Fruit Quality Detection using Convolutional Neural Network-based Cosine Annealing Learning Rate

Fruits are essential foods consumed in day-to-day lives of most people and are primarily recommended as a source of nutrient supply by experts and nutritionists. Computer vision technologies employ numerous approaches for fruit detection. However, detecting fruit quality remains difficult because of the variations in size, ripeness, and surface defects which result in them looking subtly similar. So, this research proposes a Convolutional Neural Network-based Cosine Annealing Learning Rate (CNN-CALR) to detect fruit quality effectively. CNN-CALR adapts LR dynamically, enhancing the convergence by minimizing it with a cosine curve. This increases the model’s training stability and results in better performance. Initially, the Fruit 360 dataset is applied to analyze the model’s performance. Data augmentation is applied in the preprocessing stage to increase the dataset size. Then, a ResNet50 is employed to extract the features while CNN-CALR is employed for the detection of fruit quality effectively. The proposed CNN-CALR achieves a high accuracy of 99.92% when compared to the existing methods, Cascaded Adaptive Network-based Fuzzy Inference System (Cascaded-ANFIS) and DenseNet-201.