The development of traditional object recognition systems normally involves a time consuming investigation of 'good' preprocessing and filtering methods, an extraction of appropriate image features and a creation of/search for a suitable classifier for feature vectors. The major goal of this approach is to avoid these problems and develop a different method by using raw pixel data of both small objects and background in large pictures as inputs to neural networks. This paper presents a series of algorithms and techniques for this approach, including the determination of neural network architecture, network training and testing, small object detection and finding the centres of the objects, and an evaluation of the method. Experiments arecarried out on easy, medium and difficult pictures. Both precision and recall can reach 100% for the easy and medium pictures, but the best precision is below 10% for the difficult pictures though its recall is 100% under a certain threshold. These results sugg...