Classification of Formaldehyde- and Non-Formalin-Containing Chicken Meat Using a Convolutional Neural Network (CNN) with the Mobilenetv2 Architecture in an Android Application

Abstract

Formalin is a toxic substance frequently misapplied to preserve chicken meat; however, it is difficult to identify visually. Conventional laboratory testing is time-consuming, expensive, and unworkable for traditional markets. This study establishes a classification method for identifying chicken flesh adulterated with formalin, using a Convolutional Neural Network (CNN) based on the MobileNetV2 architecture, implemented in an Android application. The image dataset is sourced from primary and secondary sources, then preprocessed and enhanced to increase variability.

The CNN model employs transfer learning and fine-tuning, achieving 98% accuracy and an F1-score exceeding 96% on the test dataset. The trained model is transformed into a TensorFlow Lite (.tflite) file of 2.93 MB, enabling effective offline operation on Android devices. The Android application is constructed with Flutter, allowing users to capture photos or choose images from the gallery for real-time classification. Results are presented with the labels "Formalin" or "Non-Formalin," accompanied by confidence levels.

This system is engineered for user-friendliness and eliminates the need for laboratory equipment, providing precise, consistent categorization outcomes through an intuitive interface. The study also delineates aspects influencing model efficacy, including dataset quality, augmentation methods, and a bifurcated training approach. The designed method aims to serve as a realistic, cost-effective, and autonomous option for consumers and suppliers to identify formalin contamination, while concurrently enhancing knowledge of food safety. Future advancements will include expanding the dataset, field testing, extending to additional meat varieties, and deployment on iOS and web platforms.