Lightweight banana ripeness detection based on improved Alexnet

Objective: To obtain a lightweight Mini-Alexnet banana ripeness grading model and apply it to Android mobile devices. Methods: Based on the external characteristics of bananas with different ripeness, the Alexnet network model was restructured, part of the convolutional layer was deleted, and the global average pooling was used instead of the full connection layer to reduce the model parameters and required memory. A larger convolutional kernel was replaced to extract the global characteristics of the banana skin to achieve an improved lightweight Mini-Alexnet network model. Then the Mini-Alexnet network model was deployed as Android mobile APP, and its feasibility and practicability were verified. Results: The Mini-Alexnet model was only 11.6 MB, and the identification accuracy rate of banana ripeness level 5 was 97.76%. The accuracy rate of local picture recognition mode, photo recognition mode and real-time recognition mode of the mobile APP banana ripeness automatic identification system was 86.66%, 79.33% and 74.00%, respectively, with an average accuracy rate of 80%. Conclusion: The improved Mini-Alexnet model occupies less memory space.