A Robust System for Noisy Image Classification Combining Denoising Autoencoder and Convolutional Neural Network

Abstract

Abstract-Image classification, a complex perceptual task with many real life important applications, faces a major challenge in presence of noise. Noise degrades the performance of the classifiers and makes them less suitable in real life scenarios. To solve this issue, several researches have been conducted utilizing denoising autoencoder (DAE) to restore original images from noisy images and then Convolutional Neural Network (CNN) is used for classification. The existing models perform well only when the noise level present in the training set and test set are same or differs only a little. To fit a model in real life applications, it should be independent to level of noise. The aim of this study is to develop a robust image classification system which performs well at regular to massive noise levels. The proposed method first trains a DAE with low-level noise-injected images and a CNN with noiseless native images independently. Then it arranges these two trained models in three different combinational structures: CNN, DAE-CNN, and DAE-DAE-CNN to classify images corrupted with zero, regular and massive noises, accordingly. Final system outcome is chosen by applying the winner-takes-all combination on individual outcomes of the three structures. Although proposed system consists of three DAEs and three CNNs in different structure layers, the DAEs and CNNs are the copy of same DAE and CNN trained initially which makes it computationally efficient as well. In DAE-DAE-CNN, two identical DAEs are arranged in a cascaded structure to make the structure well suited for classifying massive noisy data while the DAE is trained with low noisy image data. The proposed method is tested with MNIST handwritten numeral dataset with different noise levels. Experimental results revealed the effectiveness of the proposed method showing better results than individual structures as well as the other related methods.

The Medical Image Denoising Method Based on the CycleGAN and the Complex Shearlet Transform

“…In Equation (2), u is the real image, u′ is the generator-generated image, and n is the number of a batch in training. www.ijacsa.thesai.org After obtaining the four losses, the loss loss_G of the noise extractor model can be obtained by Equation ( 3) to (7).…”

“…Nowadays, many advanced denoising methods have been proposed, all of which can be generally divided into four categories: the filter-based methods [4], the model-based methods [5], the multi-scale geometric transform-based methods [6] and the deep learning-based methods [7,8]. For the filter-based methods, they typically implement the lowpass filters to replace the noisy or suspected pixel by their locally averaging value or energy in the neighboring region.…”

Application of Machine Learning Techniques in Injection Molding Quality Prediction: Implications on Sustainable Manufacturing Industry

“…An autoencoder consists of an encoder that maps the input to the hidden layer and a decoder that maps the encoded data back to the reconstruction [49]. First, it compresses the original input data to a vector of lower dimension and then decodes this vector to the original representation of the data [50]. A stacked autoencoder is an autoencoder with multiple hidden layers.…”

Visible watermarks inpainting using deep learning for multimedia application

“…Autoencoders with convolutional layers were developed based on previously used autoencoders with fully connected layers. In some cases, autoencoders built based on convolutional neural networks make it possible to reduce noise in the output images [15,16]. Unlike GANs, which require data labeling and the creation of a mask indicating an area with an object overlapping the image, autoencoders do not require such labeling.…”