An Image Pixel Interval Power (IPIP) Method Using Deep Learning Classification Models

The implementation of deep learning (DL) in various fields is becoming common. In addition, demand for higher accuracy models is increasing continuously at the same rates as the growth of other fields of science. Using all DL tools in the development of computer vision (CV) is a fundamental aspect of its future development. Considering all such tools, we conducted research on the effect of data representation on the final classification accuracy and proposed image pixels’ double representation (IPDR) and image pixels’ multiple representations (IPMR) for skipping certain pixels in the images in a dataset. Because the image pixel values range from 0 to 255, we proposed including all knowledge from different intervals of pixels. With IPDR, we trained the model using a dataset and obtained the prediction probabilities for the classification task. Next, we created two different datasets from an existing dataset. The first dataset took only image pixels of higher than 127, with all other image pixels in the dataset changed to zeros. The second dataset took only image pixels equal to or lower than 127. These two created datasets were trained on the same model architecture and their prediction accuracies for classification were ensembled with the prediction accuracies of the main model. With the IPMR method, we applied the same method as previously described, although instead of two intervals, from 0 to 127, and 127 to 255, we used, multiple intervals of 50 (i.e., [0:50], (50:100], (100:150], (150:200], and (200:255]) for the Cifar10 dataset. The number of intervals depends on the dataset, and applying our method, we achieved 89.46%, 98.90%, and 73, 38% accuracies on the Fashion MNIST, MNIST, and Cifar10 datasets, respectively, whereas their original classification accuracies under classic training were 89.27%, 98.65%, and 71.29%, respectively. As the advantage of using this method, it can be applied to any classification task and gives only extra knowledge on the trained data. As another simplicity of this method, it can be used with other DL ensemble models simultaneously.