Effective Date Fruit Type Detection Using Activation Configuration Tuned State of the Art Deep CNN Architectures

The classification of date fruits is an important step in automated agriculture and food processing. In this work, we introduce a comparative assessment of deep learning models for detecting and classifying six varieties of date fruits: Ajwa, Medjool, Rutab, Shaishe, Sokari, and Sugaye. Using images that are high-resolution and taken in a controlled environment, Several architectures of CNNs, such as MobileNetV2, InceptionV3, and ResNet-50 have been used in this work. Fine-tuning strategies are applied to explore the best activation functions toward achieving higher accuracy with the models. In our study, InceptionV3 demonstrated the best classification performance, achieving an accuracy of 1.00 when combined with Leaky ReLU and Swish activation functions. The Feature extraction from InceptionV3 using Leaky ReLU and Swish activation functions exhibited better generalization capability in date fruit detection. Our experiments show that Swish, in particular, enhanced the learning process through better gradient flow but achieved more robust decision boundaries. This approach is highly suitable for date based industries and automatic harvesting and quality control systems. Selecting the appropriate architecture and activation function for a specific domain is crucial, as this will help in understanding appropriate advancements in such smart farming technologies.