| Summary: | Accurate defect detection of navel oranges is the key to ensuring the quality of navel oranges and extending their storage life. An improved DeeplabV3+ model integrating attention mechanism is proposed to increase the current low recognition accuracy and slow detection speed of defect detection in navel oranges grading and sorting process. The improved lightweight backbone network HECA-MobileV3 is applied in the DeeplabV3+ model to reduce the amount of computational data and improve the image processing speed. In addition, the Convolutional Block Attention Module (CBAM) and Channel Space Parallel Mechanism CSPM are integrated to the DeeplabV3+ model. ASPP structure is redesigned and the low feature extraction network is optimized to enhance the capture of target edge information and improve the segmentation effect of the model. Experimental results show that the proposed model exhibits a better MIoU and MPA with 89.50% and 94.02%, respectively, while reducing parameters by 49.42M and increasing detection speed by 55.6fps, which are 7.27% and 3.51% higher than the basic model. The results are superior than U-Net, SegNet and PSP-Net semantic segmentation networks. As a results, the proposed method provides better real-time performance, which meets the requirements of industrial production for detection accuracy and speed. © (2024), (Science and Information Organization). All Rights Reserved.
|
|---|