A Baking Maturity Detection Method Based on Single Shot Multibox Detector and Center Feature Fusion for Mobile Application

Abstract

Maturity has a significant impact on the quality and safety of baking foods. Current maturity detection of baked foods faces the problems of insufficient detection capability for small objects, low accuracy, and poor detection efficiency. This paper proposes a real-time classification method for Baking Maturity Detection (BMD) by improving the baseline (SSD). Firstly, a dataset with 29,360 images of baked foods is constructed using small-sample enhancement strategies, including Generative Adversarial Network (GAN), Random Augmentation (RA), and multiple samples fusion based. Secondly, the backbone network of SSD is replaced with a synergetic backbone network by combining Residual Network 50-layers (ResNet50) and Visual Geometry Group 16-layers (VGG16) in parallel, in order to fuse the high-level and low-level features effectively. Thirdly, a center feature fusion module is introduced to enhance the small object recognition ability through top layer linear interpolation and bottom layer downsampling. A new weighted loss function based on confidence loss, localization loss, and regularization is designed to balance the convergence rate and overfitting. In addition, an embedded mobile detection platform is designed, and the ablation study and comparative experiments are conducted to verify the superiority. The average maturity detection accuracy for five typical bakery goods reaches 93.85%, which improves by 6.22% compared to the baseline model. A detection speed of 447 ms is achieved on the mobile platform, showing great application potential in fast and accurate BMD for various food products.