Image compression approach for improving deep learning applications

Abstract

In deep learning, dataset plays a main role in training and getting accurate results of detection and recognition objects in an image. Any training model needs a large size of dataset to be more accurate, where improving the dataset size is one of the most research problems that needs enhancement. In this paper, an image compression approach was developed to reduce the dataset size and improve classification accuracy for the trained model using a convolutional neural network (CNN), and speeds up the machine learning process, while maintaining image quality. The results revealed that the best scenario for deep learning models that provided good and acceptable classification accuracy was one that had the following parameters: 80×80 image size, 10 epochs, 64 batch size, 40 images dataset quality (images compressed 60%), and gray image mode. For this scenario a Dog vs Cat dataset is used, and the training time was 48 minutes, classification accuracy was 86%, and images dataset size was 317 MB on storage device. This size makes up 58% of the size of the original image’s dataset, saves 42% of the storage space and reduces the processing resources consumption.