With the popular use of remote sensing satellite images, a huge research has been conducted to solve the problem of change detection (CD) from bi-temporal remote sensing images. The improvement of the final result depends considerably on the method used to extract the features of the pairs of images. The development of machine learning (deep learning) has largely contributed to the design of features by learning the hierarchical representation in an unsupervised manner directly from data without human intervention. This approach effectively captures middle and upper-level representations. In this thesis, we propose a new semi-supervised change detection technique for unlabeled bi-temporal remote sensing images based on deep convolutional neural networks (CNN). The goal is to produce a change detection map directly from two images using CNN. First, one of the two images (usually the images before change) is divided into patches of the same size that will be classified by CNN later. Patch classification ensures that features extracted from patches are as discriminating as possible. Then the trained CNN will be used to extract the features of patches of the second image. The assumption is that the same unchanged area is expected to have relatively similar CNN features in both images. Consequently, the difference between the features of the same unchanged area is small compared to that of the modified area. Experiments are carried out on three different real datasets. We used Kappa Coefficient (KC) and F score to compare the performance of our proposed method with three different change detection techniques, Absolute Difference, ResNet and VGG19. The results obtained by our proposed method show the superiority of the method, which outperforms all the other methods and achieves the best F-score and KC on all the datasets.