descriptor (EHD) are the low level features that can represent the texture of an image. This descriptor enhances image search and retrieval process. Dissemination of texture in an image is described by Homogenous Texture Descriptor (HTD). It improves classification accuracy of objects in image retrieval. HTD is composed of 62 coefficients with first two as mean, standard deviation and remaining are energy and energy deviation of channels in frequency domain of Gabor filter responses. Feature vector of size 60 (30 means and 30 variances) for 5 scales each with 6 points of references, are obtained, scaled to analog [0-1] and placed into global feature vector.
Edge Histogram Descriptor (EHD) captures spatial dissemination of edges. An image is partitioned into sixteen equal isolated blocks covering the complete image. Edge information is then estimated for each block in five edge categories: 0˚, 45˚, 90˚, 135˚ and non-directional edge, each with one bin of local histogram. Therefore EHD gives 80 (16 non overlapped partitioned image blocks X 5 bins for 5 categories) of its coefficients to global feature vector. Thus global vector of size 464 is created to symbolize each image in the data set.
3. Fuzzy ARTMAP
Fuzzy ARTMAP architecture [13] derived from ART, has been widely adopted in literature for classification. It performs faster and expedites stable learning. The architecture includes two unsupervised fuzzy ART modules. These modules partitions input and output spaces. Fuzzy ART modules and Fuzzy ARTMAP are briefed in this section.
3.1 Fuzzy ART
Fuzzy ART is a binary ARTMAP that performs unsupervised learning implemented in analog domain using fuzzy and (∧) operator. The architecture comprises below fields.
Accumulation of input patterns is modeled by weights. Each input pattern has a stronger association with an output category. Thus each individual node j of F2 is associated with a weight vector wj = (wj1, wj2,..., wjM) . Weight wj represents the degree to which a particular feature is present in the image a such that it is coded by jth category. Initially the weights of output nodes are assigned to one and will not be selected by input pattern. During training an input pattern of known class is presented to the system. one of the output nodes j is selected by the pattern. The selected node is considered to represent the class of input pattern. Once a node in F2 Flayer gets selected, it is considered as committed to the input and weights are adjusted. Weights will always decrease and remains stable after convergence. Each output node is called as category and one or more categories may represent one class.
Choice Function: Consider N previously committed nodes in output layer. If an input pattern is presented, n nodes out of N committed nodes (n=1 or 2 or….or N) with varying degrees can be selected by the pattern. From n selected nodes at most one node (j) in F2 can become active, by taking on the input pattern(I). Output node (j) that can become active for input(I) is selected by evaluating choice (Target) function as given by (1).
where α ≅ 0 and Tj(I) is called activation value that determines the active category among nodes j and k committed to an input pattern ( I ) . Node that evaluates to the highest activation value becomes active node as given by (2).
Tj(I) determines strength of match between input pattern presented at an instant and weights of the jth output node. The ratio gives fuzzy subset hood of wj with respect to I. If any wj available is a fuzzy subset of I then ｜I ∧ wj｜ = ｜wj｜ and . Therefore, Tj(I) > Tk(I) for k ≠ j .
Resonance: Activity vector x in match field F1 is formed as below.
After a category J in output field matches to input pattern and becomes active, match field determines the match between input pattern and active category using (4).
If activity vector x is similar to input vector then (4) will pass the vigilance test where ρ ∈ [0,1]. If vigilance criteria is satisfied, network is said to be in resonance and the jth node in F2 is good enough to encode input pattern(I). In addition, for a node to establish resonance node j should also represent the same class as that of input pattern.
Learning: When an output node, that matches the input pattern is identified resonance happens, the weight vector wj is updated by (5)
where β is learning rate parameter. β ∈ [0,1] and β = 1 for fast learning.
Testing: Input pattern belongs to class corresponding to category which has high activation value. After training the system, using randomly selected input patterns the system is ready for classifying unknown images.
3.2 Fuzzy ARTMAP
Fuzzy ARTMAP [13], architecture performs supervised learning. Architecture shown in Fig. 1, has two ART modules, ARTa and ARTb. These modules cluster patterns from input space and output space into categories. During supervised learning, provided a set of input patterns {a,b} , ARTa receives a stream of input patterns {a} and ARTb receives a stream of patterns {b}. These two modules are linked by an associative learning network and an internal controller. Inter art module Fab (map field) links ARTa and ARTb.
