End-to-end food condition system via wavelet attention enhanced network

Automatic vegetable quality analysis systems can be useful in optimizing production processes and increasing harvest efficiency and effectiveness. For this reason, in this paper, we propose an end-to-end system that can be easily used in practical applications in production halls or even in mobile applications. The model operates on data obtained from the camera, which takes a photo of the vegetable and sends it to the convolutional neural network module. To improve classification quality, wavelet attention was proposed for better analysis of features during the processing of data in the network. For this purpose, we defined low and high frequencies that allow the creation of a score matrix showing important information. A proposed network allows for the recognition of objects with various textures and focuses on local and global feature extraction. Additionally, a system architecture that automates the addition of classes in the network to build an end-to-end solution was described. The modeled framework was tested on three publicly available datasets - PlantVillage, ATLDSD and Indigenous. The obtained results allowed for achieving higher evaluation metrics than the state-of-the-art.