Moth Images using Deep Learning Architecture

On-trap moth automated identification suffers the problems of varieties of moth pose, life stage and sex, which 5 finally lead to incomplete feature extraction and mis-identification. Due to the intra-species variance, a pose 6 estimation-dependent automated identification method using deep learning architecture for on-trap field moth 7 sample is proposed in this paper. To solve the segmentation task with cluttered background and uneven 8 illumination, a two-level moth automated segmentation method for field moth trap was developed to obtain 9 separate moth sample image from each trap image. Then based on the segmentation results, moth pose estimation 10 was conducted to assign each moth to one of the two types of moth pose, which are top view and side view. 11 Regarding to the refined moth pose types, suitable combination of texture, color, shape and local features were 12 extracted for further moth description. Finally, an improved pyramidal stacked de-noising auto-encoder (IpSDAE) 13 architecture was proposed to build a deep neural network for moth identification. The identification results on 428 14 field-based testing images of nine species with 480 lab-based training images achieved the identification precision 15 value of 98.13%, which demonstrates the effectiveness of the proposed identification method and its potential 16 application in integrated pest management. The results also indicated that the pose estimation process is effective 17 for improving the moth identification results. 18