Food waste is a significant issue affecting the global economy, human health, and environmental elements (Chigurupati et al., 2025). Every year, over one trillion dollars worth of food is wasted instead of being used to address the hunger of 782 million people worldwide (Gravert & Mormann, 2025). Furthermore, the consumption of spoiled foods has impacted the spread of foodborne illnesses, with 1.6 million fatalities attributed to food contamination per year (Chowdhury et al., 2025). Additionally, food waste affects the environment due to the release of methane and other greenhouse gases during the production process, causing resource inefficiency (Reddy et al., 2025). Currently, a prevalent issue is the wastage of food before and after the sales process, which contributes 60% of the total food waste (Gravert & Mormann, 2025).
This problem is especially prevalent in urban households with busy storage locations, such as packed refrigerators. With the amount of food available, storage in such areas becomes disorganized, leading to items being forgotten and eventually spoiling. Often, to prevent such issues, households manually record the items inside their refrigerators and the time they were stored. However, this method is extremely time-consuming and often results in individuals forgetting to record items. A common cause includes human dependency for detecting food waste, which leads to large expenses while being unreliable (Chigurupati et al., 2025). Current automatic methods are either unable to identify a variety of foods, struggle to detect spoilage, or require manual input from a user for each item. Thus, there is a need for urban households to have a camera system that detects whether fresh and cooked foods in refrigerators show signs of spoilage to reduce food waste. This project addresses various problems, including the lack of surveillance of food in refrigerators and their eventual spoilage due to the lack of reminders for usage. It also includes solutions to inaccurate prediction systems and the lack of flexibility in product type.
To identify visual spoilage, various types of deep learning models were used. Deep learning is a branch of artificial intelligence that enables computers to identify complex patterns from large amounts of data. The models automatically identify and classify objects in images without manual rules, allowing them to handle complex visual tasks such as food spoilage detection (LeCun et al., 2015). Convolutional neural networks (CNNs) are a common type of deep learning model used for image analysis, while object detection models such as YOLOv8 by Ultralytics can isolate multiple items within an image. By using such technologies, spoilage is identified automatically, which is difficult to achieve with traditional manual methods (Chigurupati et al., 2025). However, for these models to be effective in all situations, accurate, abundant, and structured data is essential.
Data preparation consists of many steps, including locating and recording the data, formatting it properly, and dividing it into categories. First, the model must have access to accurate data. Then, the data is formatted to be consistent and applicable to the specific model. After storage, the data is divided into testing, validation, and training sets with a ratio of 7:2:1 (Wang et al., 2021). This ratio allows sufficient training data while enabling predictions for validation and testing. Often, there may not be enough data available, leading to the need for data augmentation. Data augmentation involves slightly altering an image and adding it to the dataset for training and testing. Some alterations include rotations, slight color changes, reflections, translations, and adding speckles.
Convolutional neural networks (CNNs) are designed to automatically analyze and classify images through visual data. These models extract features from images to identify subtle differences that may indicate spoilage, such as changes in color or shape (Chowdhury et al., 2025; Chigurupati et al., 2025). CNNs are useful in food monitoring due to their ability to generalize across various food types without requiring specific manual rules. By learning directly from image data, these models provide a scalable solution for large datasets.
After the data is prepared, the CNN model is constructed and used for training. The CNN model consists of five main types of layers: the input layer, convolutional layers, pooling layers, a flattening layer, and dense layers. The input layer ensures the images are formatted correctly and consistently. The convolutional layers identify patterns, with each subsequent layer identifying more complex patterns. For example, initial layers detect edges and colors, while later layers analyze signs of spoilage such as spots or wrinkles. After each convolutional layer, a pooling layer summarizes the most important features within a specific grid area. The flattening layer converts two-dimensional feature maps into one-dimensional vectors, allowing dense layers to make final classification decisions. A dropout layer is used to prevent overfitting and improve generalization (Chowdhury et al., 2025). This layer ignores a percentage of learned patterns to ensure the model focuses on recurring features. Important hyperparameters include the learning rate, the number of epochs, and the dropout rate (Chigurupati et al., 2025). These methods enable the system to detect spoilage patterns and improve food storage efficiency.
Although CNNs are effective for classifying individual images, real-world refrigerators often contain overlapping items that require object detection. To address this challenge, an object detection model called YOLOv8 was used. This model provides bounding boxes, which are rectangular frames around identifiable items, along with confidence scores (Wang et al., 2021; Hemavathy et al., 2023). YOLOv8, standing for "You Only Look Once", is suitable for fast, real-time detection (Wang et al., 2021). Due to its ability to generate accurate bounding boxes, YOLOv8 identifies foods in images and saves a frame of each item for further analysis.
Despite advances in deep learning and object detection, current spoilage detection systems still have limitations. Many struggle with cluttered and realistic environments. As a result, most retail applications require manual input for each item or are limited to specific food types (Chigurupati et al., 2025). Additionally, user interfaces are often unclear and rarely provide real-time actionable alerts. These gaps justify the development of a deep learning based system that detects early signs of spoilage across multiple food items and provides reliable alerts to reduce household food waste.

• Use an apple as the fruit for the preliminary model
  • Use the Fruits Fresh and Rotten Kaggle Dataset to organize the information into accurate folders for training and identification.
  • Create a CNN model to train with the data from the dataset, focusing on preventing overfitting.
  • For more advanced models, it may be wise to use more complex CNN network structures, like VGG, ZFNet, GoogLeNet (Liu et al., 2016).
• Identify whether certain examples of the product have spoiled with an accuracy of greater than 90% within a training period of 10 days
  • Identify spot spoilage and color changes to predict spoilage.
  • Use the barcode or the use-by-date if applicable to predict spoilage.
  • Use the insertion and removal time of the product to predict spoilage.
  • Combine the three situations to identify the most probable class.
• Identify spoilage in at least 5 more items with an accuracy greater than 90% for each within a training period of 10 days
  • Use the Fruits Fresh and Rotten Kaggle dataset with access to bananas and oranges as well.
  • Allow the model to contain multiple classes and classify each product.
• Allow the user to input full images of new products, provide a label for the product, and get a prediction from the model after 10 inputs with an accuracy of 70%
  • Collect the data and augment it to prevent overfitting.
  • Create an autonomous system where the model can begin to train itself and adjust hyperparameters until it can predict the new product.
  • Allow the model to make predictions from the information provided and store the learning data.
• Display the model through an easy-to-understand application which streamlines the process of checking for food spoilage
  • Create an interface which displays the items present in the refrigerator and their approximate spoil-by date.
  • Alert the user when the spoilage date is approaching.
  • Provide a user-friendly system to capture images for future predictions with new products.
  • Assist in recipe choice through existing recipe AI models and use a probe object to identify the amount of food as it is used for volume calculation of food items (Wasif et al., n.d.).