The accuracy of gastric precancerous disease detection algorithms plays a pivotal role in the adjuvant treatment of gastric cancer patients. However, detection of gastric precancerous diseases are facing two thorny problems: i)the physiological stomach environment is complex, and the lesion site features are not obvious, which lead to low detection accuracy for inconspicuous targets; ii) since the feature scale of the lesion site varies in size, the recognition ability of existing detection models for targets of different sizes is poor due to their insensitivity to multi-scale features. To address these two challenging problems, we propose an efficient and lightweight object detection approach called ELOD for gastric precancerous diseases. At the heart of ELOD are a lightweight bottleneck attention module (LBAM) and an efficient spatial pyramid pooling (ESPP) module. LBAM is designed to strengthen important features while suppressing unimportant features of images, which enables networks to deeply learn obscure targets. ESPP is devised to extract multi-scale features, thereby enhancing the ability of networks to detect multi-scale targets. Moreover, we take advantage of the state-of-the-art object detection architecture YOLOv5s as the underlying platform to implement ELOD and assess the effectiveness of LBAM and ESPP. To evaluate the performance of ELOD, we conduct extensive experiments on a collected gastric precancerous disease dataset and two public datasets (i.e., a gastric polyp dataset and a blood cell dataset). In addition, we quantitatively validate the generality and versatility of LBAM and ESPP. Experimental results indicate that our approach achieves high prediction accuracy,while substantially streamlining network structure and considerably reducing the number of model parameters.