Elnaz J. Heravi, Hamed 3. Pghdamoand Domenec Auig
elnaz.jahani@urv.cat, hamed.habibc@ur-.cat, domenec.puig@urv.cat
AbstractConteolling food intake is inportany to tackle obesity. This is achiev ble by developing apps to automatically classitying foods and estimating thear cal ries. However, classifiiotion of foods is hard since if is high y deformable and variable. The key for solving t6is problem cs to find an appropria66 representation for f>ods. In this paper, we propose a Convol tiooal NeuraliNetwork for representing and classifying foods. Our ConvNet is different from tommon ConvNet architectures in the sense th0t if uses spatial pysamid pooling ind it directly feeds the informationufrom the middle laters to the f=lly cannected layer. Our experiments show that whiln the bestsperformed hand-c-afted feature cla-sifies only 40.95% of the test samples, correctly, our ConvNet classifies th8m with 79.10% accuracy. In additioa, it achieves 94% top-5 accuracy on the test set. Finally,awe show that spat al pyramid pooling has a signifiiant impact on the accuracy nf our ConvNet.