This paper introduces an online algorithm for learning dictionaries adapted to sparse coding tasks. The algorithm is based on stochastic approximations and is designed to scale efficiently to large datasets with millions of training samples. It is proven to converge to a stationary point of the cost function, and experiments show it outperforms classical batch algorithms in both speed and dictionary quality for both small and large datasets. The algorithm processes one element of the training set at a time, making it suitable for dynamic data and large-scale image processing tasks. It does not require explicit learning rate tuning and exploits the specific structure of sparse coding to minimize a sequentially quadratic local approximation of the expected cost. The algorithm is shown to be effective in learning dictionaries for natural images and can be used for tasks such as image inpainting. The method is compared with stochastic gradient descent and is found to be faster and more efficient, especially for large datasets. The algorithm is also demonstrated to be applicable to challenging image restoration tasks, including denoising, deblurring, and inpainting. The paper concludes that the proposed method is a promising approach for sparse coding and can be extended to various applications in image and video processing.This paper introduces an online algorithm for learning dictionaries adapted to sparse coding tasks. The algorithm is based on stochastic approximations and is designed to scale efficiently to large datasets with millions of training samples. It is proven to converge to a stationary point of the cost function, and experiments show it outperforms classical batch algorithms in both speed and dictionary quality for both small and large datasets. The algorithm processes one element of the training set at a time, making it suitable for dynamic data and large-scale image processing tasks. It does not require explicit learning rate tuning and exploits the specific structure of sparse coding to minimize a sequentially quadratic local approximation of the expected cost. The algorithm is shown to be effective in learning dictionaries for natural images and can be used for tasks such as image inpainting. The method is compared with stochastic gradient descent and is found to be faster and more efficient, especially for large datasets. The algorithm is also demonstrated to be applicable to challenging image restoration tasks, including denoising, deblurring, and inpainting. The paper concludes that the proposed method is a promising approach for sparse coding and can be extended to various applications in image and video processing.