Robust Filtering Technique for CNN-Based Image Classification under High-Density Salt and Pepper Noise Interference

The performance of deep learning convolutional neural network (CNN) could suffer severe misclassification and performance degradation under high-density salt and pepper noise interference. The investigation studies the classification performance of AlexNet Deep Learning Network (DLN) which is a type of deep learning CNN consisting of 8-layer deep. Conventionally median filtering (MedF) could be applied to filter out the salt-and-pepper noise to improve the image classification performance of the AlexNet. However, MedF may fail to work under high-density salt and pepper noise interference. Hence, a robust recursive weighted myriad filtering (RWMyF) for solving the problem is proposed in this work. A novel and robust filtering system model for CNN-based image classification is devised to mitigate the high-density salt and pepper noise interference for better and more reliable performance. Overall, higher peak signal-to-noise ratios (PSNRs) are obtained from the robust RWMyF filtering for Alexnet with more precise classifications obtained for those images severely distorted by high-density salt and pepper noise.