Development of a mixture of experts with optimized EfficientNet features (MoEffNet)-powered automated identification system for images of agricultural equipment

Visual data can be used to identify agricultural machines automatically, which is needed to facilitate precision farming, asset management, and autonomous fleet management. Nevertheless, agricultural real-world imagery poses some serious problems because of great intra-class variability, inter-class similarity, and uncontrolled environmental factors, including occlusion, glare, and motion blur. To solve these problems, this paper suggests MoEffNet, a new deep learning architecture that incorporates a Mixture of Experts layer with an EfficientNet backbone, which is with its training hyperparameters optimized by a new metaheuristic algorithm, the Refined Addax Optimization Algorithm (RAOA). MoEffNet is a dynamic system that routes inputs into specialty expert subnetworks; it is capable of extracting features adaptively for visual subclasses of machinery. The RAOA optimizes EfficientNet scaling coefficients, router temperature, and learning rate to achieve the highest accuracy without violating computational constraints. MoEffNet was also tested on a publicly available image dataset of 18,742 images of 12 different types of machinery and obtained 98.72% Top-1 accuracy, which is better than seven state-of-the-art methods such as EfficientNet-B4, Vision Transformers, and AMTNet, and it is also extremely fast in terms of inference throughput, making it suitable to run on the edge. Qualitative Grad-CAM analysis showed that the model concentrates on the discriminative machinery parts rather than background artifacts. The work defines a new standard in the area of agricultural equipment recognition through vision and shows the effectiveness of bio-inspired metaheuristics in the refinement of domain-specific neural architectures.