Efficient Weed Segmentation with Reduced Residual U-Net using Depth-wise Separable Convolution Network

Selective weed treatment is a cost-effective method that reduces manpower and usage of the agrochemical, at the same time it requires an effective computer vision system to identify weeds and should be smaller in size to run in resource-constrained devices. To accomplish this, a convolution neural network named Reduced Residual U-Net using Depth-wise separable Convolution (RRUDC) network is proposed in this paper. Residual Depth-wise separable Convolution Block (RDCB) is introduced as a functional unit in both contractive and expanding paths. Residual connection is incorporated inside every RDCB unit. This network employs semantic segmentation to analyze the crop field images pixel-wise. To reduce the parameter size, a depth-wise separable convolution technique is used which curtail the number of parameters generated by the model at a ~1/9 ratio with a very negligible drop in accuracy. The model is trained using Crop Weed Field Image Dataset (CWFID) and then the trained model is pruned to reduce the model size further. It compresses the final model size by around ~70% without affecting the performance. It has achieved segmentation accuracy of ~96%, a lesser error rate with a model size less than 3 MB. It can be compatible with converting the proposed deep learning model into a real-time computer vision application that seems more convenient for farmers in their resource-constrained devices on their agricultural land.