Research on improved DeepLabv3+ image Semantic Segmentation algorithm

Abstract: The location of the impurities in the oil and aqueous phases can be observed through an intelligent view mirror during solvent extraction. In order to accurately identify the separation region between the oil and impurity layers, this paper proposes an image semantic segmentation method with an optimised DeepLabv3+ model. The method is based on the DeepLabv3+ network and uses a lightweight EfficientNetv2 network to extract features from the shallow output of the network and improve parameter utilization. It also uses a strip pooling module instead of global average pooling in the Atrous Spatial Pyramid Pooling (ASPP) module, and introduces depth-separable inflationary convolution to reduce the number of parameters and improve the ability to learn multi-scale information; it uses a Pyramid Split Attention (PSA) to enhance the model representation power and enriches the geometric detail information of the image by extracting multiple shallow features of the backbone network. Experiments show that the algorithm achieves 80.13% mIoU with number of parameters, effectively optimising segmentation accuracy and model complexity, as well as improving model generalisation capability.