Development of statistical object discrimination system using image characteristics

在本研究中，我们将已经在统计学研究中发展起来的数据识别技术应用于图像中的物体识别，并通过评估给定的两幅图像来自同一物体的概率，开发了一种图像物体识别系统。我们实现了(1)使用逻辑判别分析开发目标识别系统，(2)将分层人工神经网络引入(1)开发的系统。我们还开展了以下相关研究：(3)纺织品图像特征与人类感性的关系研究；(4)统计判别的基础研究。我们在下面解释(1)和(2)，这里省略(3)和(4)的解释。(1)我们开发了一个系统，该系统可以根据图像中的物体与另一图像中的物体是属于同一类别还是不同类别的置信度进行区分。该系统通过测量不同图像特征下的每对已知图像之间的差异，利用最大似然方法推导出最优区分同一类别和不同类别的逻辑函数。利用该逻辑判别函数对新的图像进行置信度判别。我们将该系统应用于树叶图像的识别。(2)将(1)开发的系统与三层神经网络进行了统一。我们仅将神经网络应用于错误率较高的情况，并开发了一个具有清晰的逻辑函数判别模型和神经网络高判别能力的判别系统。该混合系统具有与原始神经网络基本相同的识别能力，并将计算时间减少到原来的五分之一。

We developed in this research a discrimination system of image objects by applying the data identification, which had been already developed the statistical research, to object identification in images, and by evaluating the probability that two given images are derived from an identical object. We achieved that (1) development of an object identification system using the logistic discrimination analysis, and (2) introduction of the layered artificial neural network into the system developed in (1). We also carried out related researches as follows : (3) investigation of relationship between image characteristics and human kansei on textiles, and (4) fundamental research on statistical discrimination. We explain (1) and (2) in the following, and explanations of (3) and (4) are omitted here.(1) We developed a system that discriminates with the degree of confidence whether the object in an image and that in another image are in the same category or in the different categories. This system measures the differences between each known pair of images subject to several image characteristics, and derives using the maximum likelihood method the logistic function that optimally discriminates pairs belonging to the same category and those belonging to the different categories. A new pair of images is discriminated with the confidence degree by this logistic discrimination function. We applied this system to the discrimination of images of leaves.(2) We unified the system developed in (1) and the 3-layer neural network. We applied the neural network only for the cases that the error ratio is high, and developed a discrimination system that has the clear discrimination model of the logistic function as well as the high discrimination ability of the neural network. The new hybrid system achieved almost the same discrimination ability as the original neural network and reduced the computation time to one-fifth.