Design for Embedded Image Processing on FPGAs, 2nd Edition

13Blob Detection and Labelling

Segmentation divides an image into regions which have a common property. One common approach to segmentation is to enhance the common property through filtering, followed by thresholding to detect the pixels which have the enhanced property. Filtering has been covered in Chapter 9, and various forms of thresholding in Sections, 8.1.5, 8.2.3, and 9.3.6. However, these operations have processed the pixels individually (using local context in the case of filters). To analyse objects within the image it is necessary to associate groups of related pixels to one another and assign them a common label. This enables object data to be extracted from the corresponding groups of pixels.

The transformation from individual pixels to objects is an intermediate‐level operation. Since the input data is still in terms of pixels, it is desirable where possible to use stream‐based processing. The output consists of a set of blobs or blob descriptions and may not necessarily be in the form of pixels.

Of the many approaches for blob detection and labelling, only the bounding box, run‐length coding, chain coding, and connected component analysis (CCA) will be considered here. Field‐programmable gate array (FPGA) implementation of other region‐based analysis techniques such as the distance transform, watershed transform, and Hough transform is considered at the end of this chapter.

13.1 Bounding Box

The bounding box of a set of pixels is the smallest rectangular box ...