Retail product recognition on grocery shelf photographs is a challenging task due to the variety of products and lighting conditions. This research proposes a novel method for completing this task efficiently and in a short amount of time. The object detection and recognition module is deployed on a jetson Nano board, which is trained to trace a path for real time object image capture. The proposed method consists of two stages: detection and recognition. The detection stage is performed by a generic product detection module that has been trained on a particular class of products (e.g., tobacco packages). This is accomplished using the Viola and Jones Cascade Object Detection Framework. The recognition stage is performed using two FCN designs: U-Net and Fully Convolutional Regression Network (FCRN). We extract both shape and color information by applying feature-level fusion to two distinct descriptors computed using the bag of words method. We evaluated our method on a dataset of grocery shelf photographs. We achieved a detection accuracy of 95% and a recognition accuracy of 90%. The proposed method is efficient and accurate, and it can be used for a variety of applications, such as inventory management, price tracking, and fraud detection.
Part of the book: Deep Learning