Real-time Recognition of Handwritten Numbers: A LeNet-5 Based Approach and Its Improvements

The main idea of this project is to use edge detection to construct the shapes of each number and segregate each contour into separate images in order to isolate the numbers. Then feeding the images through a neuro network to perform recognition. Finally, the information is printed onto a video showing both the paper and boxes, boxing the individual numbers. The results achieved are fairly accurate, however needs improvements on the accuracy of the number seven. The goal of this project is to enable the real-time recognition of various handwritten digits presented on a physical paper medium. This method places no restrictions on the colors of the paper or pen, as long as the digits remain legible. The dataset employed originates from the MNIST collection, and the model architecture used closely mirrors that of LeNet-5. The project is categorized into three main sections: model development, image parameter assessment, and implementation for real-time video analysis.