Information Fusion Based Object Detection And PET Image Reconstruction

Object detection and PET image reconstruction are two hot topics. Object detection aims at determining the position, contour or gesture of object or its components, so it is the prerequisite and foundation of many problems in computer vision. PET can determine the distribution of radionuclide by receiving the gamma photons, so it can provide the important diagnostic information for forecasting cancer. Thus, both object detection and PET image reconstruction play significant roles in computer science.Owing to presence of abundant information, object detection based on information fusion achieves a better performance. The part-based method which is the focus of this article is a typical kind of method in them. However, the detection accuracy and robustness of part-based methods are not good enough to be applied in the complex environment now. For the problems of detection accuracy and robustness, some improvements for the part-based methods are proposed in this paper, these belong to the feature-level informa-tion fusion. PET image reconstruction based on information fusion usually achieves the accurate reconstruction by acquiring the projection in multi-views. However, there are some disadvantages existing in the present rotating PET systems. For example, it is diffi-cult to estimate their SRMs accurately, and a very high mechanical precision is required. For these problems, a new rotating PET system is designed, it belongs to the data-level information fusion. This paper focused on the following four parts:1) This paper proposes a part-based model to address the part-occluded problem. For robustness to the occlusion, the traditional shape model and local appearance model are modified. In addition, the occlusion prior and a penalty term are also introduced to regularize the argument space of the objective function. In detection, the outer inference and the inner inference are carried out alternately to optimize the object function. In this procedure, we also incorporate the validity test mechanism to avoid the invalid inner inference results. The results of non-occlusion detection show that our method can achieve a fairly decent detection rate within small number of iteration. The results of occlusion detection show that our method achieves a good performance even though most parts are occluded.2) This paper proposes a part-based model with rotation invariance to address the problem of object rotation. To get rid of the influence of the object rotation, we design a novel shape model with rotation invariance which only refers to the dis-tances among the parts. In detection, BP algorithm is employed. Furthermore, we generalize the generated distance transforms, which makes Beliefs calculated in a nearly linear time. The results show that the performance of our method is constant under different angles and different weight. Meanwhile on the rotated dataset, the performance of our method is better than that of sparse shape prior model, with a good enough initial localization. On the non-rotated dataset, the performance of our method is not worse than that of Pictorial Structures Model when the part number is larger than 2. But if the part number is 2, due to the absence of angle informa-tion, the performance of our method will be lower than that of Pictorial Structures Model. The results also show that the shape model designed in this paper can be combined with various local appearance models.3) For the problem of low constraint of existing shape model, this paper proposes a new graphic model to describe the relationship among the parts. In this way, the dynamic programming is not applicable any more. For this problem, a hierarchical propagation algorithm is proposed in this paper, whose computational complexity is linear. The results show that the performance of our method is higher than that of k-fan (k=1,2) on the motorbike and airplane dataset.4) For the problems of estimating SRM and high mechanical precision, a 90° rotating PET system is designed in this paper, it improves the spatial resolution along x-axis greatly, reduces the engineering complexity and requirement of mechanical preci-sion. To our knowledge, our PET system achieves the best geometrical symmetry, which reduces the estimated time of SRM to the greatest degree. Moreover, this designment avoids the large amount of redundant data. The results show that our PET system can resolve the contour of phantom clearly under the detector spac-ing of 5cm,10cm and 15cm, and achieve a x-axis spatial resolution of 0.79mm, 0.74mm and 0.72mm under these detector spacing respectively, which are far bet-ter than that of static system. Finally, we conclude that the key element affecting the resolution is the SRM correlation between the adjacent voxels.