Reconstruction of Image using Swin Transformer Encoder and Convolutional Neural Network Decoder

In recent years, reconstruction of image has crucially advanced because of Deep Learning (DL) and AI-driven techniques. Traditional methods such as Fourier transform is replaced by Neural Networks (NN) which are used to predict the missing information with incredible accuracy. Even though, previous researchers have suggested various advanced model still there are challenges such as computational complexity and ineffective image reconstruction. Therefore, this research proposes a robust model for image reconstruction using the proposed Swin-CNN model. Initially, the data is collected from Heron Island Coral Reef Dataset (HICRD) that contains underwater images collected in various conditions. Further, the collected data is embedded into preprocessing stage which utilizes data augmentation and min-max scaling for improving model generalization and to ensure uniformity in pixel intensity respectively. Additionally, Swin transformer encoder is employed for extraction of features through hierarchical processing and finally, CNN decoder is incorporated for reconstruction of image effectively. The proposed swin-CNN acquired greater results in terms of Structural Similarity Index (SSIM) of 0.9351, Peak Signal-to-Noise Ratio (PSNR) of 34.9713 and Learned Perceptual Image Patch Similarity (LPIPS) 0.0501 respectively when compared to existing Denoising Diffusion Probabilistic Models (DDPM) model.