Developing a CNN-Based Model for High Accuracy Fruit Image Classification

Recently, we have been developing a fruit and veg classification model inspired by the need to make scanning groceries more convenient. Our goal is to develop a convolutional neural network (CNN) that is capable of accurately classifying the different types of fresh produce.

Dataset Development

For us to train the model, we gathered publicly available and manually collected images to form a comprehensive dataset. Throughout the process, the dataset grew to several gigabytes in size with the focus of ensuring diversity in the training data to improve the model's ability to generalise on unseen data. This is key to solving the challenge of identifying the produce across varying conditions.

Model Architecture

The CNN architecture was used as it is effective in image classification tasks. At the start, the performance of the model was suboptimal however after several training iterations and fine-tuning we saw significant improvements. By the third training cycle, the model achieved an accuracy of approximately 87% on unseen data.

With further adjustments and data augmentation techniques, we were able to push the accuracy above 99% on unseen data for certain categories showcasing the model's potential and impact of iterative refinement in machine learning.

Limitations

While significant progress was made, the model still faces limitations in certain scenarios like variations in the image backgrounds, lighting conditions and other real-world factors that could affect the classification.

What's next

Looking ahead, we plan to continue developing and improving this model and integrating it to help solve real-world problems.