A semi-supervised detection method for multi-scale tea geometrid by integrating intra- and inter-class information in natural scene images

Tea geometrid is a pest that is very harmful to tea plants. Images of tea geometrids taken in natural scenes suffer from out-of-focus blurring. The different morphologies and scales of tea geometrids in the images cause intra-class differences and inter-class similarities. In addition, the tea geometrids are too similar to the backgrounds, and accurate data labeling is a major challenge. To address these problems, this study proposed a semi-supervised learning method for multi-scale tea geometrid detection by integrating intra-class and inter-class information in natural scene images. This method solved the blurring problem caused by out-of-focus blurring by constructing a Zoom Enhance Module (ZEM). The semi-supervised learning network, Integrated Information Multi-scale Network (IIMNet), adopted an architecture based on teacher–student networks, and training this semi-supervised network required only a small number of labeled samples and a large number of unlabeled samples. The teacher–student network used ResNet50 as the backbone, and the Scale-Aware Feature Pyramid Network (SAFPN) was introduced to perceive targets and detail information on different scales, effectively solving the problem of scale differences in tea geometrids. The Intra-class and Inter-class Information Integration Network (IIINet) in the teacher–student network integrated the intra-class similarity information of tea geometrids with different morphologies and the inter-class difference information of tea geometrids and similar backgrounds, solving the problems of intra-class differences and inter-class similarities in tea geometrids. Our experimental results showed that compared with the existing supervised detection network, the method proposed in this study improved detection accuracy by 7.63, 9.93, 14.80, and 4.81% using only 5, 10, 20, and 40% labeled samples. Compared to the existing Semi-Supervised Object Detection (SSOD) methods, the proposed method improved the detection accuracy by 1.03, 0.67, 1.02, and 1.63% with 5, 10, 20, and 40% labeled samples, respectively. By integrating intra- and inter-class information, this study achieved effective detection of tea geometrids in natural scene images using a small number of labeled and unlabeled samples.