Elinda Kajo Meçe Journal of Engineering Research and Application www.ijera.com ISSN: 2248-9622 Vol. 10, Issue 01 (Series -III) January 2020, pp 08-13 www.ijera.com DOI: 10.9790/9622-1001030813 8 | Page Towards Guidelines for Assembling Image Datasets for Transfer Learning Techniques Elinda Kajo Meçe*, Evis Trandafili*,Sibora Theodhor* *Department of Computer Engineering, Faculty of Information Technology, Polytechnic University of Tirana, Albania Corresponding Author; Elinda Kajo Meçe ABSTRACT With the latest breakthroughs in deep learning, the field of machine learning for image recognition has been attracting increasing attention. Neural networks for image similarity and image classification problems have yielded impressive results. Another successful breakthrough on the field is the application of transfer learning, to reduce training time. Collecting data and preprocessing datasets is the most expensive task of the process. Our work aims at observing the transferability of the dataset characteristics and outlining guidelines for the effort of image data collection on a transfer learning scenario. Keywords - Machine Learning, Transfer Learning, Image Recognition, Classification, CNTK, ImageNet --------------------------------------------------------------------------------------------------------------------------------------- DATE OF SUBMISSION: 27-12-2019 DATE OF ACCEPTANCE: 15-01-2020 -------------------------------------------------------------------------------------------------------------------------------------- I. INTRODUCTION In the last few years machine learning and the field of computer vision have made great progress. Models used to recognize images like neural networks were first described by Patterson, Dan W. in 1930 [15] but it’s only in the last four or five years that the hardware advancements like GPU utilization have made it feasible to effectively run and discover their full potential. This evolution of the hardware combined with several breakthroughs in the field of computer vision, like Krizhevsky, Sutskever, and Hinton [2] have laid out the path for tackling new challenges in the field of image recognition and image classification. Deep learning and models like convolutional neural network (CNN), has proven to achieve good performance in difficult visual recognition tasks - matching the human performance in some fields [17]. The layer of neurons in the CNNs are designed similarly to how vision in animals and humans works. One of the challenges of scaling such techniques is the hardware requirements. Toolkits like Theano, Tenserflow, CNTK are making use of GPU programming and leveraging the highly parallel computational architecture of the GPUs to parallelize the many matrix operations utilized on various computer vision techniques. The need to overcome the high costs sacrificing little accuracy spawned the efforts that produced the next breakthrough: transfer learning [22]. The transfer learning approach is to get a pre-trained model (parameters and weights of an NN that has been trained with large dataset by others) and regulating the model with your dataset in the last layers [1]. By using transfer learning a significant amount of labeling effort can be saved, reducing the data gathering cost and most importantly saving a lot of training time. Similar work has been conducted by [Bengio, Yosinski] who explored and indicated that the features of image recognition and classification techniques are transferable. On our paper, we compose a set of experiments to investigate how two characteristics of datasets: the number of samples, and the ratio of sample affect precision and accuracy on image classification through transfer learning. We generate several datasets of images, that are subclasses of ImageNet and take a systematic approach at determining on whether the task of collecting images should continue or it is sufficient. Also, we make observations about the ratio of the positive samples and negative samples concluding with a discussion of future work. OVERVIEW A. Image Recognition/Classification Image Classification involves taking a set of pixels representing a single image and as signing a tag to it from a given set of categories. This is one of the essential problems in Computer Vision that has many practical applications [11, 12, 13]. The first step of Image Classification process consists on the input, often referred as the training set. A set of M images, each labeled with one of K different classes. The next step is training RESEARCH ARTICLE OPEN ACCESS