Research On Butterfly Images Segmentation And Recognition Based On Deep Learning

Biodiversity in nature is very important for maintaining ecological balance and promoting the harmonious development of human and nature.Insects are the largest group of creatures on the planet,and butterflies as one of them are closely related to human production and life.Because butterflies in nature often have complex color and texture characteristics,manual recognition has a high error rate and low efficiency.Therefore,the automatic butterfly image recognition method has very important application value in many scenarios,such as butterfly population protection,crop disease and pest control,border quarantine and so on.According to different application requirements,the research on butterfly images can be divided into two tasks: segmentation task and classification task.In the butterfly image segmentation task,the difficulty lies in how to segment the butterfly target with high accuracy from the complex environmental background.In the butterfly image classification task,the difficulty lies in how to effectively distinguishing inter-class difference and intra-class difference.For this reason,this paper uses deep learning techniques to study butterfly image segmentation task and classification task.The main work of this paper includes:(1)A butterfly segmentation method based on Mask RCNN is proposed.Due to the lack of available butterfly segmentation annotations,a small butterfly segmentation dataset was first constructed from the existing butterfly dataset,and the segmentation labels of the dataset were annotated manually.Then transfer training is performed on the classic Mask RCNN instance segmentation algorithm using the manually annotated dataset.The entire butterfly dataset is segmented by the trained Mask RCNN to obtain the final segmented butterfly image dataset.The dataset contains 4353 segmentated butterfly segmentation images,which is essential for butterfly image analysis in the future.(2)Aiming at the problems of over-segmentation and under-segmentation in the segmentation results of Mask RCNN algorithm,a refined segmentation method combining multiple image segmentation algorithms is proposed.In this method,the mask edge segmented by Mask RCNN algorithm is first extended by SLIC superpixel method,and the segmented ternary image is obtained by morphological processing.Finally,the original image is segmented twice using the improved Grab Cut algorithm combined with the ternary image.Experiments show that the butterfly image testset has a segmentation accuracy improvement of about 2% compared to Mask RCNN.(3)Aiming at the problems of large intra-class difference and small inter-class difference in butterfly image classification,a hierarchical attention bilinear pooling method is proposed,and the data augmentation method is used to improve the accuracy of butterfly image recognition.By replacing the three projection matrices in the HBP method with selective kernel convolution which have an attention mechanism,local features in the butterfly image are located,and the local features are further strengthened by a quadratic product method.Finally,the Mix up data augmentation method was used to improve the generalization ability of the model.The recognition accuracy on the butterfly dataset was 96.01%.