Aiming at the problem of low average precision of small target detection in industrial large-size images, an improved Faster R-CNN-Tiny model is proposed. Firstly, the feature pyramid structure is used to improve the second-order detector Faster R-CNN to enhance the feature expression capability and increase the resolution of small target feature mapping to improve the prediction accuracy; secondly, the last piece of the original ResNet structure is changed to deformable convolution to automatically calculate the offset of each point and take features from the most suitable place for convolution, which is used to enhance the small target region Finally, when extracting the features of the region of interest, the contextual information of the content is introduced to improve the accuracy of small target detection. The comparison tests are conducted on the representative satellite remote sensing UCAS-AOD dataset in industry and the quality inspection dataset of surface defects of tiles in Tianchi. The results show that the improved FRC-Tiny model improves the mean average precision of detection by 5.57% and 14.25%, respectively, compared with the original model.