CA-EBM3D-NET: a convolutional neural network combined framework for denoising with weighted alpha parameter and adaptive filtering

Block-matching and 3D filtering (BM3D) is a popular denoising technique, renowned for its efficiency in enhancing image quality. However, traditional BM3D method involves hard thresholding and Euclidean distance, which shows limitation in yielding effective results due to constant thresholding. Hence in this paper, we provide an improved variant of the BM3D algorithm that incorporates the cosine distance metric for patch similarity measurement and includes exponential moving average (EMA) of noise and non-zero similar patch. EMA acts as a weighting factor to balance the contributions of similar patches. The weighting parameter of EMA is adjusted according to the unique characteristics of the input image. This adaptability improves BM3D’s efficiency to maintain image details while effectively reducing noise. Further, when identifying similar patches for collaborative filtering, we also replace the Euclidean distance metric with the cosine distance, which is renowned for being robust to variations in brightness and contrast, making it particularly ideal for natural images with a variety of textures and contents. Our cosine adaptive enhanced BM3D (CA-EBM3D) delivers improved patch selection and enhanced denoising performance by using cosine distance. CA-EBM3D-NET also incorporates convolutional neural networks (CNNs), which can capture complex image features and improves denoising performance. CNNs have demonstrated remarkable proficiency in learning complex images and performing denoising tasks. Experimental results against benchmark images show that the proposed CA-EBM3D-NET consistently outperforms traditional BM3D and other cutting-edge denoising approaches. The evaluation measures are peak signal-to-noise ratio (PSNR), mean squared error (MSE), structural similarity index (SSIM), and entropy. The results consistently demonstrate improved performance, demonstrating the effectiveness of the proposed method in a variety of image denoising applications. The combined use of EMA, cosine distance, and CNN-based denoising improves noise reduction while preserving fine image details.