Lightweight convolutional neural network models for semantic segmentation of in-field cotton bolls

Robotic harvesting of cotton bolls will incorporate the benefits of manual picking as well as mechanical harvesting. For robotic harvesting, in-field cotton segmentation with minimal errors is desirable which is a challenging task. In the present study, three lightweight fully convolutional neural network models were developed for the semantic segmentation of in-field cotton bolls. Model 1 does not include any residual or skip connections, while model 2 consists of residual connections to tackle the vanishing gradient problem and skip connections for feature concatenation. Model 3 along with residual and skip connections, consists of filters of multiple sizes. The effects of filter size and the dropout rate were studied. All proposed models segment the cotton bolls successfully with the cotton-IoU (intersection-over-union) value of above 88.0%. The highest cotton-IoU of 91.03% was achieved by model 2. The proposed models achieved F1-score and pixel accuracy values greater than 95.0% and 98.0%, respectively. The developed models were compared with existing state-of-the-art networks namely VGG19, ResNet18, EfficientNet-B1, and InceptionV3. Despite having a limited number of trainable parameters, the proposed models achieved mean-IoU (mean intersection-over-union) of 93.84%, 94.15%, and 94.65% against the mean-IoU values of 95.39%, 96.54%, 96.40%, and 96.37% obtained using state-of-the-art networks. The segmentation time for the developed models was reduced up to 52.0% compared to state-of-the-art networks. The developed lightweight models segmented the in-field cotton bolls comparatively faster and with greater accuracy. Hence, developed models can be deployed to cotton harvesting robots for real-time recognition of in-field cotton bolls for harvesting.