LCRM: Layer-Wise Complexity Reduction Method for CNN Model Optimization on End Devices

The increasing significance of state-of-the-art convolutional neural network (CNN) models in computer vision tasks has led to their widespread use in industry and academia. However, deploying these models in resource-limited environments, such as IoT devices or embedded GPUs, presents challenges due to increased complexities and resource consumption. This research paper proposes an optimization algorithm called Layer-wise Complexity Reduction Method (LCRM) to address these challenges by converting accuracy-focused CNNs into lightweight models. It evaluates the standard convolution layers and replaces them with the most efficient combination of substitutional convolutions based on the output channel size. The primary goal is to reduce the computational complexity of the parent models and the hardware requirements. We assess the effectiveness of our framework by evaluating its performance on various standard CNN models, including AlexNet, VGG-9, U-Net, and Retinex-Net, for different applications such as image classification, optical character recognition, image segmentation, and image enhancement. Our experimental results show up to a 95% reduction in inference latency and up to 93% reduction in energy consumption when deployed on GPU. Furthermore, we compare the LCRM-optimized CNN models with state-of-the-art CNN optimization methods, including pruning, quantization, clustering, and their four cascaded optimization methods, by deploying them on Raspberry Pi-4. The profiling experiments performed on each model demonstrate that the LCRM-optimized CNN models achieve comparable or better accuracy than the parent models while providing added benefits such as a 62.84% reduction in inference latency on end devices with significant memory compression and complexity reductions.