RESEARCHES

In this study, we introduce a cell-based labeling method that can accelerate multi-object recognition to extract locations and features of multiple objects in an image, and multi-object feature extraction is performed at 2000 fps for 512×512 images by implementing a cell-based labeling algorithm as hardware logic on a high-speed vision platform.

A concept of cell-based labeling can reduce the number of scanned pixels for labeling and memory size to store label equivalences without accuracy degradation in space resolution by dividing an image into subimage regions as cells, assuming the additivity in image feature calculation. Our cell-based labeling algorithm has two sub-processes: (1) cell-level image feature calculation for divided cells, and (2) object-level labeling and updating of cell-based image features. The upper figure shows the concept of cell-based labeling algorithm for moment feature calculation.

To accelerate calculation of additive image features such as moment features, higher-order local autocorrelation (HLAC) features, and color histograms, we have already implemented several hardware logics of cell-based labeling on an FPGA-based high-speed vision platform, IDP Express. In the implementation, additive image features of 1024 objects in an image can be simultaneously extracted for multi-object recognition by dividing the image into 8×8 cells concurrently with calculation of the 0th and 1st order moments to obtain the sizes and locations of multiple objects.

Their performances that can simultaneously execute multi-object feature extraction at 2000 fps for 512×512 images, were already verified by showing several experimental results for high-speed moving objects.


Cell-based labeling for moment features