Field detection of small pests through stochastic gradient descent with genetic algorithm

Pest invasion is one of the main reasons that affect crop yield and quality. Therefore, accurate detection of pests is a key technology of smart agriculture. Pests often exist as small objects with limited features in the actual field. Deep neural networks, as promising small object detectors, are adopted to fully acquire the feature information. The pest detection network has a large number of parameters to be trained, where the current stochastic gradient descent method may tend to fall into local optimum and lead to poor pest detection precision. To solve the above issue, we propose the GA-SGD algorithm to help the SGD jump out of the local optimal trap. It consists of selection operation, crossover operation and mutation operation. The selection operation selects fine solutions from the parent population, crossover operation exchanges and combines the two solutions to generate the new offspring, mutation operation replaces the original value with the random value to produce the new solutions. Experiments show the proposed GA-SGD achieves higher detection accuracy and stability than five algorithms on three object detectors. The results indicate small pests are detected with superiority. It also proves the effectiveness and value of the proposed algorithm.