Abstract
Deep learning based techniques are contributing value to the supply chain of the crops by executing tasks such as crop inspection, grading, maturity detection, disease recognition. Designing lightweight (low trainable parameter count) deep learning models for fruit maturity detection is an active research area and to address this, the following objectives are framed (i) to design a deep learning model for fruit maturity detection (ii) to reduce the trainable parameter (TP) count and model size in the proposed model than existing models under comparison. (iii) to achieve equivalent or better model performance than existing models under comparison. MultiNet is proposed to achieve the said objectives. MultiNet is a group of deep learning models i.e. (i) MultiNet-none (ii) MultiNet-eca (iii) MultiNet-cbam, trained for fruit maturity detection of guava, date, pomegranate, tomato. The architecture of MultiNet consists of proposed MNet modules designed to reduce TP count and model size (in MB). Between cascading MNet modules, the attention modules named eca, cbam may be present or absent (none) based on which the variants are named. The trained models are evaluated for (ii) reduction in model size along with TP count (iii) recall, accuracy, F1-score, precision. MultiNet group of models are compared with existing Resnet18, MobileNetV3, Xception, Squeezenet. MultiNet reports upto 99.27% reduction in model TP count and upto 99.22% reduction in model size making MultiNet “lightweight”. MultiNet has achieved accuracy value of upto 99.39% which is on par with existing models.