Lightweight Tomato Leaf Intelligent Disease Detection Model Based on Adaptive Kernel Convolution and Feature Fusion

Abstract

Timely detection and prevention of tomato leaf diseases are crucial for improving tomato yields. To address the issue of low efficiency in detecting tomato leaf diseases, this article proposes a lightweight tomato leaf disease recognition method. First, enhanced intersection over union is introduced in the you only look once v8 (YOLOv8) model to replace the complete intersection over union loss function, enhancing the accuracy of bounding box localization. To solve the problem of fixed sample shapes and square convolution kernels not adapting well to different targets, lightweight alterable Kernel convolution (AKConv) is introduced, providing arbitrary parameters and shapes for the convolution kernel. Inspired by the lightweight characteristics of AKConv, the C2f module is improved by integrating AKConv to reduce floating-point operations and computational complexity during the convolution process. Second, as it is not feasible to construct a lightweight model with a large depth to achieve sufficient accuracy, a new lightweight convolution technique is introduced. GSConv, combining the GS bottleneck and the efficient cross stage partial block (VoV-GSCSP), replaces the feature fusion layer to achieve lightweight feature enrichment. To test and train the model, a tomato leaf disease dataset was constructed. The improved model demonstrated higher accuracy and fewer parameters on the tomato leaf disease dataset. The improved model achieved an mean average precision 50 (mAP50) of 94.9<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\%$</tex-math></inline-formula> and an mAP50:95 of 75.6<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\%$</tex-math></inline-formula>, representing increases of 1.9<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\%$</tex-math></inline-formula> and 2.8<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\%$</tex-math></inline-formula> over the original model, respectively. The number of parameters is only 2 322 262, a reduction of 22.8<inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$\%$</tex-math></inline-formula> compared to the original model. This method meets the daily needs of tomato leaf disease detection, providing technical support for agricultural spraying robots to quickly and accurately detect tomato leaf diseases and precisely spray pesticides.

Topics & Concepts