چكيده به لاتين
Depression is one of the main causes of sadness and inability of people to do the simplest daily
tasks in the world and it has increased by about 50% during the last three decades. Depression
and anxiety can lead to serious health crises and even suicide attempts. Although various
treatments are available for depression, such facilities are not available in underdeveloped
countries and even developing countries which are not so advanced. Therefore, to treat
depression in these countries, first of all, it is very necessary to diagnose depression in the
shortest possible time otherwise starting the necessary treatments and dealing with the disease
will be very difficult and long for the patient.
In this research, a comparative study focusing on existing depression diagnosis models is
presented to examine the strengths and weaknesses of each and it was concluded that the
methods based on deep learning have better performance.
To implement the proposed research model, the image data is first pre-processed before
entering the network, which includes things such as changing the size of the input images and
labeling in this first research, a comparative study focusing on existing depression diagnosis
models is presented to examine the strengths and weaknesses of each and it was concluded that
the methods based on deep learning have better performance.
To implement the proposed research model, the image data is pre-processed before entering
the network, which includes things such as resizing the input images and labeling then the
additional information of each image, such as the information related to the filter and the name
of each channel, is removed and only the signal values are kept, and the information of each
data is stored in small periods including 100 seconds of signals(batch size is 100 s) , and the
call is made in periods smaller than that i.e. batch size 80 seconds which In a way, it is
considered a data augmentation process that improves network learning and finally the
implementation of convolution neural network and deep learning techniques is performed on
the pre-processed and trained data set. In this respect, this study simultaneously evaluates the
accuracy of existing models and presents its proposed model. In this research, the dataset
prepared from the Figshare site has been used.
Implementation and comparison of deep learning models in three neural networks VGG16,
residual neural network and Inception neural network and each of them reached the test
accuracy of 0.962, 0.846, 0.575 respectively, which shows that VGG16 neural network is more
accurate than other networks.
The reference article uses the data set of Lanzhou University and the convolution neural
network method achieves an accuracy of 0.943, which is less accurate than this research.
