Improving Deep Learning-Based Image Classification Through Noise Reduction and Feature Enhancement

The proposed study is related to the development of deep learning-based image classification with the help of various pre-processing techniques applied within the RESNET50 architecture. The purpose of this research is to check the impacts of different pre-processing techniques like noise reduction, histogram equalization, random cropping, and rotation on training efficiency and accuracy of the model. We had to reduce the complexity of configuration of this system and speed up training: we used the open-source deep learning framework PyTorch and CUDA-enabled GPUs; each experiment had a consistent dataset. Our results demonstrate how pre-processing techniques dramatically improve the model's performance: in our case, the combination of techniques doubles the accuracy achievable with the best possible baseline model, versus only attaining 83% for the baseline. Accuracy of validation further exhibits the advantage of learning through pre-processed images by having better accuracy rates for models at each training step. Normalization was found as the most influential factor in improving convergence and generalization ability on unseen data. Overall, results manifest the optimality that very well-chosen pre-processing strategies can provide when used as precursors to deep learning-based image classification.