IMAGE CLASSIFICATION THROUGH TRANSFER LEARNING & FINE TUNING

Abstract

The advancement of sophisticated scientific techniques in today's world is pushing the limits of human outreach in various fields oftechnology further. One such field is that commonly known as OCR (Optical Character Recognition) field of character recognition. The goal ofthis project is to develop algorithms for image recognition to identify the handwriting on postal letters and parcels. This article explores the various techniques used for handwriting recognition. Different stages involving the processing of images, such as pre-processing, extraction of features and prediction will be studied and addressed. Finally, the algorithm end product will be written in the software named Ying. The program is created using the Artificial Neural Network technique. The main advantage of using this strategy is that it offers extraction and identification capabilities which are ideal for recognition of character. VGG16, Convolutionary Neural Network (CNN) architecture, model is addressed and Error-back propagation algorithm was used because of its capacity to shape internal character representations in classification. A suitable collection oftraining parameters is specified after trials and errors and a network configuration is generated that consists of 1 input layer, 2 hidden layers and 1 output layer with 69 input neurons, 324 neurons for both hidden layers and 38 neurons for output layer. The system initially proceeds with threshold, inverting, and smoothing the pre-process of the captured image. In the phase sampling resizing, and selection of features are often done. The feed forward process is then invoked through the network to yield an output matrix. The recognized character can be determined based on output matrix. This system is designed to personalize an individual user's network. The report also contains recommendations for future development and conclusions