Lightweight Saliency Detection Network-based Target Recognition Method for Livestock Carcass Chopping Robot

Existing methods for salient object detection mainly concentrate on how to improve the prediction quality. The computational overhead and hardware requirements of the model are ignored. The result is that it cannot be effectively applied in practice on industrial equipment. To solve this problem, a lightweight saliency detection algorithm for livestock carcass chopping robot target detection is proposed in this paper. First, a redundant feature generation module is introduced in the encoding stage. To ensure that richer salient features are extracted with fewer parameters. Secondly, a selection residual module is introduced in the decoding stage. Compress the number of channels while maximizing the use of multi-scale features. Finally, a hybrid distance loss function is designed in this paper to facilitate the model to focus more on the overall salient regions. Experimental results on public datasets and self-made livestock carcass dataset showed that. The proposed method uses a smaller number of parameters and achieves higher detection accuracy than other lightweight saliency detection models. And the target recognition of livestock carcasses can be effectively performed on the self-made livestock carcass dataset.