LeanConv-SENet: A novel lightweight Neural Network Architecture for Handwritten character recognition

Recognition of handwritten characters and conversion into digital form is a requirement of various domains like healthcare, banking, education, etc. Deep learning models have shown high utility to achieve this requirement; these models need to be fast and lightweight to support all types of system configurations and processing needs. This study proposes a novel Convolution Neural Network architecture to recognize handwritten Chinese numbers titled LeanConv-SENet (Lean Convolutional Squeeze-Excite Network). LeanConv-SENet is based on Squeeze Excite blocks and Cheap Convolution layers and focuses on the training process with minimal memory requirement and computational complexity while maintaining the classification accuracy of handwritten numbers. 15,000 grayscale images of Chinese numbers from the MNIST dataset are used for experimentation. When compared with ANN and standard CNN architectures, the proposed LeanConv-SENet has shown similar results with a 92% reduction in trainable parameters. This study proposes an architecture with the potential to be deployed on mobile devices and machines with limited hardware configuration without compromising classification efficiency.