Deep learning has been used in many applications where patterns from past-trained data can be extracted to predict future outcomes. Deep learning is characterized by training and testing data with the identical input feature space and same data distribution. However, whereas the data distribution is same between the training and testing data, the results might be different. This study introduces input image preprocessing, an enhanced neural network optimization method, and prediction probability ensemble to minimize the number of trainable parameters but maintain the outcome accuracy. In the suggested methodology, input images are separated into pixel interval and the fully connected layer jointly used with saved weights. Outcome results of separated input images are ensembled to the corresponding class probabilities of the original image. The results of the proposed method were compared with those of other previous methods in the image classification task and achieved successful performance accuracies in several datasets.

The application of computer vision (CV) in many fields is becoming an important factor in development, and the demand for high accuracy CV products is increasing. Therefore, a new method that ensembles three trained models with the same architecture but different inputs is proposed. Each of the different channels of the images were used as input. The first channel of the image was used in the dataset for the first model, the second channel was used for the second model, and the third channel was used for the third model. The classification probabilities were achieved after training three different models with three channels. The model with the highest accuracy was selected as the main model, and it was added to the classification probabilities of certain classes of the main model. Applying this method to the classification task resulted in 75.45% accuracy, while using the entire image for the model resulted in 70.9% accuracy. The advantage of using this ensemble method is that it can be used simultaneously with other ensemble methods and achieves better results. Dividing an image into three different channels than simply using the entire image helps the model learn the image better.