Today, various photography devices such as digital cameras, web cameras and mobile phones have caused the rapid growth in the number of digital images and consequently, voluminous digital image databases have been constructed. Therefore, saving, searching, organizing and management of digital databases have become significant and indispensable. Image retrieval is considered as an important searching subject in this field, and it is generally divided into two main groups: text-based image retrieval (TBIR) and content-based image retrieval (CBIR). TBIR was initially introduced for searching images in the 1970s. In this method, for description of each image, one or a number of words are allocated manually. Then, these words are used by a database management system to perform the image retrieval. TBIR is based on keywords, so its implementation is an easy task. However, TBIR methods have some disadvantages: first, it is hard or sometimes impossible to assign features for large databases manually and second, different individuals may never assign the same words to an image. To solve the problems of TBIR methods, Kato first introduced CBIR in 1992. The main difference between TBIR and CBIR methods is that in TBIR, human interference is necessary while in CBIR, retrieval process is done automatically. Color, texture and shape are the typical image features that represent and index an image in CBIR systems. In CBIR, each image is represented by a feature vector and then all the images in the database are indexed using a similarity measure. Finally, the images with minimum distance (maximum similarity) are retrived. Feature extraction and feature matching are very important in CBIR. A feature vector with higher information leads to better precision in retrieval. On the other hand, an appropriate similarity measure can be very effective in CBIR. Many different studies have been conducted to extract color features. Dominant Color Descriptor (DCD) is of particular importance because of its simplicity and compactness. In this research, new DCD features are proposed as well as a new similarity measure based on Euclidean distance. On the other hand, the multi-resolution analyses, such as wavelet and curvelet transforms, are the most significant methods for the extracting of the efficient texture features in CBIR. So, the features based on the wavelet and curvelet are improved in this research. Finally, a new CBIR system based on optimized integration of DCD, wavelet and curvelet is proposed using Particle Swarm Optimization (PSO) algorithm. The proposed CBIR system has outperformed the prior methods by 4% in terms of average precision. Texture is one of the most important features of an image which has a high potential to discriminate the images from each other. Since different definitions for texture is done several methods have been introduced to extract it. Among them, local features are very important and have been used in many applications such as face recognition. LBP, LTP, LDP, LTrP and LVP methods extract local textures from the image. In general, LBP, LTP, LDP, LTrP and LVP are based on gray-level difference of the referenced pixel and its nearest neighbors. On the other hand, these patterns are defined based on various combinations of gray-level difference of pixels located in a square or circle. Since many of the natural textures can be shown by the relationship of pixels intensities along a line, these methods have limited ability to represent the texture information. Besides, the difference between the referenced pixel and its adjacent pixels is encoded with two, three or four values in the aforementioned methods, and this may lead to losing much of the image information. Therefore, more image information will be preserved if multi-level coding is used instead of binary coding. In this research, we propose the Local Derivative Radial Patterns (LDRP) to obviate the previous patterns drawbacks. LDRP is based on gray-level difference of pixels along a line and their weighted combinations. Therefore, these patterns are descriptors for extracting meaningful texture information from the image. In addition, multi-level coding in different directions is used instead of binary coding which leads to higher precision in image retrieval. Experimental results on Brodatz and Vistex databases show that proposed LDRP outperforms the prior methods by at least 3.5% in terms of average precision.