Text images subjected to unbalanced lighting exhibit irregular feature distributions, making it challenging for traditional recognition methods to attain low loss values. Consequently, this article introduces a deep learning-based OCR recognition approach tailored for such conditions. This method seamlessly integrates the strengths of lightweight convolutional neural networks and recurrent neural networks. A key innovation lies in modifying the convolution operation within LSTM residual blocks, incorporating a combination of 1 × n and n × 1 convolutions. Additionally, we incorporate Mobile Net normalization after the initial convolution combination to further enhance the model’s adaptability. During the recognition process, RMSProp optimization algorithm is employed to assign distinct learning rates to individual parameters within the enhanced model. By adjusting the step size of gradient descent, we optimize the model’s training efficiency. Ultimately, the Mobile Net network generates OCR recognition results for text images under unbalanced lighting. Experimental findings reveal that variations in Epoch have minimal impact on recognition outcomes. Notably, the loss value remains consistently low, hovering around 2.0 with a peak of 1.75. These results demonstrate a clear advantage over the control group, establishing the proposed method as a viable solution for text image recognition in challenging lighting conditions.