|
The connected-component labeling (CCL) is a technique for extracting connected pixels having the same value. It is mainly used for abnormality diagnosis of products, and for extracting noise areas of products. In the extraction of product areas in product diagnosis, a hole filling processing (HFP) is used to complement discolored areas. However, the HFP is inefficient, because the CCL needs to be executed twice in the foreground and background, and half of the threads are idle during each process. In this study, we propose a rewriting method for continuous label IDs with pixel-by-pixel parallelism, and a HFP method using simultaneous CCL of foreground and background. We implemented and evaluated these methods on Jetson TX2. The rewriting process to the continuous label ID is 3.7-13.8 times faster than the conventional method of sequential processing on the CPU, and on average 9.2 times faster. For the HFP using simultaneous CCL, we implemented and verified the conventional method that requires twice the CCL and the proposed method that can extract the foreground and background with one CCL. The performance of the proposed method is about 13-27% better than that of the conventional method. In addition, in the lightweight object detection method that is an application using the proposed method, the facial detection time is about 14 ms, which is about 60 times faster than the conventional method. As a result, the facial detection processing with high computational complexity can be operated practically even on an inexpensive and small processor. The CCL process for GPGPU has little room for optimization, and it has been difficult to achieve higher speeds. However, we focused on wasted idols in the HFP, proposed a method to reduce and supplement them, and realized a faster HFP than the conventional method.
|
