Multi-Label Weather Image Classification for Embedded Computing Platforms

In this work, we propose a Multi-Label Weather Classification framework (MLWC) designed for embedded computing platforms. The MLWC framework consists of five Convolutional Neural Network (CNN) models, where each one is trained with a specific adverse weather condition, i.e., clear skies, cloudy, rainy, foggy, snowy. In order to develop lightweight CNN models, we have applied the novel Sparse-Split-Parallelism (SSP) framework to DenseNet-121, achieving a model size reduction of 55%. The experiments prove the effectiveness of the multi-label framework, by correctly predicting multiple weather conditions in a single image. In particular, the weather classification models, lightened with the SSP method, achieved accuracy rates of 90.95% on clear skies, 59.24% on cloudy weather, 91.54% on foggy weather, 73.33% on rainy weather, and 95.22% on snowy weather. Following that, their performance is compared with other lightweight models, such as, MobileNetV2, MobileNetV3 and EfficientNetB0, confirming the advantages of the SSP framework. Lastly, the proposed MLWC framework is evaluated on the NVIDIA Jetson Nano board, which achieves multi-label classification results in 4.2 seconds per image. With the accomplished inference time, we prove that the MLWC framework is suitable to be deployed in small embedded platforms that can be used in autonomous driving systems, like vehicles or drones, and maintain a good performance.