In order to solve the problem of difficult detection due to the existence of dense workpiece with high similarity and disorderly arrangement, an attention mechanism is proposed to optimize RetinaNet’ s dense workpiece detection method. Firstly, the attention mechanism is introduced into the RetinaNet backbone feature extraction network to reduce the influence of interfering objects on the detection effect and improve the feature extraction ability of the neural network. then a new predictive box is constructed using Soft-NMS to improve the overlap localization accuracy. Finally, the dataset is trained by transfer learning method to improve the model training efficiency. The method effectiveness is verified on the produced dense workpiece dataset; Experimental results show that the detection accuracy of the improved method reaches 98. 11%, which is 2. 59% higher in comparison with that before the improvement. The detection speed of a single picture is up to 0. 026s. The proposed method can meet the purpose of accurate detection of workpiece in actual industrial production process, which can reduce the rate of missed and false detection and assure the speed simultaneously.