基于积分图的LATCH算法改进

Improved LATCH Based on Integral Image

LATCH(基于机器学习的三像素块描述子)在原有局部二值描述子的基础上通过将二像素块比较改为三像素块提升了局部二值描述子的准确率,但提高准确率的同时带来了更大的时耗,在研究LATCH以及其他二值描述子的基础上,借鉴积分图在目标检测中的应用,将积分图的思想应用在改进LATCH描述子中,减少了LATCH描述子中各像素块内的重复计算量.实验证明,改进算法的描述子计算时间较原算法缩减了30%–40%,而配准精度与原算法保持相近.

LATCH （Learned Arrangements of Three Patch Codes） improves the accuracy of local binary descriptor by comparing three pixel blocks rather than tow pixel blocks. However, the improvement of accuracy brings a larger time consuming. Based on the study of LATCH and other local binary descriptors, using integral graph theory in improved LATCH descriptor, it reduces repeated calculation of each pixel blocks in LATCH descriptors. According to the experiment results, the computation time of the improved algorithm is reduced by 30%-40% compared with the original algorithms, while the accuracy of the improved algorithm is similar to that of the original algorithm.