Film identification method based on improved deeplabv3+ for full-film double-ditch corn seedbed

This study aimed to investigate the task demand of intelligent unmanned fertilizer application in seedling stage of corn planted in full-film double-ditch seedbed, a film identification method based on improved DeepLabv3+ identification method for full-film double-ditch corn seedbed was proposed. The differences in performance indicators of the original Deeplabv3+ network taking Xception as the backbone network and the network model that replaced three lightweight backbone networks, MobileNetV2, MobileNetV3 and GhostNet were tested. At the same time, the network models, classical semantic segmentation was introduced to PSPNet and UNet for comparative test. The MIoU of DeepLabv3+ network model that replaced its backbone network increased by 5.01%, and FPS improved by 206% compared with original network, and the model size reduced by 90.3%. The three DeepLabv3+ models after replacing the backbone network were further compressed, and the two-layer expansion convolution with low expansion rate in ASPP was deleted, and the common convolution after feature fusion was replaced by the depthwise separable convolution to obtain a lightweight network model. After testing the improved network model, it was found that the average decline of precision indicators was only 0.17%, FPS raised to 66.5, with an average increase of 25.5%, and the size of the model was compressed to 10.53 MB. Test results showed that, the improved model showed excellent performance, and could provide important technology and method support for the research and development of intelligent topdressing and field management on full-film double-ditch corn seedbed during seedling stage.