Deep Learning-based Semantic Segmentation Of Crops And Weeds

Weed control is a crucial farming operation that suppresses unwanted plants to allow the crop to utilize the soil nutrients,sunlight,and water without any competition for its growth.The weed control mechanism has evolved from the traditional,mechanization to autonomous Robotics.Currently,Intelligent farm robots empowered by proper vision algorithms are the new generation of agricultural machinery that eases weed control with speed and accuracy.The performance of the Robot depends on the robustness of the vision algorithm.Deep convolutional neural networks(DCNN)have elevated the segmentation of crops and weeds to another milestone.However,the strong greenness similarity between crops and weeds,the complex background interference from other objects,and the environmental variation such as changes of illumination are among the major challenges for proper vision in the crops and weeds environment.The existing DCNN network focuses on general detection without considering the basic details of crops,weeds,and the surrounding environment,which lead to missing essential cue that will improve the performance of the segmentation network.Based on these substantial nitty-gritties that will ultimately affect the model performance,an improved Deep neural networks are proposed to generate the essential shape and textual cues to improve the contextual aggregation,which expose the crops,weeds boundaries,and leaf fractals for the precise pixel-wise semantic segmentation of crops and weeds.Our main contribution is as follows:1)A lightweight backbone for feature extraction is designed based on Res Net basic block that reduces the number of channels to lower the model complexity.Also,an extra residual connection is utilized to adjust the signals of the feature patches considered irrelevant and further used as input to the subsequent layer of the backbone to extract more meaningful information.In addition,a multi-level feature re-weighted fusion(MFRWF)module was proposed to combine only the relevant information from every backbone layer output to improve the crop and weed objects contextual learning.2)An attention-based decoder is suggested to combine the rich multi-level features context and efficiently aggregate the global context information to generate a single feature output that exposes the targets shape boundaries and leaf fractals.An alternative simple and robust convolutional weighted fusion(CWF)decoder is also designed to speed up the inference during the prediction.3)A dual encoder that functions as two separate MFRWF is designed to incorporate two distinct feature extraction algorithms that extract different sequential semantic cues of the target objects in two-fold of feature outputs.A hybrid feature selection module(HFSM)is suggested to iteratively align the distinctive semantic cues and allow the two-fold feature outputs to complement one another for better model prediction.