Object detection for blind inspection of industrial products based on neural architecture search

Object detection is a key technology to realize the blind inspection of industrial products. To improve the automation degree of building deep convolutional neural networks (CNNs) for object detection and further improve the detection accuracy, this paper proposes an improved neural architecture search method using exclusive-OR (XOR)-based channel feature fusion. First, an XOR-based channel fusion module is designed; it can fuse the feature mapping of different scales at the channel level in the case of multibranch access complementarily. Then, an improved cell pruning strategy is proposed to efficiently prune the connections between cells by setting the architecture parameters of the candidate operations to 0 s, which are in the alignment layers of the subsequent cells. The cell pruning strategy can directly search the multibranch CNN models and narrow the neural network architectures’ gap between the search stage and the evaluation stage. The experimental results show that the proposed method takes approximately 0.75 GPU days to search the optimal neural network on a dataset including six classes for blind inspection of industrial products, and the mean average precision (mAP) is approximately 99.1% on a test dataset, which is higher than those of state-of-the-art methods, e.g., DenseNAS and CSPDarknet53.