A Novel Image-Invariant Feature Extraction Using Sliding Window for Machine Learning

Feature extraction is an important task in building machine learning and deep learning applications. A gray-level cooccurrence matrix, a histogram of oriented gradients, a local binary pattern, principal component analysis, and linear discriminant analysis are some of the feature extraction methods that are used in a lot of research studies. However, these methods are sensitive to image quality, have trouble with non-linear relationships, are difficult to compute, and cannot capture global or contextual information. These limitations often require additional preprocessing or modifications to enhance their performance in practical applications. The goal of this study is to come up with a feature extraction method that works with all kinds of image changes and can pick up both local and global features to make machine learning classifiers work better. To achieve this, we propose in this research a new feature extraction method that is based on the concept of a sliding window to extract local and global image-invariant features. For evaluating the proposed method, we have used the chest X-Ray medical images from the publicly available Novel COVID-19 Chest X-Ray Repository dataset at Kaggle. We conducted experiments using five benchmark feature extraction methods and eight state-of-the-art machine learning classifiers to assess the significance of the proposed feature extraction. For binary classification, the tests indicated that MLP had better accuracy, recall, precision, specificity, and balanced accuracy than other methods (96.25% for accuracy, 96.05% for recall, 92.4% for precision, 96.34% for specificity, and 96.19% for balanced accuracy). The dense MLP neural network, which has two hidden layers with 1024 and 512 neurons each, was able to correctly classify with a 93.98% accuracy, 92.07% recall, 93.21% precision, and 95.32% specificity. It also had a balanced accuracy of 93.69% when it came to multiclass classification.