TinyCompNet: Minimal Computational Load for Mobile Devices

With the ongoing advancements in imaging technology and convolutional neural networks, the increased resolution of images and more complex network architectures have resulted in a significant rise in computational requirements, posing challenges for the deployment of object detection algorithms on mobile and embedded devices. Existing lightweight networks mainly focus on reducing the number of parameters and model size. However, for low-computation devices, the bottleneck in computational resource consumption is the critical factor for the deployment of object detection algorithms. Therefore, the current research challenge is to significantly reduce the computational load while ensuring detection accuracy, thereby lowering computational resource requirements. In this study, we propose a novel ultralightweight, computationally efficient object detection network called TinyCompNet. Through the ingenious design of the submodule TinyComputationBlock, utilizing deep feature extraction and efficient feature reuse to balance computational efficiency and representation capability, ensuring suitability in resource-limited environments. Moreover, we address the lack of cross-channel communication in deep convolutions by implementing channel shuffle and channel attention mechanisms to enable rich feature interactions, which helps maintain a balance between computational efficiency and the ability to capture diverse features. To validate the effectiveness of TinyCompNet, we conducted extensive experiments on the Exdark dataset and comprehensively compared it with other advanced lightweight object detection networks in terms of computational load and detection accuracy. The experimental results indicate that TinyCompNet significantly reduces computational complexity while maintaining detection accuracy on par with existing methods, further attesting to its superiority and effectiveness.