关键词: 蛾类, 幼虫, 姿态, 姿态校正, 扇形变换, 不变矩, 欧式距离

Abstract: 【Aim】 In the case of in-situ recognition of pest larva, larvae often appear in bending posture, which makes the extracted feature vectors distorted and affects the results of recognition. To solve this problem, this paper proposes an eigenvector extraction method of posture invariant based on sector-shaped transform. The extracted eigenvectors of pest larvae have the properties of translational, proportional, rotational and posture invariance, which facilitates the automatic identification of fat and short larvae in bending postures. 【Methods】 Firstly, the bending region and non-bending region of the pest larvae are determined by optimal consistent approximation after image thinning. Then, the bending region is straightened through sector-shaped transform, and rotation and translation operations are performed on the non-bending region which is combined with the straightened bending region to form a complete larva. The eight-neighborhood mean method is used to fill up the pixels of gaps in the combined larva, and the automatic correction of bending larva is finished. The invariant Hu’s moment extracted from the image has the property of posture invariance. The minimum distance classifier was used to realize automatic identification of larvae with multiple postures. Finally, experiments were carried out to verify the efficacy of the method by using larvae of Prodenia litura, Heliocoverpa armigera, Spodoptera exigua and Ostrinia nubilalis moths as the target. 【Results】 The identification of 24 kinds of bending larvae of the moths in different postures was carried out on condition that the recognition threshold is 80%. The identification rates based on the classic Hu’s moment and the invariant Hu’s moment were 25% and 100%, respectively. 【Conclusion】 The experimental results demonstrate that the method can effectively identify the bending postures of fat and short larvae.

Key words: Moth, larva, posture, posture correction, sector-shaped transform, invariant moments, euclidean distance