Ensemble of multi-task deep convolutional neural networks using transfer learning for fruit freshness classification

Automatic classification of fruit freshness plays an important role in the agriculture industry. In this work, we propose an ensemble model that combines the bottleneck features of two multi-task deep convolutional neural networks with different architectures (ResNet-50 and ResNet-101). In our proposed multi-tasking framework, there are two classification branches: a binary classifier to distinguish between fresh and rotten fruits, and a multi-class label classifier to identify the kind of fruit. Since the features (e.g., color, texture, and shape) of rotten fruits are different from each other depending on the kind of fruit, the input of the first branch is combined with the kind of fruit information from the second branch to classify the fruit freshness more accurately. Transfer learning technique has been applied during the model training since transfer learning has been shown to be effective transfer learning has been shown to be effective in many applications in which training data for the target problem are limited. To evaluate our proposed model, we use simple images from the existing dataset and real-world images crawled from the web, both representing fresh and rotten fruits for different fruit categories as our dataset. Our proposed model achieved average accuracies of 98.50% and 97.43% for freshness classification and fruit classification, respectively, demonstrating that our transfer learning-based ensemble model outperforms other transfer learning-based models.