Detection of peanut surface residual film in complex environments based on HCM-UNet

Residual plastic film on the peanut field surface has limited the subsequent large-scale production in farmland. Existing segmentation of residual plastic film is also confined to low accuracy and slow detection speed. An accurate and rapid detection is often required to assess the residual plastic film pollution. In this study, a segmentation approach was proposed for the residual plastic film on the peanut field surface using HCM-UNet. Both performance and efficiency were optimized to detect the residual plastic film on the peanut field surface under complex environments. Furthermore, MobileNetV3 was integrated into the UNet model as a lightweight backbone feature extraction network in order to enhance the overall performance of the HCM-UNet model. Meanwhile, a Haar wavelet downsampling (HWD) module was adopted to mitigate the accuracy degradation caused by the lightweight model. The loss of details was then reduced during downsampling. Secondly, the multi-scale attention aggregation (MSAA) architecture was introduced to fuse the multi-scale features, in order to improve the detection performance for the small-target residual plastic film. Finally, the convolutional block attention module (CBAM) was incorporated into the decoder to strengthen the focus on the edges of residual plastic film, thereby further improving the segmentation accuracy. A total of 1 340 images were collected from the peanut field surface. The dataset of residual plastic film was then constructed to expand the image quantity to 4 880 after data augmentation. Subsequently, the dataset was divided into 3 416 training samples, 976 validation samples, and 488 test samples in a ratio of 7:2:1. A series of ablation experiments was conducted to verify the effectiveness of the HCM-UNet model. The experimental results demonstrated that the detection accuracy was significantly improved with the MSAA feature fusion module and the CBAM attention mechanism. Compared with the original UNet model, the HCM-UNet model increased the mean intersection over union (mIoU), mean pixel accuracy (mPA), and F1 score by 3.27, 3.80, and 4.79 percentage points, respectively. In addition, the mIoU of HCM-UNet was 1.30, 0.87, 2.15, and 1.90 percentage points higher than that of UNet, respectively, in the four scenarios of "Postharvest-sunny", "After plowing - sunny", "Postharvest - cloudy", and "After plowing - cloudy". The mIoU of HCM-UNet was 1.22 and 2.93 percentage points higher than that of UNet, respectively, under sunny and cloudy lighting conditions. Furthermore, the mIoU of HCM-UNet was 1.38 and 0.86 percentage points higher than that of UNet, respectively, in the two operation periods of post-harvest and after plowing. The mIoU of HCM-UNet increased by 4.38, 9.48, 3.27, 3.84, 1.65, and 5.85 percentage points, respectively, compared with the mainstream lightweight models, such as Deeplabv3, PSPNet, UNet, Segformer, Mask2former, and HRNet. Meanwhile, the inference time was accelerated by 66.62, 99.66, 118.47, 36.07, 1.15, and 149.34 ms, respectively. Overall, the HCM-UNet model achieved an mIoU of 85.72%, an mPA of 84.26%, and an F1 score of 83.68%, with a model parameter count of 43.22 M and an inference time of 127.17 ms, indicating excellent performance in both accuracy and speed. Visual analysis confirmed that the improved model also exhibited high stability and robustness under different scenarios. There was an intelligent and high-precision detection of the residual plastic film on the peanut field surface. In conclusion, the HCM-UNet model can provide a promising solution to the accurate and rapid detection of the residual plastic film on the peanut field surface under complex environments. The finding can accurately capture the distribution, coverage, and fragmentation of the residual plastic film at different stages. Agricultural operations can be adjusted to improve the recycling machinery of the residual plastic film. Intelligent technical support can also offer to accurately assess and efficiently control the residual plastic film pollution.