Steganalysis researches for the techniques used to reveal the embedded messages that is hidden in a digital medium -in most cases in images. The research and development activities in Image Steganalysis has gained more traction in recent years. Although machine learning techniques have been used for many years Deep Learning is a new paradigm for the Image Steganalysis domain. The success of the deep learning process is based on the training of the model for a sufficient amount of and with a high quality, diverse and large-scale data set. When the training process lacks dataset in terms of quality, variety and quantity, Transfer Learning emerges as an effective solution from Deep Learning methods. In Transfer Learning, an untrained model benefits from a previously trained model and its dataset. Base function is defined to transfer the parameters from the trained model to the untrained model. Hence, it would increase the success of deep learning model on Image Steganalysis. In this work, we compare the results of two series of models that are trained both with and without Transfer Learning method. The optimization method of the model training process is selected as experimental AdamW optimization method. Comparison of training, testing, evaluating and F1 scoring are based on the models trained with different steganography payload values which starts from easy to hard to detect. We investigated for the best possible ways of increasing the success rate and decreasing the error rate on detecting stego images and cover images separately with this study. Results showed that transfer learning applied model is more successful on detecting stego images on every different rated payload dataset compared to the normal trained model.