关键词: 双模态MobileViTv2, 饲料剩余量, 非接触式估算, RGB图像, 深度图像

Abstract: Aiming at the problems of traditional feed residue detection methods relying on contact sensors, high cost, and the need to modify feeding troughs, a lightweight convolutional fusion regression model （dual-modal MobileViTv2 + CMFIM + SE） based on dual-modal MobileViTv2 was proposed to achieve non-contact and high-precision automatic estimation of feed residue. Taking RGB images and depth images as input, the model extracted multi-scale features respectively through the dual-modal MobileViTv2 and introduced a cross-modal multi-scale feature interaction module (CMFIM) at four levels to achieve spatial-channel dual interaction between RGB and depth features. An SE module was employed to adaptively calibrate channel weights and enhance high-level semantic representation capability. The prediction results were output through a multilayer perceptron regression head. On the self-built dataset, the mean absolute error (MAE) and root mean square error (RMSE) of the dual-modal MobileViTv2 + CMFIM + SE model were 98.24 g and 140.21 g, respectively, which represented reductions of 21.65% and 16.73% compared to the dual-modal MobileViTv2 model without the CMFIM and SE modules, and the parameter size of the model was only 9.9×10⁶. The model combined the advantages of high accuracy, strong robustness, and lightweight design, providing a feasible technical pathway for precision feeding in intelligent livestock farming.

Key words: dual-modal MobileViTv2, feed residue, non-contact estimation, RGB images, depth images