Research On Multi-angle Fly Species Face Recognition Based On Deep Learning

With the continuous development of international trade,the substantial increase of entry-exit personnel and goods lead to the invasion of alien species.In particular,invasive fly species will cause harm to agricultural production and human health.Although the customs of various countries continue to improve the construction of quarantine and inspection facilities at ports.However,there are many kinds of fly species.Fly species are similar in size and difficult to recognize,which increases the difficulty of customs inspection.Therefore,it is necessary to study the recognition of fly species.In view of the characteristics of fly species,such as high similarity of overall body shape,large differences of their faces and changeable angles.A multi-angle fly species face recognition method based on deep learning is proposed in thesis.Because the fly face image angle is changeable.Especially when the rotation angle of fly is large,the recognition effect will be affected.Therefore,the fly species faces images are divided into two cases: small-angle fly species face image and large-angle fly species face image.The specific contents include:(1)Aiming at the situation of small-angle fly species face recognition,thesis proposes a fly species face deep convolutional neural network model.The network is used to extract rough contour features and fine local features.In the rough extraction of contour features,Inception-ResNet network and Reduction network with depthwise separable convolutions are used.In the fine extraction of local features,Inception-ResNet network and Reduction network are formed into a module.And multiple modules are spliced together by Filter Concat to become an optimized Inception-ResNet network.(2)Aiming at the situation of large-angle fly species face recognition,thesis proposes a fly species face generative adversarial network model.It synthesizes the large-angle fly face image into the frontal face image.First of all,through the transfer learning,the fly species face keypoint detection network is built.And the fly face is more suitable for the human face keypoint detector.It can detect and locate the fly face keypoints.Secondly,the detected fly face keypoints are processed as pose embeddings.And the large-angle fly face image is guided to synthesize positive face image in the generator of fly species face generative adversarial network.And in the discriminator of fly species face generative adversarial network,the cascaded double proxy discriminator network.Finally,the synthetic frontal fly face image is put into the fly species face deep convolutional neural network for training and testing.Compared with other advanced methods,the experimental results show that the recognition accuracy of this method is 97.14% in small-angle fly species face recognition.In the large-angle fly species face recognition,the recognition accuracy of thesis is 96.74%.And the synthesis effect of thesis is better than other generative adversarial networks.These meet the needs of multi-angle fly species face recognition.