Abstract: The existing logging image interpretation work highly relies on manual experience and expert opinions, which can not quickly understand and give the gist meaning of the image, seriously affecting the depth mining and utilization of the information contained in the logging image. A logging image description method based on the ConvNext network coding-decoding architecture is proposed. The ConvNext network with mixed dilated convolution encoder is adopted to enhance model’s ability to extract low resolution image detail information. Then, the mechanism of additive attention is used to replace the original model for long attention mechanism, cooperated with LSTM(Long Short Term Memory) capable of temporal information memory, enhancing the model’s ability to capture long rely on information, which can generate descriptions of logging image more accurate and more natural. The experimental results show that the evaluation indexes of BLEU-4 (Bilingual Evaluation Understudy-4), METROR(Metric for Evaluation of Translation with Explicit ORdering) and CIDEr (Consensus-based Image Description Evaluation) are improved by 3. 8,4. 0 and 5. 3, respectively, compared with the baseline model method. This research scheme using ConvNext architecture to describe log image information is feasible. 

Key words: image description, ConvNext network, mixed cavity convolution, multi-head attention mechanism