An End-to-End Deep Residual Learning Paradigm for Robust Visual Defect Detection and Quality Assessment in Baked Product Processing Systems

Abstract

In contemporary food manufacturing environments, the large-scale production of baked goods, necessitates highly reliable and efficient quality inspection mechanisms. Even minor visual imperfections, such as color inconsistencies, surface cracks, irregular shapes, or the presence of foreign objects, can result in significant product rejection, financial losses, and diminished consumer confidence. To overcome these challenges, this study presents an Automated Visual Inspection (AVI) framework designed for the classification and analysis of defects in baked goods using advanced computer vision and machine learning techniques. The proposed system processes image data collected from production lines and applies standardized preprocessing methods, including resizing, normalization, and enhancement, to ensure visual consistency. Feature extraction is performed using a pretrained Residual Network (ResNet), leveraging transfer learning to capture critical attributes such as texture, color, and structural patterns. These extracted features are then classified using Linear Discriminant Analysis (LDA), Light Gradient Boosting Machine (LGBM), and Extra Trees Classifier (ETC) to enable comparative evaluation. Furthermore, a ResNet-based Convolutional Neural Network (CNN) is implemented for end-to-end image classification. The system categorizes products into four classes: color defect, non-defect, object-related defect, and shape defect. A Graphical user interface (GUI) developed using Tkinter facilitates real-time image input and result visualization. Experimental findings demonstrate that the proposed framework enhances accuracy, consistency, and operational efficiency in industrial quality inspection.