Many species of pests have posed a serious threat to the yield and quality of rice and ,thereby, caused huge economic losses every year in the world. Accurate, real-time forecast of rice pests is highly demanding to take controlling measures in time. Since the commonly-used light-traps in China can automatically trap and kill the trapped insects, the subsequent procedure is still labor-intensive and time-consuming with low efficiency and less objectivity, causing a delay in the artificial identification and counting of rice pests. The reason is that the trapped insects in a day are often collected in one insect bag, and then the insect bags were taken back to the pest identification. Although deep learning has widely been used in pest identification, high accuracy is still lacking in the detection of rice planthopper, due to the small area proportion of planthoppers in the whole image. Moreover, great similarities of various rice planthoppers have made it much more difficult to detect light-trapped rice planthoppers in the complicated field environment. In this study, an improved automatic detection was proposed using CornerNet for higher precision and recall rate of rice planthoppers on light-trap insect images. 12 megapixel images were also captured for object detection. Data enhancement was first employed to expand the training set, further improving the generalization ability of the model. Next, an overlapping sliding window was applied to select fixed size regions by scanning the overlapped in the sliding direction from the edge to the center of one image. Subsequently, these size-fixed areas were fed into a cross-layer connected hourglass network with a high symmetry for the extraction of features. The area proportion of planthoppers on one image increased in sliding window selection for the improved detection rate. A large number of candidate boxes were obtained after the corner matching and coordinate correction in the prediction module. Four approaches were also developed for the detection box suppression to remove redundant detection boxes. Finally, the detection boxes with the highest score were selected as the detection data in the overlapping candidate boxes. The same image sets from the light traps were used to verify three detection models and their improved ones. The results demonstrated a higher accuracy detection of rice planthoppers was achieved under the overlapping sliding window and the detection box suppression in the different improved models. The improved CornerNet model presented an excellent performance with an average precision of 95.50% and a recall rate of 95.53% for the two species of rice planthoppers. This work can be applied in smart light traps to enhance the forecasting accuracy of rice planthoppers.