Comparative Performance Analysis of Fine Tuned Optimized Deep Transfer Learning Techniques for Fruit Quality Assessment

Automatic and accurate detection of quality of fruits is critical need of the food industry to save manpower and time. The Transfer learning models are state-of-the-art deep learning techniques which offer high performance for classification chores. They make use of the knowledge from pre-training and require a smaller number of labelled training datasets. This reduces the learning time and cost of preparing large sized labelled dataset. The objective of the current research is to study, apply and evaluate the performance of various contemporary transfer learning models in classification of the quality of five different types of fruits into fresh unspoiled and rotten spoiled classes. Optimize the result using fine tuning and to investigate the most suitable model with highest accuracy, precision, and recall. The outcome of the research exhibited that MobileNet transfer learning model performed with highest classification of accuracy of 99.76%, with 99% precision and recall. The EfficientNet model achieved second highest accuracy of 99.73% and AlexNet exhibited third highest accuracy of 99.56%. The experiment is bench marked with the open-source dataset of fruits. The developed model is suitable for real time use.