نشریه ماشین بینایی و پردازش تصویر
سال یکم شماره 1 (بهار و تابستان 1392)

Research activities in wireless sensor networks have been growing in recent years. In such a network, sensor nodes collect scalar data such as temperature, pressure, humidity and etc. Scalar data are not sufficient for some applications like automatic surveillance and environmental monitoring. With recent advances in the technology of image sensors and embedded processors, most of attentions have been concentrated on camera sensor networks. Therefore, these networks are being utilized in many applications such as environmental monitoring and target tracking. Target detection, localization and tracking in a specific region are the most important issues in these applications. Due to quantization in CCD cameras, the obtained information from these nodes is not very accurate.
In this paper, we present a geometrical model to analyze the quantization error. The proposed model can be generalized to a multi-camera system, where more than two cameras are used to have more accurate estimation of the target location. This error can be decreased by selecting cameras in the network with appropriate positions. Camera selection problem in camera sensor networks is essential not only to improve the accuracy of the network but also to compensate for the processing, energy and bandwidth limitation of each sensor node. Hence, for more accurate estimation of the target height and to prolong the lifetime of the network, we propose the priority as well as a genetic search algorithm. In these algorithms, the precision of height estimation and also the resource constraint of the network are considered. Therefore, the accuracy of the measurements and also the lifetime of the network are increased. Simulation results show that the proposed metrics decrease the computational overhead and energy consumption of the network.

Rice is one of the most important stable foods in Iran. Sometimes, for reason such as illegal profit, it is probable a commercial rice variety with good quality properties be mixed with some low quality properties that have great similarity in appearance. In this paper, an expert system for rice purity detection based on extracted texture features of bulk samples and modeling by a multilayer neural network has been introduced. First, images of bulk samples are taken using a black box. Then, texture features is extracted. In the next step, the best features are selected using a genetic algorithm approach. Finally, a neural network based regression is used for modeling of proposed approach. The best performance is obtained using local binary pattern. To increase the efficiency of the proposed approach, the results of previous section is combined using a majority voting approach. The result of this study can be used for construction of rice purity detection system.

Content-based image retrieval (CBIR) is a major challenge in the field of pattern recognition. The CBIR systems attempt to bridge the semantic gap by employing relevance feedback (RF) methods like short term learning (STL). This paper proposes a similarity refinement based STL method in CBIR systems. In this method, the weights of the feature’s components and also the weights of each type of features are optimized by minimizing an error function. The proposed method is examined in a standard public dataset with 10000 color images. The proposed error function improves the precision and computational time in comparison with the similar methods. The experimental results and comparison with the competing methods confirms the effectiveness and efficiency of the proposed method.

Since, most of the descriptors of 3D models are not invariant to various transformations and differentiations, the alignment of 3D models is one of the most important steps to achieve high precision 3D model retrieval system. In this paper, a method is presented to estimate the different pose of triangular mesh model in 3D space using Nelder-Mead optimization algorithm with non-overlapping pixels of each pair of 2D silhouettes for many viewing angles as cost function. So, after applying the translation and scale standardization for 3D models, in each class of considered database, a favorite model is selected as the example and the other models are rotated in such a way to reach the most similar 3D pose of example model. The overall performance of the suggested framework is evaluated using McGill 3D models Database. The numerical results obtained from different experiments prove the ability of proposed algorithm in 3D model alignment. For example, in airplane 3D model with silhouettes size, the error of alignment is 36437 pixels. This error equals to 6.8% of total area of whole 2D silhouette views of fixed moving 3D models.

Goal is the most important event in soccer matches; thus, goal event detection is very useful for video summarization and video retrieval. In this article, we propose a new goal event detection method in the broadcast soccer videos using a fuzzy inference system. In this method, audio-visual data is processed and low-level and mid-level features are extracted to detect the goal event as a high-level concept. At the low-level processing stage, audio energy is extracted from the audio signals and 3D RGB histogram is computed for each frame. Additionally, boundaries of shots are detected as low-level features. Then, mid-level processes are accomplished. This stage contains view type recognition, logo detection, and replay boundary detection. Finally, the video is segmented to some semantic parts and a fuzzy inference system investigates the content of each semantic part to detect goal events. The main contribution of our method is presentation of expert's knowledge and heuristic rules in the form of fuzzy rules for goal event detection. This method benefits from fuzzy modeling and fuzzy inference systems while presents heuristic rules in a simple and understandable form. A soccer video data set containing 12 videos related to FIFA 2010 (South Africa) was used for experiments. Experimental results show that precision and recall of our method is 90.9% and 90.9%, respectively. They also illustrate that the proposed method outperforms other methods for goal event detection.

Face retrieval is an important research topic in image processing and aims finding face images similar to a query image. In this paper, a novel method is proposed to retrieve face images using gradient histogram and local binary pattern (LBP). The combination of these two techniques will increase the robustness against face variations and thus improve system performance in face retrieval. In order to increase system ability, a relevance feedback scheme based on support vector machine (SVM) is proposed. The Experiments have been conducted on the AR face database in two modes: without occluded images and with occluded images. Experimental results show that the proposed method can retrieve face images effectively. In the next, the proposed method is compared with several successful methods in face researches. Mean average precision (MAP) metric for the proposed method in two experimental modes is equal to 94.40% and 68.12% , while the best results for compared methods is 90.37% and 61.99%, respectively. The results show that the proposed method is superior to these methods and is a good method to retrieve the face images.