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

Aiming at the problems of low accuracy and slow detection speed of existing residual film segmentation methods, a peanut surface residual film segmentation method based on Haar-Wavelet CBAM Multi-Scale Attention Aggregation UNet (HCM-UNet) is proposed. Based on the UNet model, MobileNetV3 is introduced into the UNet model as a lightweight backbone feature extraction network, and in order to minimize the effect of accuracy degradation after the model is lightweighted, the Haar Wavelet Downsampling (HWD) module is used to reduce the details lost in the downsampling process; secondly, to improve the detection effect of the model for small target residual film, MSAA (Multi-subsampling) is introduced to improve the detection effect of the model for small target residual film. model for small target residual film detection, Multi-Scale Attention Aggregation (MSAA) architecture is introduced to fuse the multi-scale features of residual film; finally, Convolutional Block Attention Module (CBAM) is introduced in the decoder to enhance the model's attention to the edges of the residual film, which further improve the segmentation accuracy of the model. The experimental results show that the HCM-UNet model has an average intersection and merger ratio of 85.72%, an average pixel accuracy of 84.26%, an F1 score of 83.68%, a model parameter count of 43.22 M, and an average segmentation time of 127.17 ms per image on a test set of peanut surface residual film images, all of which outperform Deeplabv3, PSPNet, and UNet, Segformer, Mask2former and HRNet mainstream segmentation models. The model improves the segmentation accuracy of peanut ground residual film on the basis of lightweight, and shows good stability and robustness under different operating periods and light conditions, providing data support for evaluating the contamination of peanut ground residual film.