Classification Model Based on Deep Learning with Hybrid Loss Function for Trout Freshness Analysis

The trout farming industry in Lake Titicaca, Peru, has witnessed a remarkable growth phase, with annual production surpassing 20,000 tons. Assessment of trout freshness is of great importance for both commercial value and consumer safety. This research aims to enhance the process of determining the freshness of the trout by employing deep learning techniques, specifically convolutional neural networks (CNN). The study seeks to optimize this assessment through the analysis of trout gill images at different time intervals. Customized neural network architectures are developed, leveraging transfer learning techniques, to analyze images of trout gills captured at varying intervals. The approach involves training models to recognize freshness-related features using deep learning methods. The research produces promising results. It is observed that extended intervals between image captures result in improved classification performance. Moreover, the incorporation of a hybrid loss function contributes to improved results across most experimental scenarios. Notably, the performance of the light TroutNet model is constrained due to limited available data. However, when transfer learning is applied with the ResNet architecture, satisfactory results are achieved. This study introduces an innovative methodology for determining trout freshness through gill analysis utilizing artificial intelligence techniques. The research underscores the potential of artificial intelligence in improving production processes within the trout farming industry. By accurately assessing trout freshness, this approach contributes to maintaining quality standards and overall efficiency.