Fig. 1.Fuzzy ARTMAP Architecture
Map field gets triggered, when one of the ARTa or ARTb categories are active. Mapping field contains the association between categories predicted by ARTa and ARTb for an input pattern. If prediction made by module a is disconfirmed at ARTb , map field initiates match tracking process. Match tracking increments vigilance ρa and a new search is triggered.
The process converges, if search results in any of the ARTa category predicted matches the predicted category in ARTb or a previously uncommitted ARTa category node. This process ensures that the category that resonates matches to the best with input pattern. FAM system is same as Fuzzy ART. In addition; match ensuring mechanism is included.
Inputs to two FAM modules are in complement code Ia = (a,ac) and Ib = (b,bc) where vector a = (a1,a2,...,aM ) is input pattern of size M and ac = (1-a1,1-a2,...,1-aM ) expresses absence of each feature in a . Input vectors Ia and Ib will be of size 2M. The map field denotes output vector of Fab and weight vector of jth node to Fab is denoted by . Between input presentations all activity vectors are set to zero.
3.2.1 Map field Establishment
Input to the mapping field Fab is from both or any one of the category fields in ARTa and ARTb . That is its activity vector xab is constituted by and as given by (6).
If for a given input pattern, Jth category in wins and becomes active, it sends information to map field Fab through . Then Fab is active only if predicts same class as . If yb fails to confirm prediction made by , then xab = 0 and match tracking is initiated.
3.2.2. Match Tracking
When the input is first presented to the network the vigilance parameter ρa is set to its baseline value . Matching field vigilance parameter ρab ensures the matching between categories in ARTa and ARTb . if then a best match exists. If a matching error occurs i.e. , match tracking increments and search is initiated for finding node. The process continues till a correct match for Ia is identified by ARTa . Otherwise a new category is created in ARTa and committed to Ia .
3.2.3 Category Proliferation
FAM suffers from category proliferation problem. Exclusive advantage of FAM is that its weights can be easily converted into IF-THEN rules. Number of categories created during training FAM is usually large. This causes an intractable bunch of IF-THEN rules due to class mismatches and creation of new categories which leads to category proliferation problem. This section briefs about how match tracking process leads to creation of new categories.
FAM category (j) can be viewed as hyperboxes Rj with corners decided by weights wj . Weight vector () of jth node in inter art module and its complement (1-wj,m+i) gives the minimum and maximum values of ith component among patterns in input vector that has selected jth category. So category size Rj can be defined by (7).
where M = ｜I｜ = a1 + a2 + (1− a1) + (1− a2) provides the size of attributes in input pattern and lji gives ith component of patterns learnt by jth category. Thus a pattern is learned by a category if pattern lies inside the box or the box enlarges to accommodate the pattern. The box is decided by choice function (1). Vigilance parameter ρa controls the size of the box. Hyperbox with weights w will be selected by an input pattern I only if . If the hyperbox passes the test, the weights of the hyperbox will be updated. Therefore upper limit of the size is controlled by ρ and given by (8).
Hyperbox represented by weights w = (w1, w2, 1−w3, 1−w4) has corners, input patterns a1 inside and a2 outside the box is shown in Fig. 2. Fig. 3 illustrates and provides the truth that a pattern that lies in two hyperboxes will be won by a box whose size is small and have large weights. In addition to ensuring better match, match tracking mechanism also causes category proliferation due to class overlap, pattern presentation order and presence of noise in data.
Fig. 2.Representation of hyperbox with size R1 formed by weights w = (w1, w2, w3, w4) in two dimension input space and input pattern a1 inside the box and a2 outside the box
Fig. 3.Illustration to prove if input pattern a contained in one or more hyperboxes Rj , the smaller box will have higher activation value. w1 = (w11, w12, 1-w13, 1-w14) ; w2 = (w21, w22, 1-w23, 1-w24); ; Similarly, ; Therefore T1 < T2 , Since (w11 + w12) < (w21 + w22) and (1-w13) + (1-w14) = (1-w23) + (1-w24)
3.2.3.1 Problem of overlapping classes
FAM network tries to classify all input patterns during training. But a large number of small categories are formed at regions, where two classes overlap. These categories will not contribute to predictive accuracy, since patterns in overlapping region may belong to either class. This leads to creation of many hyperboxes. During learning, class mismatches may occur and match tracking increases vigilance parameter. Higher value of vigilance parameter causes difficulties for existing categories to pass the vigilance test. Hence a new category is created to represent a particular pattern. Presence of large number of categories causes category proliferation problem.
