Automated Image Preprocessing Pipelines forLarge-Scale Datasets Based on Deep LearningModels

Abstract

Improving the effectiveness of deep learning (DL) for image classification on large datasets requires not only data
preprocessing but also the right network designs. In this research, a new image classification system is introduced, which applies a
modified Inception model to a subset of the ImageNet dataset containing 10 classes. Preparation of the data includes cleaning,
augmentation, normalization, and feature extraction. These activities not only preserve the quality of data but also provide efficient
input to the model. The modified Inception architectural changes enhance the original Inception network idea, thus increasing the
power of the network for feature extraction and data classification. To evaluate the results, besides the modified Inception model,
several well-known DL architectures like DenseNet, ResNet50, and VGG16 have been employed and compared using common
performance metrics. The modified Inception model achieves better performance in the classification tasks than the baseline models.
The accuracy is at 98.7%, the precision at 98.5%, the recall at 99%, and the F1-score at 98.6%. The results reported here have
highlighted the role of architectural modifications and the impact of the thorough preprocessing pipeline in boosting the performance
of image classification tasks on large datasets. Such a system offers a viable means to address challenging visual recognition problems.