Fusion of Enhancing Target Features and Focusing Slice Features for Soil Image Classification

Soil classification based on the identification of soil surface and physical properties is a key point in calculating indicators such as stability, bearing capacity, and uniformity of construction sites in soil exploration. In this paper, soil image classification is regarded as a sub task of fine-grained image classification, and a soil image classification algorithm model integrating target feature enhancement and slice feature focusing is proposed. This model constructs a three branch network for Supervised learning. First, the backbone network obtains the overall characteristics of the image; Then, the image position is obtained through cross level feature fusion to reduce background noise interference and enhance target features; Finally, based on the feature mapping relationship of the target image, a local region with the highest degree of differentiation and the lowest degree of redundancy is proposed to achieve slice feature focusing. The proposed model has achieved competitive results on CUB-200-2011, Stanford Cars, FGVC-Aircraft and Soil.