VSTDet: A lightweight small object detection network inspired by the ventral visual pathway

It is difficult for object detection networks to effectively pay attention to the characteristics of small objects, especially when edge computing devices such as drones process their aerial images, and it is even more challenging to accurately detect a large number of different small objects with similar semantic information. Biological vision has a high degree of selectivity and sensitivity to similar objects by the interaction of visual mechanisms in the V1/V2-V4-IT ventral visual pathway, so a new lightweight one-stage model VSTDet is proposed. Specifically, inspired by the antagonistic receptive field and brightness contrast in the V1/V2 area, the orientation selection coding, and the graphic background separation in the V4 area to promote feature integration, the feature extraction network VCBNet and the multi-attribute information integration VDE module were designed. These designs deepen the prominence of small object feature information, to improve the ability of the network to extract small object feature information. In addition, inspired by the analysis of spatial context understanding in the IT area, this paper proposes a lightweight ITT head for information interaction detection head. The proposed VSTDet model has been experimentally evaluated on multiple small-object datasets. On the VisDrone2019 and AI-TODv2 datasets, VSTDet-l attains AP50 scores of 50.3% and 63.3%, respectively, with a parameter count of only 3.51M, reaching the state of the art in lightweight small object detection.