Dylan Seychell

Monoscopic inpainting approach using depth information

Cheap depth sensors that can be integrated in consumer cameras provide additional data that can be used for improved post-processing results of the captured images. Removal of objects in a scene is one such editing procedure that demands inpainting techniques that limit noticeable artifacts generated in the process. In this paper, a monoscopic inpainting technique that uses depth information to process results is presented. It allows users to select an object from the foreground that needs to be removed and then inpaints this region from the neighborhood. This approach uses texture and depth information and is pipelined in a way that allows for parallelization. Results are returned in 0.034 seconds on average. A mean opinion score evaluation was carried out and the current technique scored an average of 3.24 from a scale of 5 on the quality of inpainted regions. This exercise was held to identify the attributes that need to be improved in future implementations.

Recognition of Handwritten Characters Using Google Fonts and Freeman Chain Codes.

In this study, a unique dataset of a scanned seventeenth-century manuscript is presented which up to now has never been read or analysed. The aim of this research is to be able to transcribe this dataset into machine readable text. The approach used in this study is able to convert the document image without any prior knowledge of the text. In fact, the training set used in this study is a synthetic dataset built on the Google Fonts database. A feed forward Deep Neural Network is trained on a set of different features extracted from the Google Font character images. Well established features such as ratio of character width and height as well as pixel count and Freeman Chain Code is used, with the latter being normalised using Fast Fourier Normalisation that has yielded excellent results in other areas but never been used in Handwritten Character Recognition. In fact, the final results show that this particular Freeman Chain Code feature normalisation yielded the best results achieving an accuracy of 55.1% which is three times higher then the standard Freeman Chain Code normalisation method.

Efficient object selection using depth and texture information

Object selection is a challenge in computer vision since it is generally a trade-off between accuracy and performance. A popular approach is the use of bounding boxes around objects that are to be selected. Other common techniques provide a set of objects from which the user can then choose. The method presented in this paper is designed around the priority of performance and granular selection of objects in the scene. Experiments performed on a non-parallel implementation of the proposed solution return results in an average time of 0.043s. The technique also returned very good results in the processing of objects that are partially occluded, hence enabling future work in improved identification and recognition of such objects.

A Model for Cultural Attributes

One of the challenges when predicting human behavior is the variation of the same behavior in relation to the culture of the individual in question. This paper demonstrates a technique that can be employed in order to predict certain individual cultural attributes based on a regression model that considers age, gender and nationality. The technique makes use of the data set collected through the World Values Survey. An equation based on the Multinomial Logistic Regression was derived in order to extract the values of the designated cultural attributes with respect to the given parameters. This paper briefly explores the World Values Survey together with the theory behind this concept that lead to the composition of a set of equations. A possible application of this model is also presented.