A Learning-rate Optimization Technique for Object Detection Accuracy Enhancement

Abstract

Recently, Deep Learning [1] models are used primarily in Object Detection algorithms because of their specific capability for Image Recognition. These models identify items present in input images and videos [2] by extracting features from them. These models have a variety of applications, which include Image Processing, Video analysis, Speech Recognition, Biomedical Image Analysis, Biometric Recognition, Iris Recognition, National Security applications, Cyber Security, Natural Language Processing [3], Weather Forecasting applications, Renewable Energy Generation Scheduling etc. The Convolution Neural Network (CNN) [3], which comprises many artificial neuron layers, is employed for these models. The accuracy of Deep Learning models is determined by a number of factors, including the learning rate, the training batch size, the validation batch size, the activation function, and the drop-out rate. Hyper-Parameters are the name for these parameters. The accuracy of Object Detection depends on the choice of Hyper-Parameters. It is therefore a difficult task to find the best values for these parameters. Fine-Tuning is a method for selecting an effective Hyper-Parameter for improving Object Detection precision. Selecting an inaccurate Hyper-Parameter value, leads to Over-Fitting or Under-Fitting of data. Over-Fitting is a problem, when training data is greater than the necessary, leading to learning noise and inaccurate Object Detection [4]. Under-Fitting occurs when a model is unable to capture the data's trend, resulting in more erroneous testing or training outcomes. By changing the ‘Learning rate' of various Deep Learning Models, a balance between Over-Fitting and Under-Fitting is reached in this article. For experimentation purpose, this paper considers four Deep Learning Models such as VGG-16, VGG-19, InceptionV3 and Xception. In terms of maximal Object Detection accuracy, the best zone of Learning-rate for each model is analyzed.  The prediction accuracy of a dataset of 70 object classes is investigated in this study by adjusting the ‘Learning-Rate' while keeping the rest of the Hyper-Parameters fixed.This article focuses on the impact of ‘Learning-Rate' on accuracy in Object Detection and identifies an ideal accuracy zone.  This analysis helps in reduction of computational effort in calculation of Objection Detection Accuracy.