Abstract:

Deep learning, which hosts many subjects such as object recognition, image processing, prediction, has begun to become a need for mankind over and over the day. In this study, the process of object recognition is carried out using deep learning techniques. Faster R-CNN (Faster Region Based Convolutional Networks) is a network of 502 images. Some of the data set created was used by dividing the other part for testing and the other part for training. By conducting different experiments, it was recommended how to increase the accuracy prediction rate after observing how the glass is recognized or not recognized. The interface is designed with the web glass options that we can get photos, videos and images instantly while making object recognition. The Tkinter Library, which is the Python Library, was used during the design. The image to be detected, after the processes, is recorded a photograph that is written in how much percentage it is estimated correctly, if the photo is made into a frame above the object. The image video that will be detected will be written how much percentage the object on the video is predicted in the frame while the video will be played, and the web glass will immediately continue to display the object in the image by taking it into the frame. In this study, using the Object Detection API, we tried to train the model with different epoch values and find the epoch value that is predicted in the most accurate proportion. The 18 separate experiments were carried out through deep learning on the data set created and using Faster R-CNN. In order to find the most successful predictive rate in the training process, experiments were conducted with different epoch numbers. The result of the total 100,000 Epoch training is 0.97835 and the loss rate is 0.02165.

Keywords:

Abstract:

Deep learning, which is home to many subjects such as object recognition, image processing, forecasting, has become a human need. In this study, object recognition is performed using deep learning techniques. Faster R-CNN (Faster Region Based Convolutional Networks) Network has developed and 502 images in the study “Cup” data set has created. Some of this data set has used for testing and the other for training. After observing the ways in which the glass can be recognized and not recognized by conducting different experiments, suggestions have made on how to increase the accuracy prediction rate. The Python library Tkinter library has used when designing the interface. The image that is to be detected is taken into a frame after the operations done, and a photograph is recorded with the correct estimate of the percentage of the image. If the video is the image that will be detected, the video will be displayed on the screen, while the video will be played, and the web cam will instantly display the object in the image in the frame. In this study, which has performed using Object Detection API, we tried to find the epoch value that trains the model with different epoch values and makes the most accurate predictions. It has conducted using deep learning and Faster R-CNN on the data set generated over 18 separate experiments performed. In order to find the most successful prediction rate in the training process, experiments has conducted with different epoch numbers. A total of 100.000 Epoch's has achieved as a result of the education and the result of success is 0.97835 and loss ratio is 0.02165.