Novel Framework for Image Classification Based on Patch-Based CNN Model

Image data is being developed at a rapid speed every day due to the speedy growth of Internet; resultantly, image classification has become very crucial topic in solving computer vision-related tasks like images segmentation and object detection, etc. The convolutional neural networks (CNN) have become the preferred choice for categorizing images since 2012, because of its capability to find out hierarchical representations from the original image data without human intervention. CNNs are excellent at deciphering the intricate visual features present in images because they can capture local patterns and gradually build up more abstract and high-level representations. In several image processing tasks, patch-based solutions performed better than rival approaches. This research work illustrates the patch-based CNN model to increase picture classification accuracy, which may also serves as a model assembly without supplementary cost of model training. At the same time, applying data augmentation techniques throughout the training as well as testing phases helps assure network optimization as well as improvise generalization of the model. This methodology can transmit specific domain information effectively by processing individual patches instead of the entire image. To increase training samples, essential patches of appropriate size were extracted from the original images, improving overall model performance. Additionally, our approach is applicable to various network architectures and data augmentation techniques, making it versatile for deep learning tasks. Thorough evaluations on a real-world dataset demonstrate the superior performance of our proposed framework which also achieves an accuracy of 96.21% on the CIFAR-10 dataset, surpassing the state-of-the-art results (93.41%).