Abstract: In order to further improve the effect of image recognition, meet the current complex image recognition task, using Informer model to design a new type of image recognition model, the model by introducing probspares self-attention mechanism, effectively reduce the time complexity and memory usage, in addition, with generative decoder structure to obtain the output of image sequence, effectively avoid the cumulative error propagation of model reasoning stage. In order to verify the effect of the model, three different types of fine-grained image recognition datasets were used: Road Sign Detection data set, Stanford Cars data set, and Chinese traffic sign detection benchmark data set(CCTSDB). The Road Sign Detection data set was used to initially verify the effect of the image recognition model, while the Stanford Cars data set and the CCTSDB data set were used to verify the recognition effect of the designed image recognition model in vehicle recognition and complex scenarios. Identification results on the data set shows that in the background information is very complex, attention layer cannot quickly locate to the key feature area, after several rounds of feature extraction, the encoder got a lot of irrelevant features, decoder in reasoning generation stage cannot get enough quality information, the model training is insufficient, can reduce the accuracy of image recognition, but the image recognition model still can to overcome the interference of the noise information, can achieve high image recognition accuracy.