Abstarct: Word spotting is a way to index, search and retrieve words in document images. Word spotting aims to find multiple occurrences of a query word in document images. Word spotting is one of the major challenges in the field of document image analysis which has recently received many attentions. Several word spotting systems have been introduced on Latin scxript, but a limited number of works exist on Arbic and Persian documents and most of them only proposed for printed documents. The goal of this thesis is to present an end to end multiwriter Persian word spotting system which accepts, independently of lexicon, both word images and text strings. The number of classes in this problem is very high. Moreover very large intra-class variability due to different writing styles, illumination, typography, etc, can make the same word look very different. For this reason, a classification method using attributes is needed. Attribute-baxsed classification by sharing information between classes makes out-of-vocabulary (OOV) spotting possible. In this approach, both word images and text strings embeds in a common vectorial subspace. This is achieved by a combination of label embedding, attributes learning, and a common subspace regression. In this subspace, images and strings that represent the same word are close together, allowing one to cast recognition and retrieval tasks as a nearest neighbor problem. In the first step, an attribute-baxsed method is proposed to retrieve isolated handwritten Persian words. In this method we suggest two methods for representing Persion words using attributes. These methods are inspired from two facts, first the appearance of a Persian letter changes depending on its position in a word. Second, Persion letters can be categorized baxsed their shape of main part. To evaluate the performance of proposed methods, two isolated world Persian handwritten datasets, Iranshar and Farsa, are used. Mean average precision (MaP) for presented attributes in word spotting is 95.67%, 96.25% and precision for handwritten word recognition is 96.15% and 97.55% respectively. Then we introduce a method for segmenting handwritten documents to constitute lines. Using a new approach we slide a window on each line of a document. The proposed sliding window is baxsed on connected components, so it can overcome to the problem of basic methods baxsed on sliding window such as determining window size and movement step. FHT dataset is used to evaluate the proposed line-baxsed method. MaP criterion in this experiment for query by example and query by string is 73.67 and 81.02% respectively. Finally, a two-pass segmentation-free method is proposed for spotting of words in handwritten documents. In the first pass a number of candidate regions for the query word are selected in the document. These regions are extracted baxsed on the proposed spotting method at connected component level. Then considering neighboring connected components to the candidate regions, the regions which likely match the query word are extracted. Evaluation of this method is also performed using FHT dataset and MaP 79.45% for text query and 70.25% for image query are obtained. It is noteworthy that we trained the proposed line baxsed and segmentation free methods using Farsa and Iranshahr datasets whereas FHT dataset is used for test. This illustrates that the proposed methods are robust to writing style. Furthermore, any word from class of queries is not used in training phase so the proposed methods can spot and retrieve unseen words.