Combining YOLOv7-SPD and DeeplabV3+ for Detection of Residual Film Remaining on Farmland

• Published in: IEEE access 2024, Vol.12, p.1051-1063
• Main Authors: Huang, Deqi, Zhang, Yangting
• Format: Article
• Language: English
• Subjects: Agricultural land; Agricultural residual film leftover; Agriculture; Algorithms; attention module; Convolutional neural networks; Cotton; Crops; DeeplabV3; Farming; Feature extraction; Image segmentation; Object detection; Occlusion; Recycling; Soil measurement; Target detection; Technical services; YOLO; YOLOv7
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2023.3347588

Aiming at the problems of low pickup rate of residual film recycling machine, low recognition of background soil and residual film left in farmland under complex farmland environment, and mutual occlusion between classes, we propose a method combining YOLOv7-SPD target detection and Deeplabv3+ image segmentation. YOLOv7-SPD was first introduced to recognize and locate the residual film left in the farmland, and the detected residual film image was passed to the image segmentation algorithm, and the segmented image was processed to calculate the area of the residual film in the farmland. By improving the loss function, fusing the Coordinate Attention (CA) mechanism, and introducing the Space-to-Depth (SPD) module and Atrous Separable Convolution (ASConv) to improve the accuracy of the leftover film detection of farmland residual film. The experimental results show that the average detection precision of the final improved model recall is 87.62% and the average precision is 93.72%, which are 4.93% and 2.53%; The mIOU and F1 of the image segmentation model reached 91.55% and 94.77%, respectively, which is more significant. This research result demonstrates the potential of this algorithm in practical applications related to agricultural residue management and field cleanliness assessment, providing certain technical support to improve the recovery rate of residual film recycling machines and realizing the accuracy and efficiency of detection.