Insects are the most diverse animal group in nature. Some species are difficult to collect， which makes datasets often highly heterogeneous with long-tailed distributions. This article proposed a convolution recognition network model based on synthetic samples combined model （SSCM） to solve the problem that the uneven distribution of insect datasets leads to the poor recognition performance of recognition models in tail categories with less data. The model contains three modules including image segmentation and shuffle module， backbone network module and data fix branch module. Through the image segmentation and shuffle module， the training image was segmented and shuffled to obtain new training data and added to the training set. ResNet-50 was used as the network backbone to extract features of image. At the same time， the data fix branch module combined the mean square error and cross-entropy to calculate the error between the synthetic samples and the original image to reduce the adverse effect of the synthetic samples on the tail data. A butterfly dataset containing a total of 26 045 images of 300 species was constructed to evaluate the performance of the model proposed. The results showed that the accuracy of SSCM model was 3， 2.14 and 2.71 percentages higher than that of DRC， BBN and RIDE in the butterfly dataset， respectively. In addition， the validity of the SSCM in the public IP102 insect dataset was verified on the public insect dataset IP102. The results showed that the accuracy of SSCM model was 18.94， 3.02 and 3.36 percentages higher than that of DRC， BBN and RIDE， respectively.