Image segmentation approaches are among the crucial tasks in computer vision applications, such as object recognition, tracking, agriculture, autonomous vehicles, and medical imaging, relying heavily on deep learning neural networks (NN) for precise region delineation in images. This work implements an image segmentation task using an in-memory computing architecture on NNs. Magnetic
... [Show full abstract] random-access memory (MRAM) offers significant advantages for implementing NNs among various non-volatile memories. However, it faces challenges such as low resistance of synaptic MRAM devices, resulting in high power consumption in conventional crossbar arrays utilizing current summation. This work proposes a method for resistance summation that facilitates analog multiply and accumulate (MAC) operations. The approach leverages a differential spin Hall effect (DSHE) MRAM-based crossbar array, capitalizing on its high storage density, speed, and energy efficiency, mitigating power consumption issues in image segmentation tasks. The proposed approach achieves 11.1× and 1.83× more energy-efficient MAC computation compared to the existing spin–transfer–torque (STT) and spin–orbit–torque (SOT)-based designs, respectively. Moreover, the presented DSHE-based approach lowers the energy consumption by 11.05× compared to the STT-based implementation for MAC computation. The proposed approach has been utilized to implement image segmentation tasks on UNet and SegNet architectures. In comparison to MAC implementations using STT and SOT-based current and STT-based resistance summation operations, the DSHE-based design for image segmentation achieves reductions of 57×, 4.5×, and 622× in energy consumption and 20×, 2.8×, and 20.66× in latency, respectively, for the UNet architecture.
