Image Denoising and Dimensionality Reduction Using Autoencoder
Gedela Triveni, Sunita Nayak, Jagannath Padhy, Cherukuri Gunalakshmi, Megha Sehgal, Sudheer Choudari · 2024
A major problem in domains like computer vision, photography, and medical imaging is picture noise, which can significantly reduce image quality due to variables like inadequate sensors, dim lighting, or electronic interference. Images can be effectively denoised by autoencoders, especially Convolutional Neural Network (CNN) autoencoders. While filtering out noise, these neural networks preserve important information by compressing noisy images into a lower-dimensional feature space. The compressed form is then used by the decoder to recreate the denoised image. For picture denoising in this investigation, a CNN autoencoder with convolutional, max-pooling, and upsampling layers was used. The design consists of a decoder that reconstructs the image, an encoder that extracts features, and a bottleneck layer with 128 filters that compresses the image data. In order to lower the Mean Squared Error (MSE) between the original and denoised images, the model was trained over five epochs. The efficacy of the model was demonstrated by the consistent reduction in loss during training that this method produced. A noteworthy accomplishment of this approach was the decrease in storage needs, as the image size shrank from 1121.16 MB to 560.58 MB, indicating the autoencoder's capacity to compress data while maintaining crucial image characteristics. For domains like data storage, medical imaging, and image restoration, this effective trade-off between storage capacity and image quality has significant ramifications. The efficiency and usability of the model could be further improved by future research concentrating on real-time denoising capabilities, alternative loss function exploration, and model configuration optimization.