Ensemble of top3 prediction with image pixel interval method using deep learning

Computer vision (CV) has been successfully used in picture categorization applications in various fields, including medicine, production quality control, and transportation systems. CV models use an excessive number of photos to train potential models. Considering that image acquisition is typically expensive and time-consuming, in this study, we provide a multistep strategy to improve image categorization accuracy with less data. In the first stage, we constructed numerous datasets from a single dataset. Given that an image has pixels with values ranging from 0 to 255, the images were separated into pixel intervals based on the type of dataset. The pixel interval was split into two portions when the dataset was grayscale and five portions when it was composed of RGB images. Next, we trained the model using both the original and newly constructed datasets. Each image in the training process showed a non-identical prediction space, and we suggested using the top three prediction probability ensemble technique. The top three predictions for the newly created images were combined with the corresponding probability for the original image. The results showed that learning patterns from each interval of pixels and ensembling the top three predictions significantly improve the performance and accuracy, and this strategy can be used with any model.