IMPROVED SMALL AGRICULTURAL PLANT SEGMENTATION USING POST-TRAINING ADAPTIVE NEURAL NETWORK

Precision agriculture enables an automated and data-driven way of improving agricultural crop yields. Soil nutrition, spraying pesticides against pests and diseases can be applied at a large scale. However, defeating weeds poses a significant challenge. High weed localization precision is required as herbicides kill both weeds and crops, cutting or laser removal also should be performed with care. Robotic devices have been proposed to remove weeds, all relying on neural-network-based weed localization. Robots typically perform on-device processing, without Internet connection. Typically, weed removal should be performed at early stages of growth, so the plants occupy a small part of the image, which makes the segmentation task difficult. Existing weed segmentation approaches have insufficient ratio of quality to computation complexity for edge deployment. In the meantime, it is estimated that weeds are accountable for 31.5% yield loss. To solve the problem of on-device weed segmentation, we propose PAN+PTA semantic segmentation neural network, computational complexity of the network can be adjusted after training in a range from 13,08 to 18,12 GFlops. Consequently, the network can be adapted to a wide range of devices without additional training or costly redeployment. We achieve this by 1) integrating the Post-Train Adaptive (PTA) network as encoder in Pyramid Attention Network (PAN); 2) introducing width multipliers to configure initial capacity of the PTA network. To train and evaluate the neural network we use WE3DS dataset, which contains annotations of 7 crops and 10 weeds. The lightest configuration of PAN+PTA achieves higher Dice Score compared to PAN with MobileNetV2 encoder, while reducing the number of computations by a factor of 1.9. Additionally, the trained network in heavy configuration with width multiplier of 1.5 has Dice Score of 0.5112 and computational complexity can be adjusted in range of 32.34%, which is a substantial improvement over existing U-Net+PTA network (Dice Score: 0.4348, range: 3.66%), while reducing inference GFlops by 80%.