Deep Orthogonal Transform Feature for Image Denoising

Shin, Yoon Ho; Park, Min Je; Lee, Oh Young; Kim, Jong Ok

doi:10.1109/access.2020.2986827

Cited by 7 publications

(4 citation statements)

References 41 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…From Equation (24), it results that the analysis of 2D signal A is equivalently performed by using a couple of 1D dictionaries (according to Definition ( 12)) or with the help of sfas taken from a 2D dictionary (according to Definition (23)).…”

Section: Jvnt For Discrete-space 2d Signalsmentioning

confidence: 99%

“…Since their inception by Meyer, Mallat and Daubechies (more than 35 years ago), the literature reporting how these wavelets work in conjunction with images has become so vast that any attempt to encompass all the sound (and sometimes amazing) results is very likely doomed to fail. One can only cite a few interesting recent works, such as [23] (with application to image filtering), [24] (in which principal component analysis is developed by means of wavelets, aiming to achieve image denoising), [25] (where image fusion is performed) and [26] (which deals with the efficient implementation of 2D wavelet transforms).…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

John von Neumann’s Space-Frequency Orthogonal Transforms

Stefanoiu,

Culita

2024

Mathematics

View full text Add to dashboard Cite

Among the invertible orthogonal transforms employed to perform the analysis and synthesis of 2D signals (especially images), the ones defined by means of John von Neumann’s cardinal sinus are extremely interesting. Their definitions rely on transforms similar to those employed to process time-varying 1D signals. This article deals with the extension of John von Neumann’s transforms from 1D to 2D. The approach follows the manner in which the 2D Discrete Fourier Transform was obtained and has the great advantage of preserving the orthogonality property as well as the invertibility. As an important consequence, the numerical procedures to compute the direct and inverse John von Neumann’s 2D transforms can be designed to be efficient thanks to 1D corresponding algorithms. After describing the two numerical procedures, this article focuses on the analysis of their performance after running them on some real-life images. One black and white and one colored image were selected to prove the transforms’ effectiveness. The results show that the 2D John von Neumann’s Transforms are good competitors for other orthogonal transforms in terms of compression intrinsic capacity and image recovery.

show abstract

Section: Jvnt For Discrete-space 2d Signalsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

John von Neumann’s Space-Frequency Orthogonal Transforms

Stefanoiu,

Culita

2024

Mathematics

View full text Add to dashboard Cite

show abstract

“…The commonly performed image preprocessing is the image scaling and image quality improvement with certain filters. Meanwhile, for image feature extraction, the Fourier transform [1], wavelet, and principal component analysis (PCA) [2], [3] are often used. Image scaling is usually carried out by resizing the original image so that the size (dimensions) of the image for training and testing becomes identical.…”

Section: Introductionmentioning

confidence: 99%

Citra Tekstur Terbaik Untuk Gaussian Naïve Bayes Dengan Interpolasi Nearest Neighbor

Irwan Budi Santoso,

Shoffin Nahwa Utama,

Supriyono

2024

Jurnal Nasional Teknik Elektro dan Teknologi Informasi

View full text Add to dashboard Cite

Salah satu faktor yang berpengaruh terhadap kinerja Gaussian naive Bayes classifier (GNBC) dalam klasifikasi citra tekstur adalah ukuran (dimensi) citra. Ukuran citra adalah salah satu kriteria citra tekstur terbaik di samping nilai pikselnya. Pada penelitian ini, diusulkan metode untuk mendapatkan ukuran citra tekstur terbaik terhadap GNBC dengan optimalisasi interpolasi nearest neighbor (NN). Ukuran citra tekstur terbaik dengan nilai piksel hasil interpolasi tersebut membuat GNBC mampu membedakan citra tekstur pada setiap kelasnya dengan kinerja paling tinggi. Langkah pertama metode usulan tersebut adalah menentukan ukuran citra tekstur untuk pelatihan melalui kombinasi ukuran baris dan kolom pada proses optimalisasi. Pengubahan ukuran (resizing) semua citra tekstur asal dengan interpolasi NN adalah langkah penting berikutnya untuk mendapatkan citra tekstur baru. Langkah selanjutnya adalah membangun GNBC berdasarkan citra baru hasil interpolasi dan menentukan akurasi klasifikasinya. Langkah terakhir yaitu memilih ukuran citra tekstur terbaik berdasarkan nilai akurasi klasifikasi terbesar sebagai kriteria pertama dan ukuran citra sebagai kriteria kedua. Evaluasi terhadap metode yang diusulkan tersebut dilakukan menggunakan data citra tekstur dari dataset publik CVonline dengan melibatkan beberapa skenario uji coba dan metode interpolasi. Hasil uji coba menunjukkan bahwa pada skenario yang melibatkan lima kelas citra tekstur, GNBC dengan interpolasi NN memberikan nilai akurasi klasifikasi terkecil 89% dan terbesar 100% pada ukuran citra terbaik, masing-masing 14 × 32 dan 47 × 42. Pada skenario yang melibatkan jumlah kelas kecil hingga besar, GNBC dengan interpolasi NN memberikan akurasi klasifikasi 81,6% - 95%. Dari hasil tersebut, GNBC dengan optimalisasi NN memberikan hasil lebih baik daripada metode interpolasi nonadaptif lainnya (bilinear, bicubic, dan lanczos) dan principal component analysis (PCA).

show abstract

“…Researches based on artificial neural networks have made considerable progress, a batch normalization denoising network was designed by Tian [1], the problems of covariant displacement and small batches are solved, but the method is only suitable for high-light images, it is not suitable for low-light and blurred images. The separable noise orthogonal transform feature was introduced into the neural de-noising framework by Shin [2]. Compared with single-input networks, multi-input networks have good denoising characteristics.…”

Section: Introductionmentioning

confidence: 99%