Parallel R(2+1)D CNN Acceleration Cores Using Winograd for Fish Appetite Detection

In recent years, the development of artificial intelligence (AI) in aquaculture technologies has reached an unimaginable level of complexity with vast variations. One of the AI applications in aquaculture is smart fish appetite detection using convolutional neural networks (CNNs). This article specifically discusses the hardware implementation of the R(2+1)D CNN model, a variety of CNN that deconstructs 3-D convolutions into separate spatial (2-D) and temporal (1-D) components, achieving a balance between computational efficiency and accuracy. In this article, we provide several ideas and insights for the R(2+1)D model hardware implementation, highlighting its design complexity, performance, and potential applications in multimodal deep learning architectures by using the Winograd domain convolution. We also implement the R(2+1)D cores in parallel two and four. Experimental results show that the parallel four cores achieve up to 112.29× and 183.37× faster speed than the reference Tensorflow program for the 1-D and 2-D CNNs, respectively. The proposed parallel R(2+1)D CNN cores are implemented using the PYNQ FPGA board. The work is useful for embedded systems, FPGA implementations, and application-specific IC applications in aquaculture.