3.2.3.2 Problem due to order of presenting
ARTa module performs an unsupervised clustering of input patterns. Correctness of category that resonates to an input pattern is examined by match tracking mechanism. This ensures that if the same pattern is presented subsequently, same category has to be selected. If selected category J does not provide a better match, ρa is increased, Jth category is reset and a new category is selected say k . It is verified that ｜RJ ⊕ a｜ > ｜Rk ⊕ a｜ . After learning, ｜Rk ⊕ a｜ will be the smallest hyperbox containing the input pattern. If the pattern is selected during subsequent instances, the system will select this category k .
Consider each category in Fig. 4 represented by regions R1 and R2 has a different associated class label. Let a1 belongs to class predicted by R1 . If pattern a1 is presented R2 is selected as it offers higher choice value. Since R2 predicts a wrong class, match tracking mechanism is triggered by incrementing ρa to a value such that . R2 is reset and then category R1 is evaluated. As ρa has been incremented R1 also will not meet the vigilance criteria and thus R1 is also reset. A new category is created and a1 is assigned to it. If R2 is not created and baseline vigilance = 0 i.e. if input patterns that have created R2 has not presented to the system before a1 , then a1 will select the class predicted by R1 . Thus match tracking mechanism that ensures prediction accuracy also can cause category proliferation based on order of presentation of input patterns.
Fig. 4.Geometrical representation of two hyperboxes (R1 and R2) associated for Fuzzy ART categories in a two dimensional input space. For an input a either R1 or R2 has to be enlarged to accommodate a , such that R1 ⊕ a > R2 ⊕ a for a → R1, R2 ⊕ a > R1 ⊕ a for a → R2
4. Non-Proliferation Fuzzy ARTMAP
NPFAM architecture is a modified version of FAM. Proposed NPFAM includes an inter ART reset mechanism which avoids category proliferation problem and also does not raise ARTa vigilance. But by performing an off-line learning stage, predictive accuracy is ensured. One-to-many → relationships are allowed and their probabilistic information is stored in . An offline map field with weight is included in FAM. stores probability of → , when inter-ART is reset during prediction mode. Entropy of training set is projected by the system using these weights. Also a vigilance parameter is assigned to each category node in ARTa .
4.1. Learning Phase
Before training, all weights are initialized to one. But is initialized to zero; j = 1,2,...,Na ,k = 1,2,...,Nb . A baseline parameter () is set as starting vigilance. is set to zero to minimize the number of categories. Vigilance parameter is raised in case the input space demands. In addition two parameters hmax and Hmax is defined to fix upper limit of hj and H respectively for restricting growth of hyperboxes. After initialization and identifying active node, training proceeds with vigilance test, entropy test and finally learning is completed with offline evaluation.
4.1.1. Vigilance Test
Training is continued by presenting input-output pairs, (a,b) to the network. When a pattern {a} is presented to ARTa , category J is selected according to (1). If the category is newly committed i.e. the category is not assigned to any other input pattern previously then ρJ = ρa . Vigilance test is performed to evaluate the reset condition using ρJ in (4). If category J fails vigilance test, this node will be suspended and new search is initiated. Pattern {b} is presented to ARTb and category k, that best matches to pattern {b} is selected. Then match field activity vector xab is formed by (7).
4.1.2 Entropy Test
Smaller hyperboxes representing the categories will provide high activation values. The part of the hyperbox, which is not available or contribute in predictive accuracy, is called entropy. The size of the entropy has to be restricted to ensure efficient classification and also avoids overlapping. The total entropy H is given by,
Probability of occurrence of Aj is denoted by Pj and Pjk is the probability of Bk assuming Aj . Contribution of jth node to total entropy hj is given by,
From activation vector xab of map field, hj is calculated. If Jth node is selected then
If hj > hmax thenJth node in ARTa is inhibited by setting TJ(Ia) = 0 , without raising vigilance parameter. Search process is continued to choose the other categories in ARTa , till entropy condition hj < hmax is satisfied.
4.1.3 Winner test
An input pattern is classified belonging to a particular hyperbox (winner node), if the pattern lies inside the box (provides higher activation value) or the box enlarges to include the pattern. Any algorithm must have the capability to restrict the growth of hyperboxes to include a single pattern that is away from existing patterns committed to the hyperbox. This ensures predictive accuracy by restricting the size of the box. This is achieved by verifying whether the node that has won the pattern satisfies (13).
