Novel deep learning frameworks for image classification-based computer vision tasks

Abstract

In recent years, there has been a notable surge in interest from the research community regarding the improvement and refinement of image classification techniques. The thesis at hand specifically delves into two pivotal areas within these domains: vehicle make and model recognition, and the classification of skin cancer. These areas have been selected due to their critical importance and the potential impact that advancements in image classification methodologies can have within these fields. Vehicle make and model recognition is crucial for automatic vehicle recognition systems. It has received significant attention from the computer vision and artificial intelligence communities in recent decades. Previous research has focused on improving recognition accuracy through attention mechanisms, which, while effective, often increase complexity and computational burden unnecessarily. A novel deep neural network model, named the Simple Sequential Attention Network (Sim- SANet), is proposed to address this challenge. SimSANet integrates a sequential multi-kernel approach to balance complexity and performance, effectively reducing computational load while efficiently capturing essential information from feature maps of varying scales. The proposed approach exhibits superior performance, as evidenced by experimental results on publicly accessible VMMR datasets such as Stanford Cars, CompCars, CompCarsSV, and the Vehicle Images dataset. For medical contexts like skin cancer classification, the accurate identification of skin abnormalities is imperative for effective diagnosis and treatment. Traditional convolutional neural network based ensemble approaches, while effective, often incur high computational costs due to the combination of multiple models. In contrast, snapshot-based approaches offer a streamlined alternative by utilizing a single baseline model. To mitigate computational burden without sacrificing performance, SEFFNet, a novel snapshot ensemble-based feature fusion network coupled with a machine learning-based classifier, is proposed. SEFFNet captures diverse feature representations at different training stages, eliminating the need for multiple models and refraining from data augmentation and pre-processing techniques. The effectiveness of SEFFNet is demonstrated through experimental assessments conducted on publicly accessible datasets such as HAM10000 and ISIC 2017.