Design and Deployment of an Image Polarity Detector with Visual Attention

Abstract Embedding the ability of sentiment analysis in smart devices is especially challenging because relies on deep neural networks, particular, convolutional networks. The paper presents a novel hardware-friendly detector image polarity, enhanced with saliency detection. approach stems from hardware-oriented design process, which trades off prediction accuracy and computational resources. eventual solution combines lightweight deep-learning architectures post-training quantization. Experimental results standard benchmarks confirmed that strategy can infer automatically salient parts polarity an high accuracy. Saliency-based solutions literature prove impractical due to their considerable costs; shows deploy perform successfully variety commercial smartphones, yielding real-time performances.