Value of δ is predefined and called as decision parameter. If input pattern fails the test, winner node cannot enclose the pattern. Other nodes are restricted from learning the pattern. This pattern is categorized as unlearned. Other input patterns are presented for learning. After all patterns have been presented, many hyperboxes must have been created to cluster input patterns. Any one of the newly created hyperbox may provide high activation value to patterns in unlearned category. Now patterns in unlearned category are presented in subsequent iterations. If still some patterns are left out unlearned after a specified number of iterations, those are assigned to a new category. This largely reduces classification error and creation of new categories. Parameter δ prevents overlapping and avoids dependency of NPFAM on training pattern presented.
4.1.4 Offline Evaluation
Once training sets are learned by the nodes in output layer, learning process is considered to be completed. After processing all patterns, offline mapping field weights = 0 : j = 1,2,...,Na;k = 1,2,...,Nb and training data is presented to revise these weights. The units from ARTa and ARTb is selected in an unsupervised manner and weights is updated. Total entropy H is computed by (9) with value of Pjk and Pj computed from (14).
Impurity of all category nodes constructed during training process is described by Hmax . It also controls overtraining and influences overall accuracy of NPFAM classifier. If H > Hmax , then mapping defined by NPFAM between input and output portions are too large. Input spaces are partitioned further to bring them within Hmax and to improve accuracy of the system.
To achieve the requirement, node j in ARTa with maximum value of h is identified and removed. Weights and are set as 1 and 0 respectively. Baseline vigilance parameter ρa is re estimated as given by (15).
Newly created categories will thus be smaller in size, which is bounded by (9).
is carried out till H < Hmax . Patterns of deleted categories are submitted in next iterations. If Hmax = 0 then training algorithm takes long time to complete. If Hmax is set to high value, the process will be terminated earlier and leads to over simplification. ∆ρ used in (15) ensures, the category with H > Hmax is not created after getting detected. Thus input space is partitioned into categories such that retrieval accuracy is preserved.
4.2 Retrieval System
The proposed algorithm is implemented in two phases. Training phase in offline and testing phase in online. The conceptual framework with proposed NPFAM algorithm is depicted in Fig. 5. The algorithm depicting both the process is given in Fig. 6.
Fig. 5.Conceptual framework for Content Based Image Retrieval using NPFAM Classifier
Fig. 6.Algorithm for Training and Testing Phases
4.2.1 Training Phase
This phase constitutes extraction of features from images and scaling them between 0 and 1. Scaled features are stored in the database as feature vectors. The proposed system, with NPFAM classifier is trained with randomly sampled images across all classes and is labeled. The feature vector of each labeled test set image forms training vector. Thus NPFAM is trained to labeled classes by fusing training vector into the classifier.
4.2.2 Testing Phase
Feature vectors of all images in database are presented to the trained NPFAM classifier. The classifier does the mapping between input vectors and classes. During testing phase, query image is provided as input. Feature vector is extracted and normalized between 0 and 1. The classifier identifies the class of the query image. Images in the identified class are evaluated using (16) and displayed in increasing order based on similarity.
where Nfd is number of feature descriptors and
where p = number of elements in feature vectors corresponding to each descriptor.
5. Experimental Results
The proposed system is developed using Java and My SQL and implemented to retrieve general purpose image database from COREL. Two Image databases are used to validate and test the System. Data set1 having 10 classes with 100 images in each class (total 1k images) and Data set2 having 100 classes with 100 images in each class (total 10k images) in JPEG format of size 384X256 or 256X384. Raw images are used to validated and test. Visually similar images are considered to be an entity in a particular class. A maximum of 10 images from each class selected randomly is considered for training. Retrieval performance of the system from both data sets is evaluated using standard measures Precision (P), Recall (R) and F-Score. F-score is the harmonic average of recall and precision which reflects the degree of similarity and order of answers. Retrieval accuracy for query images with Gaussian noise of 0.1 mean, 0.04 variance is also verified.
5.1 Query Examples and System Demonstration
For evaluating the system performance a query image is submitted to the system. NPFAM classifier identifies the class of the query image. Feature vector of images in the particular class and query image are matched using similarity metric. The most similar images are displayed to the user. In every experiment, the system is evaluated by five images selected at random from each class as query image. Fig. 7 and Fig. 8 show the display of the query image and retrieved images for the class Rose and Sunset from Corel 10k data set, Horse and Dinosaur (with Gaussian noise in the query image) from Corel 1k data set.
