With the continuous progress of image retrieval technology, the speed of searching for the required image from a large amount of image data has become an important issue. Convolutional neural networks (CNNs) have been used in image retrieval. However, many image retrieval systems based on CNNs have poor ability to express image features. Content-based Image Retrieval (CBIR) is a method of finding desired images from image databases. However, CBIR suffers from lower accuracy in retrieving images from large-scale image databases. In this paper, the proposed system is an improvement of the convolutional neural network for greater accuracy and a machine learning tool that can be used for automatic image retrieval. It includes two phases; the first phase (offline processing) consist of two stages; stage1 for CNN model classification while stage 2 for extracts high-level features directly from CNN by a flattening layer, which will be stored into a vector. In the second phase (online processing), the retrieval depends on query by image (QBI) from the system, which relies on the online CNN model stage to extract the features of the transmitted image. Afterward, an evaluation is conducted between the extracted features and the features that were previously stored by employing the Hamming distance to return all similar images. Last, it retrieves all the images and sends them to the system. Classification for images was achieved with 97.94% deep learning results, while for retrieved images, the deep learning was 98.94%. For this paper, work done on COREL image dataset. The images in the dataset used for training are more difficult than image classification due to the need for more computational resources. In the experimental part, training images using CNN achieved high accuracy, proving that the model has high accuracy in image retrieval.