Weed Identification And Application In Sweet Potato Seedling Stage Based On Improved YOLOv5s

As the largest sweet potato producer in China,sweet potato weed control is of great importance to improve sweet potato production and quality.The current reliance on manual weed identification is somewhat misleading.Driven by the trend of digital agriculture,scholars have used target detection techniques to achieve automatic weed identification,but most of the current studies only consider weed detection with a single background and single target,and because sweet potato seedlings and weed leaves are highly similar and the farm environment under different weather conditions is also complex,resulting in low weed identification accuracy,which makes it difficult to apply in real farm scenarios.To address these problems,it is important to study an algorithm that can identify sweet potato seedlings and weeds accurately and in real time in a real field environment for weed control in sweet potato fields.Therefore,this article implements a high-precision,real-time detection of weeds at the sweet potato seedling stage in complex field environments based on a target detection algorithm,which works as follows:(1)A sweet potato seedling weed dataset was constructed.The full cycle of sweet potato seedlings was recorded,and sweet potato seedlings and weeds were collected at different angles and under different weather conditions.The images were manually annotated and then normalised and enhanced to create a dataset containing eight targets: sweet potato seedlings,duckweed,iron amaranth,thistle,concave-headed amaranth,oxalis,bogeyweed and horsetail.At the same time,a Contrast-Limited Adaptive Histogram Equalisation(CLAHE)image pre-processing algorithm was introduced to address the problem that weed leaf texture and edge shape are more difficult to identify,and contrast enhancement was applied to images such as texture and edge blur to highlight leaf texture edge detail features.The final processed images are divided into a training set,a validation set and a test set according to a ratio of 7:2:1.(2)Construct a YOLOv5s-SKS sweet potato seedling stage weed recognition algorithm.In this article,the YOLOv5 s network with the optimal real-time performance,the smallest weight file and the most suitable for deployment to embedded devices is selected as the base network.In order to suppress non-essential feature interference information and improve the feature extraction capability of weed images,the SK attention module is added to the backbone network of YOLOv5 s,which solves the problems of low differentiation between weeds and crops and serious feature loss.On this basis,the regression loss of the model was replaced with SIo U＿Loss to further improve the model performance and the recognition accuracy of the network model.After various sets of experiments,the results show that the m AP value before optimisation is 84.3%,and after optimisation the m AP value of YOLOv5s-SKS is improved to 93.2% and the FPS is 54.6,which meets the requirement of real-time performance.The proposed method has a good detection effect on sweet potato weeds in complex backgrounds.(3)A weed identification system for sweet potato seedlings was designed and developed.Using YOLOv5s-SKS as the core algorithm,the model was transformed using the NCNN framework and deployed into the mobile terminal.The system contains modules such as image recognition,weed encyclopedia and sweet potato production knowledge.After testing the system,it was proved that the system meets the requirements of high accuracy and real-time,and can provide reference for accurate application and weed control.