FBoT-Net: Focal bottleneck transformer network for small green apple detection

Abstract

Due to the small sizes of green apples and the similar color of the fruit skin and background in real complex orchard environments, detecting small green apples remains a great challenge. In this paper, we present a focal bottleneck transformer network (FBoT-Net) to incorporate high-level semantic information with strong representation ability and global and local feature information through the focal bottleneck transformer module. We replace the multi-head self-attention with a focal transformer layer in the final three bottleneck architecture steps of ResNet. Specifically, focal multi-head self-attention based on the window is introduced to capture local fine-grained information around the window and global coarse-grained features via window-wise attention. The experimental results show that APS of small and large scale can reach 47.3% and 34.2% on the SmallApple and Pascal VOC. Overall, FBoT-Net is slightly better than BoTNet and other state-of-the-art methods when run on SmallApple. Our method also has a strong generalization ability on the Pascal VOC publicly.