Research on Pest Target Detection in Smart Kitchen Based on Deep Learning

Regarding the target detection of harmful animals in indoor enclosed environments by camera surveillance systems, there are issues of target recognition errors or failure to recognize due to factors such as changes in lighting, indoor furnishings, and the size of the identified target. A new method for indoor harmful animal target detection is designed and proposed. This method is based on the yolov5-5.0 framework and uses the k-medoids clustering algorithm to recompute the initial anchor boxes of the original algorithm, resulting in new anchor box sizes. At the same time, an SENet layer is added to the C3 module to increase the weight of the algorithm when extracting features of small targets, thereby improving the detection and recognition performance. The experimental results show that the improved algorithm achieves an average precision of 99.2% at 0.5 IOU, and an average precision of 88.7 ^% at 0.5 IOU with a detection speed of 140 frames/s. The model size is only 13.5M, which has better recognition accuracy and success rate than the original yolov5 algorithm.