Fig. 7.Retrieval results for Corel 10k query images. a) Rose b) Sunset
Fig. 8.Retrieval results for Corel 1k query images. c) Horse, d) Dinosaurs with Gaussian Noise(Mean=0.1,Variance=0.04)
5.2 Retrieval Precision/Recall and F-Score
To evaluate the effectiveness of the proposed approach the retrieval performance can be defined using Precision ( P ) and Recall ( R ). Average value of results obtained in five runs of different query images selected at random from a particular class is reported. For a particular query image, relevant images are considered to be those that belong to the same class as that of the query image. P and R is calculated by (18).
where NIA(q) is the number of retrieved images relevant to the query image. NIR(q) is number of images actually retrieved and NIt is number of images in the database which are relevant to the query image. Table 1a and 1b provides the precision and recall values for all 10 classes in data set 1 and selective 10 classes in data set 2.
Table 1a.Precision (P) and Recall (R) Values for Corel 1k data set
Table 1b.Precision (P) and Recall (R) Values for selective classes among100 classes in Corel 10k data set
Fig. 9 gives the precision at different recall values for FAM and NPFAM. Fig. 10 shows the comparison between the classification accuracy for different sets of images selected at random using FAM classifier and NPFAM classifier. From the results it is evident that NPFAM performs better than its unmodified version. Fig. 11 provides the F-Score plot for both methods. F-Score is estimated by (19).
Fig. 9.Precision Recall Plot for FAM and NPFAM
Fig. 10.Classification Accuracy for FAM and NPFAM
Fig. 11.F-Score Plot for FAM and NPFAM
5.3. NPFAM Parameters
The retrieval results of the system depend on NPFAM parameter values. Table 2 gives the optimal parameter values used in experiments for various generated categories (differing in number) in all training sets. Each set consists of 20 images selected at random from the database. A FAM-based system highly depends on the order of training pattern. Thus, large number of categories would be required. This produces more overlapping categories leading to misclassification and larger network. Developing a optimum network, to ensure better performance with a smaller network is a seperate task.
Table 2.FAM and NPFAM Parameters used in the experiments
Moreover, redundant categories are generated as it raises ρa after implementing Inter-ART process. In NPFAM, ρa is raised in offline mode and can be ignored when not necessary. This guides classifier to find optimal number of categories. Also, NPFAM is observed to be less susceptible to order of input pattern presented during training. Fig. 12 provides accuracy of the proposed system for various values of vigilance parameter. Optimum results are shown to be achieved when vigilance parameter is assigned to 0.2, 0.4 and 0.8. Comparison of NPFAM with other methods is provided in Table 3. Retrieval accuracy for proposed technique is achieved using raw images for querying, classification and matching.
Fig. 12.Accuracy of NPFAM classifier for different values of Vigilance Parameter
Table 3.Comparison of Accuracy between proposed NPFAM and other methods
5.4 Speed
NPFAM has been implemented on pentium III 700 MHz PC running on Linux operating System. Speed of proposed CBIR system using NPFAM is compared with that using FAM [23]. Proposed system reduces search space size by labeling images prior to retrieval. This speeds up querying process and enables easy and effective searching of large image database. Fig. 13 gives accuracy of NPFAM and FAM classifiers used in a CBIR system for different sizes of training sets. It has been observed that system accuracy is directly proportional to training set size. This is due to expansion and sharpness of training boundaries, resulting from the large number of training sets. Table 4a gives a comparison of the CPU time for NPFAM and FAM based systems. Table 4b provides the CPU time for Corel 10k data set in NPFAM system.
Fig. 13.Classification Accuracy Vs Training set size
Table 4a.CPU Time of NPFAM and FAM for Corel 1k data set
Table 4b.CPU Time of NPFAM for selective classes among100 classes in Corel 10k data set
6. Conclusion and Future Enhancement
The implementation of proposed system includes a comfortable GUI. The CBIR system with proposed NPFAM classifier is used to retrieve similar images from database. The classifier generates small categories assuring high accuracy. The work has also experimented influence of values assigned to parameters on system performance. Results achieved prove that NPFAM classifier is a solution to category proliferation problem, yields good results for query images with noise and also without implementing region-based classification. Proposed algorithm also produces good results for different data sets and is independent on training pattern presentation. NPFAM based retrieval system uses a large size of feature vector. It is successfully fused into classifier, trained and minimum target error is achieved. The system performance can be improved for small training sets by including relevance feedback.