This study aims to develop a method based on vision transformers to classify various objects. Currently regarded as the most sophisticated models in natural language processing, these networks do not rely on convolution layers like traditional convolutional neural networks (CNNs). Alternatively, they use multiheaded attention processes as the basic building block to infer long-range contextual associations between image pixels. The pictures that are being studied are first segmented into patches, and then they are transformed into sequences via the procedures of flattening and embedding. This is the first step in the analysis process. To preserve information on the position, these patches add an embedding location. The final representation is produced by passing the sequence formed via several different attention levels, including many heads. During the processing that takes place during the classification stage, a softmax classification layer is the one that may receive the initial token sequence. To increase classification performance, we study several alternative data augmentation processes. These procedures are designed to provide more data for training purposes. Furthermore, through experiments, we show that it is feasible to minimize the network size by deleting half of the layers while preserving the accuracy of competing classifications. When the model's capabilities are compared to those of the most sophisticated methodologies that are now accessible, the experimental results collected from flower image datasets demonstrate the model's promising potential. More specifically, vision transformer obtains an average classification accuracy of 88.49%. In contrast to the original form, which produces 90%, the compressed version, made by removing half of the multiheaded attention layers, delivers less accuracy.