Raising pattern density on the integrated circuit layout enables reduction of the chip size and cost, while fabrication becomes more difficult. Both number and type of hotspot are increasing during the scaling down. Hotspot detection usually relies on photolithography simulation, that consumes large number of massive computational resource and often takes a long time. To address these limitations, we propose a lightweight hotspot detection model, the Lithography Hotspot Detection (LHD) model, based on deep learning technology. In order to enhance the training of the model, we employed Mentor Calibre's SONR tool to sample the hotspot and non-hotspot graphics and prepare the input dataset. Experimental results show that the hotspot detection accuracy of the LHD model reaches 95.25%, which is 5.07% higher than that of the traditional classification model Convnext.