Concrete Crack Identification Using Convolution Neural Network

This study presents a CNN-based model for detecting cracks in concrete images. Even though the model is simple and has only a few layers, it provides significant accuracy. This type of CNN model is easily deployable in mobile devices with limited computation and memory power. The proposed model outperforms the results of the previous literature survey. The proposed CNN-based model is simple yet efficient, requiring only nearly 32000 parameters, far fewer than most pre-trained models. Unlike previous research, which has relied on pre-trained CNN architectures, the proposed model addresses the limitations of global feature representation and computational resource requirements. The model is made up of five convolution layers, followed by max pooling layers, with a 3x3 kernel size and a ReLU activation function. For classification, the final layer employs global average pooling and sigmoid activation. With approximately 32,865 parameters, the architecture is computationally efficient. It outperforms previous studies in the literature with a validation accuracy of 99.18%.