Nowadays, application-specific integrated circuits (ICs) of convolutional neural networks CNN are used to make an automated and intelligent system. However, the high memory and computational requirements make deploying CNN challenging in resource-constrained or edge IoT devices. Since CNN exhibits several computational operations (MAC) and parameters, the implementation complexity of CNN can be reduced by reducing these MAC operations and parameters. This paper proposes an energy-efficient CNN architecture with reduced parameters and MAC operations for image data classification. The reduction in complexity is achieved by using depth-wise separable convolution and reducing the size of the fully connected layers by adding one extra convolution layer. Post-training quantization of the proposed architecture is also done to reduce the model size further. The classification accuracy of the architecture with and without quantization is tested and compared on the MNIST dataset. The simulation results show that the proposed LeNet CNN architecture significantly reduces the parameters (i.e., by 69.3% and 51%) and MAC operations (50% and 97%) while only compromising 0.2% accuracy over the previously proposed LeNet CNN architectures. Further, the FPGA implementation of depth-wise separable convolution using the proposed approach shows significantly lower memory, execution time, and total on-chip power consumption over the normal convolution.