A neural network-based segmentation tool for color images

The research paper focuses on the development of an efficient and accurate tool for segmenting color images. The segmentation is a problem that has been widely studied since machine vision first evolved as a research area. However, the vast majority of this research has predominantly focused on segmentation as a means of preprocessing to enhance the performance of subsequent recognition mechanisms. Thus, the needs of applications such as OCR (optical character recognition) or handwriting recognition and as well as other recognition applications have largely shaped the requirements and targeted performance of segmentation mechanisms. Specifically, these recognition applications do not require single pixel or sub-pixel accuracy to function properly and since providing this accuracy is difficult and computationally expensive, most segmentation approaches do not provide it. In effect, for recognition applications it does not matter if one or two pixels have been improperly categorized as belonging to inappropriate segments within the image. Subsequent recognition processing is able to effectively compensate for this type of noise that may be introduced and hence still recognize a shape or set of shapes as intended. However, for the embroidery application proposed here, the artwork being scanned is often quite unique and does not usually conform to specific predefined recognizable objects. Additionally, since the ultimate goal is to produce an accurate reproduction of the scanned image, the aesthetic accuracy of the segmentation approach is very important. Similar requirements exist in the area of medical imaging where segmentation must provide the highest possible precision. Although in this area, still different goals are dominant. For example, there have been many neural network approaches to segmenting medical images such as x-rays or MRIs. However, here a more important goal has been to classify pixels into known predetermined categories to allow easier detection of anomalies indicating disease or other malignancies. Thus the area of study presented here still remains open and largely un-addressed by previous research.