This paper proposes a cost-sensitive deep neural network for learning robust feature representations from imbalanced data. The method jointly optimizes class-dependent costs and neural network parameters during training, enabling the model to learn discriminative features for both majority and minority classes. Unlike data-level approaches that alter the original data distribution, this method preserves the data distribution, reducing computational cost. The approach is applicable to both binary and multiclass problems without modification. Experiments on six major image classification datasets show that the proposed method significantly outperforms baseline algorithms and data sampling techniques. The method introduces a cost-sensitive loss function that automatically adjusts based on data statistics, allowing for more accurate classification of minority classes. The cost-sensitive loss function is shown to be classification calibrated and guess-averse, ensuring better performance. The method also introduces a new cost matrix that is suitable for CNN training and modifies the output of the last layer before the softmax and loss layer. The cost matrix is used to adjust the classifier confidences, encouraging correct classification of infrequent classes. The method is evaluated on several datasets, including DIL, MLC, MNIST, and CIFAR-100, showing improved performance in terms of classification accuracy, recall, and F-measure scores. The results demonstrate that the proposed method effectively addresses the class imbalance problem in image classification tasks.This paper proposes a cost-sensitive deep neural network for learning robust feature representations from imbalanced data. The method jointly optimizes class-dependent costs and neural network parameters during training, enabling the model to learn discriminative features for both majority and minority classes. Unlike data-level approaches that alter the original data distribution, this method preserves the data distribution, reducing computational cost. The approach is applicable to both binary and multiclass problems without modification. Experiments on six major image classification datasets show that the proposed method significantly outperforms baseline algorithms and data sampling techniques. The method introduces a cost-sensitive loss function that automatically adjusts based on data statistics, allowing for more accurate classification of minority classes. The cost-sensitive loss function is shown to be classification calibrated and guess-averse, ensuring better performance. The method also introduces a new cost matrix that is suitable for CNN training and modifies the output of the last layer before the softmax and loss layer. The cost matrix is used to adjust the classifier confidences, encouraging correct classification of infrequent classes. The method is evaluated on several datasets, including DIL, MLC, MNIST, and CIFAR-100, showing improved performance in terms of classification accuracy, recall, and F-measure scores. The results demonstrate that the proposed method effectively addresses the class imbalance problem in image classification tasks.