Abstarct: Considering the fact that digital images are all over the internet and shared in the media, image classification has been recognized as one of the basic needs of digital world. Automatic content descxription is one of the major issues of image classification which has resulted in establishing a higher relationship between machine vision and image processing. Automatic content descxription has many applications in the web-baxsed systems and search engines such as filtering unconventional images, identification of subjective images and detection of human behaviors. Automatic content descxription problem can be studied from two points of view. In the first viewpoint, dataset includes only small images with specific applications such as recognition of unconventional images. Recognition of unconventional images is a descxription problem with two classes and minor complexities. In this thesis, a new structure for deep neural networks was proposed for detection of unconventional images. This new structure works principally baxsed on the extraction of high-level features from human body in unconventional images. The results obtained from applying this new approach on two datasets indicated that the performance can be improved by 4% compared to other methods presented in the recent years. In the second viewpoint, the dataset includes images with more classes of different scenes. In this thesis, an approach was proposed for an automatic content descxription baxsed on the extraction of high-level semantic regions which was applied to obtain the label of objects. This method consists of several steps. In this first step, regions and label of objects are identified. To do this, a new structure baxsed on the deep neural network was proposed which employs the R-FCN method. A loss function which has not been used before was also part of this method. Considering the approach proposed in this step, the time required for training and testing was reduced. A major step for the automatic descxription of images is the extraction of latent topic from the image scenery. Thus, in the next stage, latent topics baxsed on the high level of extracted semantic regions were obtained by the help of bag of word and improved k-means methods. Considering the bag of visual word obtained from the region proposal, their locations and label of objectives were extracted using the supervised document neural autoregressive distribution estimator for the latent topics in the visual images. At the end, with the combination of these high-level features, classification of the image scenery was done. This proposed approach was applied to Scene-15, UIUC sports, MIT-67 datasets with 15, 8 and 67 scenes, respectively. The results obtained indicated that this method can improve the accuracy and reduce the time of the testing stage in these datasets.