Visual Recognition of Local Kashmiri Objects with Limited Image Data using Transfer Learning

Learning to recognize object categories is a challenging task for computers and the task becomes more difficult if the image data is small in size. Traditional machine learning methods require extensive training data to generalize and produce accurate results. Seeking inspiration from human perception, it has been found using prior knowledge about related tasks helps in learning new tasks. Transfer Learning is based on this natural learning process and can help to reproduce the remarkable human capability of recognizing objects from just one single view. In this manuscript we explored transfer learning techniques along with state-of-the-art object recognition models to discover improved ways of performing Visual Object Recognition for Object categories with limited image data. A small data set was built from scratch having images for four local object categories and then a model was developed using pre trained Inception-v3 model for classifying them. The results were compared with stock 3 Layer Convolutional model. The proposed model obtained a respectable accuracy result of around 90% while the stock model had an accuracy of 70%. Considering the fact that the dataset used is very tiny (training portion of the data set had only 320 images for all categories, 80 per category) the results obtained are encouraging. Thus, this work further strengthens the fact that transfer-based techniques can be utilized for computer vision tasks with limited data.