| With the broadening of deep learning technology,natural scene text recognition based on deep learning is a key field in computer vision research at this stage.It has a wide range of application prospects in many aspects such as medical treatment,life,human-computer interaction,unmanned driving,information entry,and information retrieval.Texts in natural scenes contain complex backgrounds and a lot of noise.It is of great function and application value to study methods that can accurately detect and recognize texts in natural scenes.According to the characteristics of natural scene image text,this paper selects the end-to-end ABCNet model to improve it.Aiming at the problem that the ABCNet model will miss detection and misdetection in the detection and recognition of text in complex backgrounds,which influences the accuracy of the end-to-end model,this paper proposes an attention mechanism-based ATT-ABCNet model,ATT-ABCNet Compared with the classical model,the model improves the accuracy of text recognition in complex backgrounds.The pertinent research contents are as follows:(1)Introduce an attention mechanism to the backbone network of the ABCNet model detection branch,and use the attention mechanism to improve the feature distillation capability of the residual network.The introduction of the attention mechanism enables the network to allocate and process information more accurately.Important features and suppress unimportant parts to improve text missed detection.(2)When the detection branch introduces attention features,the channel attention SENet suitable for convolutional networks and the mixed attention CBAM are selected to compare the gain effects of the two kinds of attention on the feature extraction of the Res Net network,and will be more suitable for combining The attention of the residual network for feature extraction is used to build the ATT-ABCNet model.(3)The attention mechanism is introduced into the decoding network of the recognition branch of the ABCNet model to enhance the inference and recognition ability of the loop layer of the decoding network and improve the phenomenon of text error detection.(4)When the attention mechanism was introduced into the identification branch,the soft attention Luong and Bahdanau attention mechanisms suitable for neural networks were selected to compare the gain effects of the two types of attention on the LSTM network,and would be more suitable for combining with the LSTM network.The attention mechanism for text sequence recognition is used to build the ATT-ABCNet model.For the curved and deformed text,there are still problems that are difficult to accurately identify.An improved ATT-ABCNet model is proposed.The specific study contents are as follows:(1)Add the FAM feature enhancement module to solve the multi-feature fusion of the ATT-ABCNet model.The feature pyramid network structure cannot achieve good fusion of multi-scale features,and the problem of lack of underlying features;and on this basis,the AFF adaptive feature fusion module is added,so that the FPN network can focus on small-sized scene texts and improve fusion.completeness of characteristics.(2)A spatial correction network is added before the text recognition network to correct the skewed and deformed text.The recognition network can directly recognize the corrected text and improve the accuracy of the recognition of curved and deformed text.In order to confirm the advantage of the model,test experiments were carried out on the Total Text and CTW1500 datasets.The ATT-ABCNet model based on the attention mechanism showed a certain degree of improvement in various evaluation indicators of the end-to-end recognition task,and improved the Accuracy of end-to-end text recognition in complex contexts.Compared with the traditional model,the improved ATT-ABCNet model has been significantly improved in various evaluation indicators of end-to-end recognition,and the accuracy of end-to-end recognition of curved and deformed text has been improved. |
PDF Full Text Request