Introduction The objective of this project is to design and implement an order generation application that detects a hand written grocery shopping list and converts it to a digital item list using linear SVM algorithm coupled with HOG feature extraction. The data set contains different 62 different characters (English alphabets capital and small, Digits 0-9) written in 55 different styles. In particular, we apply the tools of machine learning to detect the handwritten characters. We have used segmentation, HOG feature extraction and linear SVM model to obtain better performance from any of the constituent machine learning algorithms.

Problem Definition and Algorithm The input here is the images of characters written in 55 different styles. The algorithm learns the original character for the different styles of characters in the phase by using the HOG features to label the characters. This is later used when a character in a different style unseen by the algorithm is given as input. The final output, the digitalized form of the recognized characters, is obtained by extracting its HOG features and detecting the characters by matching the HOG features from the trained linear SVM model.

Tools Used: OpenCV2 Python 3.0 Libraries: Scikit Learn for HOG, SVM, KNN

Dataset: The Chars74K dataset http://www.ee.surrey.ac.uk/CVSSP/demos/chars74k/EnglishHnd.tgz

Experiment Evaluation We experimented the recognition with SVM and KNN classifiers. Accuracy of SVM is approximately 60% whereas KNN accuracy was very low.

KNN is slow because the model cannot be trained and saved, every time it trains based on the feature generated. Linear SVM saves the trained model and uses it for training directly. SVM gives good better performance than KNN classifier.