Semantic food segmentation is an important task for the development of nutritional systems that effectively manage daily diets. Recent advances in semantic segmentation have brought great performance improvements. However, these methods require high computational resources that limit their use in a mobile application without relying on external servers to perform the segmentation. Lightweight Convolutional Neural Networks (CNNs) have emerged as an efficient alternative for deploying deep network-based models on mobile devices, a solution not yet applied to semantic food segmentation. In this paper, we propose a lightweight variant of DeepLabv3+, replacing the standard backbone with the lightweight CNN EfficientNet-B1 and the Atrous Spatial Pyramid Pooling (ASPP) with the Cascade Waterfall ASPP (CWASPP). Validation of the proposed lightweight DeepLabv3+ method, in terms of mIoU, parameters and FLOPs, was performed on two public food datasets: UNIMIB2016 and UECFoodPixComplete. The experimental results show a better tradeoff between segmentation performance and computational resources than the state-of-art methods.