Abstarct: Abstract Fire in the living environment usually leads to life and economic losses; in such a way that compensation for these losses is time-consuming and sometimes impossible. Accordingly, timely detection of fire is an important necessity. Since the presence of smoke in the atmosphere is the indication of fire, so smoke detection is of special importance in fire alarm systems. Unlike the fire flames, smoke can be identified from far away, as it moves in upward direction and is detected faster. Sensors sensitive to smoke and fire have the ability to detect these two undesirable factors, but implementing a huge network of sensors in an open space like a forest is not economical, and on the other hand, since the detection of fire or smoke by the sensor is delayed, as a result, using sensors in sensitive places, where the slightest trace of fire can lead to an explosion in facilities and infrastructures, is not considered a wise thing.There are various methods for detecting fire and smoke, and among these cases, the methods baxsed on image and deep learning exhibit bigger advantages in terms of accuracy and speed in segmentation. In this thesis, we have focused on this topic. To train the deep neural network, 1200 images are prepared and appropriate labels for smoke and fire are applied to each pixel of them. From the total images, 80% were selected for training and 20% for testing. UNet, UNet++ and UNet3+ networks, with innovations in their structure, have been used for smoke and fire segmentation. Programming has been done in pytorch, the version compatible with CUDA11.6. According to the results, the best accuracy can be seen in UNet++ with EfficientNet.B0 and EfficientNet.B7 as the encoder’s backbones, and they are 96.42% and 96.37%, respectively. However, if small-scale of fire for segmentation is considered, according to the IoU of fire, UNet3+ with VGG16 as encoder’s backbone has the best result and this evaluation criterion is 76.35%. Also, if the aim is using of networks with a small number of parameters and good accuracy, UNet3+ with the proposed method and VGG16, UNet++ with EfficientNet.B0, and UNet with the proposed method as encoder’s backbone are good structures for implementation.