Application Of Deep Learning In Lithofacies Recognition

Rock section images are necessary data sets for oil and gas exploration research.The lithofacies recognition studied in this thesis is to automatically annotate such images,and convert complex rock images into simple text.It can not only help researchers quickly grasp useful information in images,improve research efficiency,but also assist beginners in understanding image content.The thesis uses an Encoder-Decoder model.The encoder in the model is created using the GoogLeNet network as the prototype,and the decoder in the model is created using the LSTM as the prototype.When a rock image is input to the model,the encoder extracts image features and outputs feature vectors to represent the image.At the same time,the feature vector is also the input of the decoder.The decoder outputs a corresponding word for each feature vector until it outputs the end-of-sentence.The thesis not only successfully applied the automatic annotation of images in deep learning to lithofacies recognition,but also proved that the model is better than the LabelImg image annotation tool.The thesis compares the experimental results of three different pore rock images and finds that the intergranular pore rock image has a better recognition effect in the model due to its simple color characteristics and clear shape characteristics.