Pattern recognition, in particular image classification, is a computationally intensive process. The best performing algorithmic models in this field are convolutional neural networks with a multi-layer convolution architecture, such as ResNet, GoogleNet or DenseNet. Most of these algorithms are computationally intensive and have high algorithmic complexity. In this paper, we propose KDZnet, a simple CNN architecture that is shallower and more robust and gives results as satisfactory as the CNN architectures in the literature. Two main technologies are used in our model, making it performant. First, we use the acyclic learning technique, which achieves fast convergence rates and high learning performance. We then apply the Batch normalization technique to optimize the network. Our model was tested on Fruits-360 dataset, mainly on tomato plants. The results obtained show that KDZnet achieves satisfactory results comparable to DenseNet or GoogleNet, and takes 10 times less time to run, and the algorithmic complexity is lower.