| 摘要 |
A method and apparatus are provided for reducing the complexity of two-dimensional linear convolutions. All possible pixel patterns for a row of a pixel array are determined and multiplied by a corresponding row in a convolution kernel matrix. The partial convolution results are stored in a lookup table. In a processing step, each row of an input pixel window is compared to the possible pixel rows. Partial results associated with a matching pixel row are accessed from the lookup table. In one preferred embodiment of the invention, each row of an input pixel window is stored in a line cache. In another, equally preferred embodiment, a hash function is optionally used to derive an index value for each possible pixel row. The partial result associated with an equal index value is accessed. In another, equally preferred embodiment, double buffer stores two consecutive input pixel windows for processing. After each row of the first buffer is processed, the double buffer is shifted by one pixel. After the entire contents of the second buffer have been shifted into the first buffer, the next input pixel window is moved into the second buffer. Thus, the invention requires only one main memory access for each input pixel window. The partial convolution results are preferably calculated in floating point arithmetic and converted to a fixed point precision before they are stored as table entries. |
