Görüntüleme
12
İndirme
1
Bu çalışmada derin öğrenme metodu olan evrişimli sinir ağları (Convolutional Neural Networks - CNN) ve transfer öğrenme metodu kullanılarak tütün mamulleri, alkollü içecek ve silah gibi istenmeyen nesnelerin tespitini ve sınıflamasını yapan bir model geliştirilmiştir. Bu model Tensorflow JS’e dönüştürülerek, internet tarayıcısı için bir eklenti olarak geliştirilmiştir. Bu eklenti ile izlenen videolardan anlık görüntüler alınarak eğitilen model üzerinde sınıflandırmalar yapılmıştır. Videolar üzerinde yapılan sınıflamalarda gerekli görülen sonuçlar, Google tarafından sağlanan bir bulut hizmeti olan Firebase’in RealTimeDatabase’ e kaydedilmiştir. Kaydedilen veri tabanı kullanılarak daha önceden kötü içerik tespiti yapılmış videoların engellenmesi sağlanmıştır. Tarayıcıdan yapılan tespitlerin son 25 tanesi bilgilendirme amaçlı kullanıcı tarafından görüntülenebilmektedir. Bu çalışmada, izlenen videolardan anlık görüntüler alınarak model ile sınıflama yapılmıştır. Gerekli hallerde videonun bilgisi veri tabanına eklenebilmektedir ve veri tabanına kayıtlı videoların görüntülenmesi filtre edilebilmektedir. Geliştirilen sistem hem fiziksel cihazlar hem de emülatör aracılığıyla test edilmiştir. CNN ile geliştirilen derin öğrenme modelinin ağ yapısı oluşturulmasında iki yol benimsenmiştir. Birincisinde, tüm ağ modeli tarafımızca oluşturduğumuz modeldir. Bu modelde, parametre sayısı 7.752.707 adettir ve %86,75 eğitim ve %88,02 test doğruluğu elde edilmiştir. İkinci olarak, transfer öğrenme metodu kullanılarak, başarısı literatürde kanıtlanmış modellerden olan MobileNetV2 tercih edilmiştir. Çıkış katmanları düzenlenmiş bu modelde eğitilebilir 593.155 adet ve toplamda 2.852.675 adet parametre ile %65,34 eğitim ve %50,35 test doğruluğu elde edilmiştir. Yapılan çalışma sonucunda, video içeriklerini filtrelemek için CNN modelinin daha verimli olacağı bulgusuna ulaşılmıştır.
This study developed a model that uses the evolutionary neural networks (CNN), the deep learning method, and the transfer learning method to identify and classify unwanted objects such as tobacco products, alcoholic beverages and weapons. This model was converted to Tensorflow JS and developed as a plugin for the Internet browser. This plugin has been classified on the trained model by taking instant images from the videos watched. The results required in the classifications on videos are recorded in the RealTimeDatabase of Firebase, a cloud service provided by Google. The recorded database is used to prevent videos that have previously been detected with bad content. The last 25 of the browsers’ detections can be viewed by the informative user. In this study, the model was classified by taking instant images from the videos watched. If necessary, the information of the video can be added to the database and the view of the videos recorded in the database can be filtered. The developed system has been tested through both physical devices and emulators. Two ways have been adopted in the networking structure of the deep learning model developed by CNN. First, the entire network model is the model we create by our side. In this model, the number of parameters is 7,752,707 and 86,75% training and 88,02% test accuracy have been obtained. Secondly, using the transfer learning method, the success was preferred by MobileNetV2, which is one of the models proven in literature. The output layers arranged in this model can be trained 593,155 and a total of 2,852,675 parameters with 65,34% training and 50,35% test accuracy were obtained. The study found that CNN model would be more efficient to filter video content.
In this study, a model that detects and classifies unwanted objects such as tobacco products, alcoholic beverages and weapons was developed using Convolutional Neural Networks (CNN), which is a deep learning methods and transfer learning. This model was converted to Tensorflow JS and developed as an add-on for an Internet Browser. With this add-on, snapshots were taken from the watched videos and classifications were made using the trained model. The results that are deemed necessary in the classification of the videos were recorded to Firebase's RealTimeDatabase, a cloud service provided by Google. By using the recorded database, videos that had previously been detected with harmful content were blocked. The last 25 of the definitions made from the browser can be viewed by the user for informational purposes. In this study, snapshots from watched videos were taken and classified with the model. If necessary, information about the video can be added to the database, and videos saved in the database can be filtered. The developed system has been tested through both physical devices and emulator. Two ways have been adopted in creating the network structure of the model developed with CNN. First, the entire network model is the one that is created by us. In this model, the number of parameters was 7,752,707 and the accuracy of 84.84% training and 79.77% testing was achieved. Second, MobileNetV2, which is one of the models whose success has been proven in the literature, was preferred using the transfer learning method. With 593,155 trainable parameters and 2,852,675 parameters in total, 65.34% training and 50.35% test accuracy was achieved in this model with output layers arranged. As a result of the study, it was found that the CNN model would be more efficient in filtering video content.
Alan : Fen Bilimleri ve Matematik; Mühendislik
Dergi Türü : Uluslararası
Makale Detay 
Benzer Makaleler
PDF Görüntüle
Dergi Bilgisi
Eseri Dinleyin
Alıntı Yap
Bu Sayfayı Yazdırın
Paylaş
Mendeley
