Object Detection Of E-commerce Images Based On Convolutional Neural Networks

E-commerce developed rapidly since it arose in China in 1999. By the end of 2013, Chinese e-commerce market transactions reached 10.2 trillion RMB, and online retail market deal size reached 1.8851 trillion RMB. People can use mobile devices to buy goods in electricity business platforms anywhere anytime. How to allow users to find their desired products quickly and easily in electronic business platforms is particularly important. Traditional way is to use the keywords to search. With respect to the information provided in the text, the image can save more high quality information. Intelligent mobile terminal equipment provides many powerful tools to capture screen and take photos. These conditions bring great convenience to use image to search products. It is foreseeable that in the future using image to search products will be popular in the future.In the e-commerce image search, we extract object information from input image using object detection technology, and then extract the image features from object instance, and finally use of image features to retrieve similar items picture. Using object detection technology can bring many benefits. Environments will be removed by object detection technology, then the system will just focus on the user's interested parts. What users care about is the goods in the picture not the environments.In the object detection task of the e-commerce images, the types of objects in the images are usually shoes, pants, shirt and so on. These objects usually consist of several parts, for example shoes usually consists of two shoe. Traditional object recognition algorithms are not very useful to deal with these situations.Therefore, we designed a new object detection algorithm. The algorithm is based on combination of the parts of object. First, we rotate the image to reduce the influence of the deflection of the objects in the images. Then we use a traditional object recognition algorithm to identify the parts of object in the images which are generated by rotating the same input image. Each object part may be a full body of object in the image or a part of a full body. Then, we cluster the object parts to find how many objects in the image. Finally, we use a method based on energy function to find object's position.We can see that the accuracy of our model is slightly higher than that of Faster R-CNN from the experimental results. The performance of our model is better than Faster R-CNN in the prediction of the position of the objects in the image. Taking into account that the image search need much better prediction of the location of the object, our model is better than Faster R-CNN.