HANDWRITTEN NANDINAGARI IMAGE RETRIEVAL SYSTEM BASED ON MACHINE LEARNING APPROACH USING BAG OF VISUAL WORDS Prathima Guruprasad1, Dr. Jharna Majumdar2 1Dept. of Computer Science and Engineering, NMIT, Université of Mysore PB No. 6429, Govindapura, Gollahalli, Yelahanka, Bangalore, India 2Dept. of CSE(PG) and Center for Robotics Research, NMIT, Université of Mysore PB No. 6429, Govindapura, Gollahalli, Yelahanka, Bangalore, India ABSTRACT Recognition is even more challenging in case of Human handwriting comes with different rare ancient handwritten Nandinagari styles, highly variable and inconsistent which manuscripts which is earlier form of Devanagari makes recognition of rare handwritten scripts. In recent years, large-scale image Nandinagari scripts extremely challenging. indexing and retrieval show sufficient scope in This paper proposes a Bag of Visual Words both corporate sector as well as in research (BoVW) based approach to retrieve similar community. The focus of this work is to use Handwritten Nandinagari character images invariant descriptors to represent such rare from visual vocabulary. This technique is manuscripts. Retrieval of similar images from a useful in the retrieval of images from a large large image database is a challenging problem. database, efficiently and accurately also helps This is because the memory requirement of the in achieving required scalability. The images, time of processing, indexing and character images are represented as retrieval time will grow as database size grows. histogram of visual words present in the Also effective and accurate retrieval when the image. Visual words are quantized images are with various sizes, orientations and representation of local regions, and for this occlusions is a matter of concern. In the proposed work, SIFT descriptors at key interest points model, the Bag of Visual Words (BoVW) are used as feature vectors. For the efficient concept is used for effective handwritten handwritten Nandinagari character retrieval Nandinagari character image retrieval as the system, the optimal number of clusters to be technique is widely used for text based system. chosen is 32 with the maximum mean Average This is the way in which human mind interpret Precision (mAP) as 0.904. the similar objects by generating the visual Keywords: Character Image retrieval, Bag of words. In BoVW, the Nandinagari character Visual words, SIFT Features, Code book image is represented as discrete visual words, the generation, Clustering, Indexing and quantized representation of large feature retrieval, Visual Vocabulary. descriptors extracted from the images. During recognition, similar images are retrieved by I. INTRODUCTION effectively computing the histogram of visual The Indian regional languages are cursive in word frequencies with the closest histogram nature so, reliable and efficient Recognition matching in very less time. In order to process system is still not available.Many of these the handwritten Nandinagari script which is languages come in different forms subjected to having complex 52 characters which include 15 changes region to region and time to time. They vowels and 37 consonants, the following steps have rich heritage which are in manuscript Palm are used. First the Nandinagari character dataset Leaf forms and also in handwritten forms. is prepared by converting them to gray scale images in different sizes and orientations. ISSN (PRINT): 2393-8374, (ONLINE): 2394-0697, VOLUME-4, ISSUE-4, 2017 163 INTERNATIONAL JOURNAL OF CURRENT ENGINEERING AND SCIENTIFIC RESEARCH (IJCESR) Then detect the interest points and generate the Step 2: Feature Extraction: This is done using invariant feature descriptors around the key SIFT method. First identify the key interest points using SIFT (Scale Invariant Feature points which usually lie on high-contrast regions Transform) technique. This is the pre-processing of the image, such as corners which remain step for BOVW approach which is based on data invariant even change in scale, rotation and visualization concept. Next is code book illumination. Then 128 feature descriptors generation by quantizing local SIFT descriptors extracted for each indentified key points on the into visual words. This is also called as images. The steps are as shown in Figure 2. vocabulary generation. The last step is indexing and searching similar images from the visual Step 3: Representation of visual words: The vocabulary. This approach is highly scalable. To large set of SIFT local feature descriptors are measure the recognition accuracy mean Average represented as visual words using BOVW Precision method is used. approach. This technique groups the similar descriptors using k means clustering [7] II. REVIEW OF LITERATURE approach and codebook of k X 128 dimension is Large scale image retrieval systems is using generated by k cluster centroids. This is based on the BOVW model in which each image can be number of clusters formed and called as k visual viewed as a sparse vector of visual words to text- words. Each