Ứng dụng mạng sâu và phép biến đổi Fourier trong khử nhiễu ảnh Application of deep learning and Fourier transformer to image denoising

0 Người đánh giá. Xếp hạng trung bình 0

Tác giả: Đức Lưu Ngô

Ngôn ngữ: vie

Ký hiệu phân loại:

Thông tin xuất bản: Tạp chí Khoa học Trường Đại học Bạc Liêu, 2023

Mô tả vật lý: tr.36

Bộ sưu tập: Metadata

ID: 345652

Thêm vào giỏ Liên kết toàn văn

Image Denoising has become a topic of great interest in various computer vision fields thanks to its practical applications. Images are often affected by noise due to various factors such as environment, lighting conditions, and camera quality. Denoising helps to increase the accuracy and reliability of these images, thereby improving their analysis, recognition, and classification. Some previous image denoising methods focus on extracting features directly from the image in the spatial domain. However, restoring a clearly image from a noisy image captured by a camera requires more information about the noise frequency. This can be achieved by utilizing the rich representation of noise images using Fourier Transformation, which transforms the image from spatial domain to frequency domain. This study proposes using DeepRFT (Deep Learning Residual Fourier Transformation) to build an image denoising model specifically for Fujifilm camera noise. This solution is based on the use of an improved Fourier Transformation for the Residual block of the MIMO-UNet network. The experimental results show that our denoising model competes well with other denoising approaches in terms of evaluation metrics PSNR in both the self-collected dataset and the SIDD benchmark dataset. Khử nhiễu trở thành một trong những chủ đề được quan tâm của lĩnh vực thị giác máy tính nhờ vào khả năng áp dụng vào thực tế, những ảnh thường bị nhiễu do nhiều yếu tố khác nhau như môi trường, điều kiện ánh sáng, chất lượng máy ảnh, các yếu tố khác. Việc khử nhiễu giúp tăng độ chính xác và độ tin cậy của những ảnh này, từ đó cải thiện khả năng phân tích, nhận dạng và phân loại. Một số phương pháp khử nhiễu hình ảnh trước đây đều tập trung rút trích đặc trưng trực tiếp trên ảnh ở miền không gian. Để tái tạo lại hình ảnh sạch từ ảnh nhiễu chụp bởi máy ảnh đòi hỏi cần nhiều thông tin hơn nữa về nhiễu. Tận dụng khả năng biểu diễn phong phú ảnh nhiễu bởi phép biến đổi Fourier từ miền không gian sang miền tần số. Trong nghiên cứu này, chúng tôi đề xuất sử dụng DeepRFT để xây dựng mô hình khử nhiễu, cụ thể ảnh nhiễu được chụp từ máy ảnh Fujifilm. Giải pháp này dựa trên sử dụng phép biến đổi Fourier cải tiến cho khối dư của mạng MIMO-UNet. Kết quả thực nghiệm cho thấy mô hình khử nhiễu sử dụng mạng DeepRFT đạt kết quả cạnh tranh với các cách tiếp cận khử nhiễu khác như Autoencoder, RidNet, DnCNN, MIMO-UNet về chỉ số đánh giá PSNR trên cả tập dữ liệu tự thu thập và tập dữ liệu chuẩn SIDD.

1. Đọc trực tuyến

1. Đọc trực tuyến

Tạo bộ sưu tập với mã QR