Candy classification using convolutional neural networks, data augmentation and transfer learning: Application and a new public dataset

Quality control is an important task in confectionery because industrially produced candies are prone to suffer from a number of defects, such as cracks, crazing, unacceptable sizes and irregular shapes. Therefore, categorizing the candies into two classes (Non-Defective and Defective) is a key issue in the manufacturing process for the subsequent stages of packaging, pricing and selling as either first-quality products or second-quality ones. According to that motivation, we present and application of deep learning techniques -namely using convolutional neural networks, data augmentation and transfer learning- to address the problem of image-based classification of manufactured candies. A comprehensive experimental evaluation was performed considering several variants for both the deep learning techniques and the set of images. Besides, a new dataset of labeled images of manufactured candies has been built and released for public usage.

I. Introduction

Confectionery is one of the main industries established in Manizales, Colombia. In this city, the largest company in that sector got -according to [1]- a total revenue of 358.216 million COP in 2021; that is, about 90 million USD to the exchange rate at the end of that year. Such amount of income illustrates the importance of this industry for the yearly production of goods in a middle-sized Latin American city as the above-mentioned one.