A Learning-rate Optimization Technique for Object Detection Accuracy Enhancement | Chapter 14 | New Approaches in Engineering Research Vol. 14

Because of their specific capability for Image Recognition, Deep Learning [1] models have recently been used primarily in Object Detection algorithms. These models extract features from input images and videos [2] to identify items present. 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, and other applications are all possible with these models. These models use the Convolution Neural Network (CNN) [3, which consists of several artificial neuron layers. The learning rate, the training batch size, the validation batch size, the activation function, and the drop-out rate are all aspects that influence the accuracy of Deep Learning models. These parameters are known as hyper-parameters. The accuracy of Object Detection is determined by the Hyper-Parameters used. Finding the appropriate settings for these parameters is thus a complex undertaking. Fine-tuning is a technique for determining which Hyper-Parameters are most beneficial in enhancing Object Detection precision.

Selecting an incorrect Hyper-Parameter value causes data to be over-fitted or under-fitted. When training data is higher than what is required, over-fitting occurs, resulting in learning noise and erroneous Object Detection [4]. When a model is unable to capture the data's trend, under-fitting occurs, resulting in more erroneous testing or training outcomes.

This article achieves a balance between Over-Fitting and Under-Fitting by adjusting the 'Learning rate' of various Deep Learning Models. This study considers four Deep Learning Models for experimentation: VGG-16, VGG-19, InceptionV3 and Xception. The best zone of Learning-rate for each model is examined in terms of maximum Object Detection accuracy. This study investigates the prediction accuracy of a dataset of 70 object classes by altering the'Learning-Rate' while keeping the rest of the Hyper-Parameters constant. The impact of 'Learning-Rate' on Object Detection accuracy is discussed in this article, which also suggests an optimum accuracy zone. This approach aids in the computation of Objection Detection Accuracy while reducing computational effort.

Author (S) Details

Chamarty Anusha
Computer Science & Systems Engineering, Andhra University College of Engineering, Andhra University, Visakhapatnam (AP), India.

P. S. Avadhani
Computer Science & Systems Engineering, Andhra University College of Engineering, Andhra University, Visakhapatnam (AP), India.

View Book:-  https://stm.bookpi.org/NAER-V14/article/view/3533