The article describes common techniques for semantic segmentation pictures containing edges of gas turbine engines blades for detecting left and right borders for further using in forming trajectory algorithms with direct metal deposition. For analysis such metrics, as pixel accuracy, mean pixel accuracy, intersection over union, frequency weighed intersection over union are used. Classic method of computer vision with threshold filters, border segmentation neural network method, fully convoluted neural network for semantic segmentation are focused on. The classic method of computer vision process image by several sequential applied filters: translate RBG to HSL, select lightness layer, threshold for this layer, morphological transformation, select top and bottom pixels in blade edge. This method gave 95,18 % pixel accuracy and 65,19 % intersection over union. Several architectures neural network for edge’s border segmentation, such as DexiNed, RCF, PiDiNet were compared. PiDiNet gave the best result: this architecture gave 96,37 % pixel accuracy and 77,57 % intersection over union. The last method in this research was fully convoluted neural network. 75 combinations of encoders and decoders architectures were trained and tested. The represented encoders were ResNet34, ResNet50, ResNet101, VGG11, VGG16, VGG19, InceptionResNetV2, InceptionV4, Efficientnet-b0, Efficientnet-b4, Efficientnet-b7, Xception. The represented decoders architectures were Unet, Unet++, MAnet, Linknet, PSPNet, FPN, DeepLabV3, DeepLabV3+, PAN. Fully convoluted neural network method gave the best result. The most accurate combination was Unet-InceptionResNetV2 model with 99,22 % pixel accuracy and 97,25 % intersection over union metric. The best method for semantic segmentation pictures contain blade edges was chosen.