Caries is a common and frequently occurring disease, and accurate detection of caries regions is an important prerequisite for caries removal treatment. To solve the problems that the existing caries detection networks repeatedly extract the unrelated regions, which leads to a large amount of computational costs, the caries regions are easy to be confused with similar texture features, and caries regions usually only account for a small proportion of pixels in a complete upper or lower jaw image, this paper proposes a small target caries detection network based on bi-directional re-convolution and adaptation. Firstly, the BRC (bi-directional re-convolution) module is designed to reduce the influence of redundant features on the detection performance of the network. Secondly, the adaptive weight downsampling module is used to replace part of the standard convolution to guide the network to focus on the caries regions and reduce the attention on similar interference features. Finally, the Slide wise loss function is designed to optimize the localization of small objects through smooth regression and dynamic constraints. The experimental results demonstrate that the network proposed in this paper has good performance in terms of precision and can provide technical support for the deployment of dental laser tips in hospitals to assist in caries removal.