Object Detection For Low Contrast Images Based On Convolutional Neural Network

“Industry 4.0” and the “Made in China 2025” plan bring both opportunities and challenges for the development of manufacturing industry It is the inevitable choice for the development of manufacturing industry to apply more automation technology.Object detection is an important technique for automatic production.This dissertation aims at designing object detection algorithm for industry environment to solve the problem of object detection for low contrast images.According to the characteristics of object detection for low contrast images,a cascade CNN object detection algorithm is proposed.As it is difficult to label a large number of images as training set in industrial applications,several relatively shallow CNNs are design for object detection in this dissertation.We only need to label a few images and then argument the labeled data as the training data.Our cascaded CNN algorithm can be divided into three parts.The first part of our method applies a shallow CNN densely scanning the whole image and rejecting most of the background regions.In the second part,two CNNs are adopted to further evaluate the pass windows and the windows around them,and then roughly locate the objects.In the third part,a relatively deep model net-3 is applied to adjust the pass windows.The adjusted windows are regarded as final detections.For accelerating,we adopted a fully convolutional neural network to eliminate the repeated calculation of the first part.We adopt a nine-classes--classifier for adjusting detection results to reduce the calculation.We oppose a hard examples mining method to improve the capability of the CNNs in the second step.At the experiment part,we evaluate our method on some images.We also compare our method with ViDi,template matching and a CNN based algorithm.We adopt accuracy and recall rate as evaluation criteria.The experiment results verify the robustness and practicability of our algorithm.With a moderate GPU,our algorithm consumes about 72 ms per image with a size of 640×480.