key point descriptor of an image is retrieval system[2]. BOVW model is designed assigned to the closest cluster centroid. The for the local descriptors of images to describe number of assignments is represented as regions around the key points which is detected histogram of frequencies of visual words and in the images [3]. Instead of global features, an generated by aggregating of local SIFT image can have a set of patches around key descriptors. interest points. And local descriptor of 128- dimension SIFT features is used to be a good way Step 4: Then indexing and retrieval of similar to represent the characteristics of these patches characters from the codebook by querying is [4][5]. But when such local descriptors for each done using nearest neighbour approach in the image is extracted, the total number of such recognition phase. Top N similar images features is very huge. And comparing and retrieved from the database. searching similar matches for each local descriptor in the query image becomes too IV. EXPERIMENTAL RESULTS cumbersome. Therefore, BOVW is proposed as As input, Handwritten Nandinagari characters a way to solve this problem by compact of different sizes 256 X 256, 384 X 384,512 X representation of these descriptors into visual 512 and 640 X 640 and 5 Rotations are 0o, 45o, words, which decreases the size of the 90o, 135o and 180oare considered. This forms a descriptors [6][8]. This is helpful in scalable 1049 characters in the database. indexing and fast search on the vector space. For SIFT feature extraction, a set of images III. PROPOSED METHODOLOGY (.jpg or .png files) contained in a folder writes the The training phase in this proposed model as extracted features into text files. For each image, shown in Figure 1 of recognizing the hand this first key points of each image is identified written Nandinagari scripts using data and 128 feature descriptors are generated for visualization has following steps. The visual each of these key points. In this way a total of vocabulary is used for matching query images in 1049 input images are fed to the feature the recognition phase. extraction module and feature files are written in Algorithm Steps a directory which represents different types and Step 1: Preparation of dataset: Standard dataset styles of images. Average number of interest for handwritten Nandinagari characters is not points for these images are ranging from 63 to available .As input, Handwritten Nandinagari 268 key points. And the average execution time characters of different orientations and styles are is ranging from 45 milli seconds to 218 milli prepared. A total of 1049 handwritten seconds and the feature descriptor size is ranging Nandinagari character images are taken for result from 35 MB to 166 MB. This feature analysis analysis with minimal or without Pre-processing statistics is as shown in Figure 3. by converting them to Greyscale. ISSN (PRINT): 2393-8374, (ONLINE): 2394-0697, VOLUME-4, ISSUE-4, 2017 164 INTERNATIONAL JOURNAL OF CURRENT ENGINEERING AND SCIENTIFIC RESEARCH (IJCESR) Input Samples Query Images Converting to Converting to Gray scale Gray scale Images Images Extracting Invariant Extracting Invariant Features using SIFT Features using SIFT method method Generate the code book Generate the query vector of the feature descriptors using K means approach Indexing and retrieval Visual Word of similar images using representation nearest neighbour in the vocabulary method database Retrieved Results Figure 1: Proposed model 1.Construct Scale 6.Generate Feature Descriptors Space 5.Assign Key point 2. Take the difference Orientations of Gaussian 4.Remove Edge and 3. Locate DOG extreme Low Contrast Responses (Min or max) Figure 2: Steps in SIFT Method In SIFT codebook generation phase K means clustering approach is used and code book is generated only once. The code book file stores only cluster centres aFnd the number of cluster centres depends on number of clusters formed based on the value of K. To analyze the performance the codebooks with 8, 16, 24, 32 and 52 cluster centroid are generated. These are called visual words. Figure 3: Average no. of interest points, Average execution time and Feature Descriptor size for Character images ISSN (PRINT): 2393-8374, (ONLINE): 2394-0697, VOLUME-4, ISSUE-4, 2017 165 INTERNATIONAL JOURNAL OF CURRENT ENGINEERING AND SCIENTIFIC RESEARCH (IJCESR) Table I. Code book size and Clustering time query image vector with nearest neighbour different number of clusters method is computed. Sl. No. of Index Indexing Here top N similar images are retrieved from No. Clusters Size(KB) Time(m.s.) visual vocabulary. From the image set a representative of 10 query samples as shown in 1 8 8,176 95250 table 3are taken for analyzing the performance of this proposed framework. 2 16 16,895 88513 Table 3: Query Image samples 3 24 24,648 93430 Query Name Query Size and Image Orientations 4 32 33,251 91417 5 52 53,761 85466 512X512, 0 A Deg. These internally get mapped to all key points present in that particular cluster.