Abstract: To balance the speed and precision of stored grain pest image detection and obtain the optimal target detection model, an improved YOLOX network model was proposed and implemented for stored grain pest detection. By introducing a new GSConv to replace the standard convolution, and combining with the introduced Res-CBAM attention mechanism, the feature extraction capability of the model was improved and the computational load was reduced. SiLU activation function and batch normalization (BN) was introduced to improve model training efficiency. Bidirectional feature pyramid network (BiFPN) was introduced to improve the feature fusion effect. The loss function was improved to enhance the stability of target box regression, and non-maximum suppression was used to solve the problem of redundant boxes. The experimental results showed that the computational load of the improved YOLOX model was reduced by 31%. The detection speed was increased by 18%, up to 60 FPS, and the detection accuracy was improved by 6.14 percentage paints, up to 97.05%. The improved YOLOX model combined the advantages of one-stage algorithm and two-stage algorithm, and achieved obvious results in the intelligent identification of stored grain pests.