Double-stage insect recognition model based on synergistic localization of key parts.

Introduction: Currently, pests and diseases have caused great losses in global agricultural production systems. However, the Solenopsis invicta has a high degree of interspecific similarity and often pose a great challenge in the accurate identification of insects. Methods: In order to distinguish closely related species of small insects, this study proposes a Double-stage Insect Recognition Model (DIRM) based on the synergistic localization of key anatomical sites. The model uses the object detection network to pinpoint and crop the key part regions of the target object, and then feed these regions into the classification network and fuse the centroid information, the Global Context Enhancer (GCE) module, and the adaptive synergistic mechanism to linearly synthesis the features of each part in order to efficiently and accurately identify morphologically similar ant species. Results: Experimental results demonstrate that this approach substantially improves the classification accuracy of morphologically similar ant species, achieving over 95% accuracy on the constructed Ant7C dataset. Compared with conventional single-stage deep learning classifiers, the proposed DIRM method exhibits greater robustness in handling small targets, complex backgrounds, and closely related species. Discussion: Beyond its applicability to ant identification, the model holds potential for broader applications in the recognition of insects and ecological monitoring.