Comparitive Analysis of Various Transfer Learning Apporaches in Deep CNNs for Image Classification

Abstract

Deep learning has become an effective approach over the past few years to addressing intricate computer vision problems, and Convolutional Neural Networks (CNNs) have been the primary driving force behind this progress. Developing CNNs, however, comes with the obstacle of requiring huge, labeled datasets. Gathering and annotating a large dataset for any specific job is costly and time-consuming. To overcome this challenge, researchers can employ transfer learning, a technique that involves using pre-trained deep learning models on extensive datasets. This study primarily aims to investigate the application of various transfer learning methods in conjunction with Deep Convolutional Neural Networks (CNNs) for image classification. The research utilizes the Visual Object Classes Challenge 2012 (VOC2012) dataset as the foundation for its analysis. To classify a diverse range of object images, the study applies well-established transfer learning techniques, specifically fine-tuning pre-trained CNN models. Model performance is assessed through metrics like FPS (frames per second) and mAP% (mean average precision). A variety of models, including VGG19, SqueezeNet, YOLOv7, and Inception-ResNet-v2, are tested to determine the most suitable model for the dataset. The study’s findings demonstrate that employing transfer learning with CNNs can substantially enhance image classification accuracy on the dataset while reducing model development time. Notably, the Inception-ResNet-v2 model emerged as the top-performer, achieving an mAP score of 77%. This research underscores the potential of transfer learning as a powerful tool in the realm of image classification.