Abstract: The production of electronic device container is a process with high requirements for safety, efficiency and integrity, which must be paid attention to by major enterprises. But in the actual production and packaging process, the stains on the container, the foreign matters in the container, and the appearance abnormalities are inevitable. These problems need to be solved urgently. At present, the main detection methods to solve these problems are manual detection and traditional machine vision. The disadvantages of manual detection are high accuracy and low efficiency. Traditional machine vision detection methods are of high efficiency and low accuracy, which are difficult to meet the requirements of high-speed automatic production line. Therefore, this paper proposes an electronic device container quality inspection method based on cascade R-CNN. In view of the actual process of the container data oriented improvement network, we add the samples that are difficult to distinguish from Focal Loss detection, use deformable convolution to extract features more efficiently, train the strong robustness model with multi-scale training method, and apply it to the multi-category detection of electronic device containers. The experimental results show that the improved model based on cascade R-CNN has high accuracy and strong robustness.

Key words: object detection, machine vision, convolutional neural network, directional detection, DCN, multi-scale