Lightweight Convolutional Neural Network with Residual Attention for Efficient Image Classification

Image classification Deep learning has enhanced image classification performance by a significant margin, but most high-accuracy convolutional neural networks have millions of parameters and massive computational requirements. This complexity restricts their application in resource-constrained systems like mobile devices, embedded systems and edge AI systems. Despite the fact that a few lightweight architectures have been suggested, there is a challenge in preserving an effective trade-off between efficiency in computation and representation of features. To overcome this shortcoming, this paper presents LightCNN-Att, a small convolutional neural network that combines residual learning with dual attention networks to operate with the purpose of providing a better feature extraction at a minimal architecture. The STL-10 dataset is used to evaluate the model and comprises of natural images of ten categories of objects. The experimental findings indicate that the given architecture attains a validation accuracy of 59.38% using around 336K trainable parameters, which competitively performs in terms of its accuracy and is much simpler to implement than traditional deep CNN models. The remaining links facilitate gradient flow within the training process, whereas the attention modules assist the network in paying attention to informative spatial and channel characteristics. The findings mean that LightCNN-Att presents an effective trade-off between accuracy and cost of computation. It is why the proposed model can be applied to edge AI systems, such as mobile vision systems, Internet of Things devices, and other real-time image classification applications with constrained computational resources and energy consumption.