Handwritten bangla-digit-recognition-using-deep-learning
- 1. Submitted By Subroto Saha Shuvro Md: Rumman Islam Nur Abdul Momin Department OF CSE WUB Supervise By Md. Ashiqur Rahman Senior Lecturer Department OF CSE WUB
- 2. Deep Learning is a subfield of machine learning concerned with algorithms inspired by the structure and function of the brain called artificial neural networks. Ref:google.com
- 3. The traditional neural network consists of at most 2 layers and this type of structure of the neural network is not suitable for the computation of larger networks. Therefore, a neural network having more than 10 or even 100 layers is meant for Deep Learning. In this, a stack of the layer of neurons is developed. The lowest layer in the stack is responsible for the collection of raw data such as images, videos, text, etc. Each neuron of the lowest layer will store the information and pass the information further to the next layer of neurons and so on. As the information flows within the neurons of layers hidden information of the data is extracted. So, we can conclude that as the data moves from lowest layer to highest layer (moving deep inside the neural network) more abstracted information is collected.
- 4. Ref:google.com Automatic handwriting character recognition is of academic and commercial interests. Current algorithms are already excel in learning to recognize handwritten characters. The main challenge in handwritten character classification is to deal with the enormous variety of handwriting styles by different writers in different languages. Furthermore, some of the complex handwriting scripts comprise different styles for writing words. Depending on languages, characters are written isolated from each other in some cases, (e.g., Thai, Laos and Japanese). The large variety of writing styles, writers, and the complex features of handwritten characters are very challenging for accurately classifying the hand written characters.
- 5. Ref:google.com Support vector machine (SVM) : In machine learning, support vector machines are supervised learning models with associated learning algorithms that analyze data used for classification and regression analysis. Deep Belief Network (DFN): is a generative graphical model, of deep neural network, composed of multiple layers of latent variables ("hidden units"), with connections between the layers. Convolution neural network (CNN) : is a type of deep learning model that combines a deep CNN with Q- learning, a form of reinforcement learning.
- 6. Ref:google.com Gaussian Filter: Gaussian being or having the shape of a normal curve or a normal distribution. Gabor Filter: Gabor, is a linear filter used for texture analysis. Dropout is a regularization technique for reducing over fitting in neural networks.
- 7. Restricted Boltzman Machine (RBM): A restricted Boltzmann machine (RBM) is a generative stochastic artificial neural network that can learn a probability distribution over its set of inputs. CNN with Dropouts: Dropouts in CNN is very powerful training techniques usually used for fully connected layers.
- 13. • Dataset Description • CNN structure & parameters setup • DBN structure & parameters setup • Performance Evaluation • Comparison with the state of the arts
- 14. In this research, we proposed to use deep learning approaches for handwritten Bangla digit recognition(HBDR). We evaluated the performance of CNN and DBN with combination of dropout and different filters on a standard benchmark dataset: CMATERdb 3.1.1. From experimental results, it is observed that CNN with Gabor feature and dropout yields the best accuracy for HBDR compared to the alternative state-of-the-art techniques. Research work is currently progressing to develop more sophisticated deep neural networks with combination of State Preserving Extreme Learning Machine (Alom et al., 2015) for handwritten Bangla numeral and character recognition.