Lightweight user-adaptive handwriting recognizer for resource constrained handheld devices

Abstract

Here, we present our recent attempt to develop a lightweight handwriting recognizer suitable for resource constrained handheld devices. Such an application requires real-time recognition of handwritten characters produced on their touchscreens. The proposed approach is well suited for minimal user-lag on devices having only limited computing power in sharp contrast to standard laptops or desktop computers. Moreover, the approach is user-adaptive in the sense that it can adapt through user corrections to wrong predictions. With an increasing number of interactive corrections by the user, the recognition accuracy improves significantly. An input stroke is first re-sampled generating a fixed small number of sample points such that at most two critical points (points corresponding to high curvature) are preserved. We use their x- and y-coordinates as the feature vector and do not compute any other high-level feature vector. The squared Mahalanobis distance is used to identify each stroke of the input sample as one of several stroke categories pre-determined based on a large pool of training samples. The inverted covariance matrix and mean vector for a stroke class that are required for computing the Mahalanobis distance are pre-calculated and stored as Serialized Objects on the SD card of the device. A Look-Up Table (LUT) of stroke combinations as keys and corresponding character class as values is used for the final Unicode character output. In case of an incorrect character output, user corrections are used to automatically update the LUT adapting to the user's particular handwriting style.