Microbial Colony Image Enhancement Algorithm Based on Composite Adversarial D-DCGAN Neural Network

Microbial testing plays a crucial role in medical and biological research, serving as a significant direction in disease prevention and control studies. Traditional manual testing, which involves identifying colonies on Petri dishes is the best method, but it is not only expensive but also time consuming. However, the application and development of deep learning can address these issues. However, a crucial condition is that deep learning models require a large amount of labeled medical imaging data. Obtaining high-resolution images is time consuming and cannot generate a large amount of data in a short time. Therefore, we designed a Double Deep Convolutional Generative Adversarial Networks (D-DCGAN) to construct a dataset. Under the DCGAN adversarial network, synthetic colonies are generated by overlapping small single colonies and composite colonies to establish a random synthetic colony model, and a DCGANB generation model is established for the synthesis of complete colonies. In this study, we first evaluated the D-DCGAN model, obtaining an initial FID score of 38.9, and validated the effectiveness of the generated dataset using Faster R-CNN and YOLOv12. Subsequently, a second-order optimization model based on multiple evaluation metrics was developed to determine the optimal augmentation level, identifying 1320 synthetic images from the original 236 samples. With this optimized dataset, the FID score decreased to 37.1, while the detection accuracy reached 0.975. These results demonstrate the feasibility of the proposed framework for effective dataset augmentation and improved downstream detection performance.