The garbage instance segmentation in complex stacked scenes is affected by serious occlusion and high density, which makes it more difficult to detect. To solve this problem, a case segmentation method combining DCT-Mask and the idea of double-layer feature fusion network were proposed, which is used for garbage case segmentation in highly stacked scenes. At the network structure level, firstly, the feature data was decoupled in the data pre-processing link, and the impact of stack on occluded object features was reduced through dual branch feature fusion, so as to solve the case segmentation problem under complex stack occlusion. To solve the problem of dense confusion in this scenario, a cascade classifier was incorporated into the candidate box classification regression part, and the loss function of dividing network branches was optimized. The experimental results show that the AP₅₀ , the average accuracy mAP and other indicators of this method have been greatly improved, and have a good segmentation effect and a certain degree of interpretability.