Towards an Efficient CNN Inference Architecture Enabling In-Sensor Processing †