Optimized Semantic Segmentation for Fish Detection Using DeepLabV3 and MobileNetV3 on Resource-Constrained Systems

Semantic segmentation of objects in aquatic environments remains underexplored, particularly in scenarios involving real-time processing on resource-constrained devices. Existing methods often struggle with varying lighting conditions, water clarity, and diverse object shapes and sizes, limiting their effectiveness in practical applications. Addressing this gap, this research aims to develop an efficient and accurate semantic segmentation model for fish detection in aquatic environments, optimized for low-resource hardware. We employ the DeepLabv3 architecture with a MobileNetV3 backbone to achieve a balance between performance and computational efficiency. The dataset comprises 390 fish images captured under diverse aquarium conditions, split into 80% training and 20% validation sets. The model is trained over 30 epochs using a combination of binary cross-entropy and Dice loss functions, optimized with the Adam optimizer at a learning rate of 0.0001. Experimental results demonstrate robust performance, achieving a Dice Score of 0.6853, Intersection over Union (IoU) of 0.5605, Pixel Accuracy of 0.9550, Precision of 0.7286, Recall of 0.6879, and an F1 Score of 0.6853. The model’s average inference time is 379.52 milliseconds per image. These findings suggest the proposed approach is well-suited for real-time applications on edge or embedded systems in aquatic monitoring tasks.