Ketk-Densenet Model for Weed Identification in Cotton Fields Based on Densenet and Kans

Abstract

Accurate weed identification in cotton fields is crucial for non-chemical weeding robotic in precision agriculture. However, this weed identification task faces challenges from complex backgrounds, environmental interference, and morphological similarities among species. This study proposes the KETK-DenseNet model based on DenseNet and Kolmogorov-Arnold Networks (KANs), focusing on the identification of cotton seedlings and weeds. Firstly, the model adopts DenseNet121 as its basic architecture and incorporates the KAN-SE-enhanced Efficient Multi-scale Attention (KS-EMA) module after each Dense Block to optimize feature channel weights and enhance the model's precise feature representation. Then, the Transition with Inverted Residual (TIR) module is employed to facilitate the transition between adjacent Dense Blocks to enhance the expression ability. It reduces the dimension and size of feature maps, preserves more critical features, and improves training efficiency. Subsequently, the model employs Instance Normalization (IN) within the Dense Layer, which helps the model better mitigate internal covariate shift and adapt to new data distributions. Furthermore, KAN is utilized as the classification layer, enabling a more refined classification of weed species. Experimental results demonstrated superior performance of the model with 99.25% accuracy and 99.24% F1-score on cotton weeds testing datasets, which significantly surpassed those of mainstream models, such as ResNeXt-50, EfficientNetB5 and ConvNeXt. Meanwhile, the model's convergence speed was rapid and stable, achieving a validation accuracy exceeding 95% in the fifth epoch. The KETK-DenseNet exhibited the capability to accurately and stably recognize weed species in natural cotton field environments, providing essential support for the development of non-chemical weeding robotic technology.

Keywords: Precision weeding, Convolutional Neural Network, Cotton weed identification, Attention Mechanism, Inverted residual structure